diff --git a/.clang-tidy b/.clang-tidy
index 1a42b9abc..3078beacc 100644
--- a/.clang-tidy
+++ b/.clang-tidy
@@ -3,6 +3,7 @@ Checks: >
     bugprone-*,
     -bugprone-easily-swappable-parameters,
     -bugprone-implicit-widening-of-multiplication-result,
+    -bugprone-misplaced-widening-cast,
     -bugprone-narrowing-conversions,
     readability-*,
     -readability-avoid-unconditional-preprocessor-if,
@@ -15,4 +16,8 @@ Checks: >
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,
     portability-*,
+    misc-*,
+    -misc-const-correctness,
+    -misc-non-private-member-variables-in-classes,
+    -misc-no-recursion,
 FormatStyle: none
diff --git a/.devops/cloud-v-pipeline b/.devops/cloud-v-pipeline
new file mode 100644
index 000000000..f3a4944f8
--- /dev/null
+++ b/.devops/cloud-v-pipeline
@@ -0,0 +1,22 @@
+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    stage('Running llama.cpp'){
+        sh'''#!/bin/bash
+            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./main -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}
diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile
index e5fcb37d6..360602d65 100644
--- a/.devops/full-cuda.Dockerfile
+++ b/.devops/full-cuda.Dockerfile
@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip
+    apt-get install -y build-essential python3 python3-pip git
 
 COPY requirements.txt requirements.txt
 
diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile
new file mode 100644
index 000000000..6c521e9b4
--- /dev/null
+++ b/.devops/full-rocm.Dockerfile
@@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT ["/app/.devops/tools.sh"]
diff --git a/.devops/lamma-cpp-clblast.srpm.spec b/.devops/llama-cpp-clblast.srpm.spec
similarity index 56%
rename from .devops/lamma-cpp-clblast.srpm.spec
rename to .devops/llama-cpp-clblast.srpm.spec
index 739c68281..076f29695 100644
--- a/.devops/lamma-cpp-clblast.srpm.spec
+++ b/.devops/llama-cpp-clblast.srpm.spec
@@ -13,12 +13,13 @@
 #    It is up to the user to install the correct vendor-specific support.
 
 Name:           llama.cpp-clblast
-Version:        master
+Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
-Summary:        OpenCL Inference of LLaMA model in pure C/C++
+Summary:        OpenCL Inference of LLaMA model in C/C++
 License:        MIT
 Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
-BuildRequires:  coreutils make gcc-c++ git mesa-libOpenCL-devel
+BuildRequires:  coreutils make gcc-c++ git mesa-libOpenCL-devel clblast-devel
+Requires:       clblast
 URL:            https://github.com/ggerganov/llama.cpp
 
 %define debug_package %{nil}
@@ -35,18 +36,43 @@ make -j LLAMA_CLBLAST=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacppclblast
-cp -p server %{buildroot}%{_bindir}/llamacppclblastserver
-cp -p simple %{buildroot}%{_bindir}/llamacppclblastsimple
+cp -p main %{buildroot}%{_bindir}/llamaclblast
+cp -p server %{buildroot}%{_bindir}/llamaclblastserver
+cp -p simple %{buildroot}%{_bindir}/llamaclblastsimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamaclblast.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
 
 %clean
 rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacppclblast
-%{_bindir}/llamacppclblastserver
-%{_bindir}/llamacppclblastsimple
+%{_bindir}/llamaclblast
+%{_bindir}/llamaclblastserver
+%{_bindir}/llamaclblastsimple
+/usr/lib/systemd/system/llamaclblast.service
+%config /etc/sysconfig/llama
+
 
 %pre
 
diff --git a/.devops/lamma-cpp-cublas.srpm.spec b/.devops/llama-cpp-cublas.srpm.spec
similarity index 71%
rename from .devops/lamma-cpp-cublas.srpm.spec
rename to .devops/llama-cpp-cublas.srpm.spec
index 75d32fbe7..f847ebb1e 100644
--- a/.devops/lamma-cpp-cublas.srpm.spec
+++ b/.devops/llama-cpp-cublas.srpm.spec
@@ -13,7 +13,7 @@
 #    It is up to the user to install the correct vendor-specific support.
 
 Name:           llama.cpp-cublas
-Version:        master
+Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
 License:        MIT
@@ -40,6 +40,28 @@ cp -p main %{buildroot}%{_bindir}/llamacppcublas
 cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
 cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
 
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacublas.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
+
 %clean
 rm -rf %{buildroot}
 rm -rf %{_builddir}/*
@@ -48,6 +70,8 @@ rm -rf %{_builddir}/*
 %{_bindir}/llamacppcublas
 %{_bindir}/llamacppcublasserver
 %{_bindir}/llamacppcublassimple
+/usr/lib/systemd/system/llamacublas.service
+%config /etc/sysconfig/llama
 
 %pre
 
diff --git a/.devops/llama-cpp.srpm.spec b/.devops/llama-cpp.srpm.spec
index c65251a5a..446213d69 100644
--- a/.devops/llama-cpp.srpm.spec
+++ b/.devops/llama-cpp.srpm.spec
@@ -6,6 +6,7 @@
 # Notes for llama.cpp:
 # 1. Tags are currently based on hash - which will not sort asciibetically.
 #    We need to declare standard versioning if people want to sort latest releases.
+#    In the meantime, YYYYMMDD format will be used.
 # 2. Builds for CUDA/OpenCL support are separate, with different depenedencies.
 # 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed.
 #    Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo
@@ -13,12 +14,13 @@
 #    It is up to the user to install the correct vendor-specific support.
 
 Name:           llama.cpp
-Version:        master
+Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
 License:        MIT
 Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
-BuildRequires:  coreutils make gcc-c++ git
+BuildRequires:  coreutils make gcc-c++ git libstdc++-devel
+Requires:       libstdc++
 URL:            https://github.com/ggerganov/llama.cpp
 
 %define debug_package %{nil}
@@ -26,27 +28,52 @@ URL:            https://github.com/ggerganov/llama.cpp
 
 %description
 CPU inference for Meta's Lllama2 models using default options.
+Models are not included in this package and must be downloaded separately.
 
 %prep
-%autosetup
+%setup -n llama.cpp-master
 
 %build
 make -j
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacpp
-cp -p server %{buildroot}%{_bindir}/llamacppserver
-cp -p simple %{buildroot}%{_bindir}/llamacppsimple
+cp -p main %{buildroot}%{_bindir}/llama
+cp -p server %{buildroot}%{_bindir}/llamaserver
+cp -p simple %{buildroot}%{_bindir}/llamasimple
+
+mkdir -p %{buildroot}/usr/lib/systemd/system
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llama.service
+[Unit]
+Description=Llama.cpp server, CPU only (no GPU support in this build).
+After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
+
+[Service]
+Type=simple
+EnvironmentFile=/etc/sysconfig/llama
+ExecStart=/usr/bin/llamaserver $LLAMA_ARGS
+ExecReload=/bin/kill -s HUP $MAINPID
+Restart=never
+
+[Install]
+WantedBy=default.target
+EOF
+
+mkdir -p %{buildroot}/etc/sysconfig
+%{__cat} <<EOF  > %{buildroot}/etc/sysconfig/llama
+LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin"
+EOF
 
 %clean
 rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacpp
-%{_bindir}/llamacppserver
-%{_bindir}/llamacppsimple
+%{_bindir}/llama
+%{_bindir}/llamaserver
+%{_bindir}/llamasimple
+/usr/lib/systemd/system/llama.service
+%config /etc/sysconfig/llama
 
 %pre
 
diff --git a/.devops/main-cuda.Dockerfile b/.devops/main-cuda.Dockerfile
index 30c01196a..2b7faf7c1 100644
--- a/.devops/main-cuda.Dockerfile
+++ b/.devops/main-cuda.Dockerfile
@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential
+    apt-get install -y build-essential git
 
 WORKDIR /app
 
diff --git a/.devops/main-rocm.Dockerfile b/.devops/main-rocm.Dockerfile
new file mode 100644
index 000000000..789deff6d
--- /dev/null
+++ b/.devops/main-rocm.Dockerfile
@@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt requirements.txt
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/main" ]
diff --git a/.devops/tools.sh b/.devops/tools.sh
index 2787c21fe..9d999315f 100755
--- a/.devops/tools.sh
+++ b/.devops/tools.sh
@@ -7,15 +7,12 @@ arg1="$1"
 # Shift the arguments to remove the first one
 shift
 
-# Join the remaining arguments into a single string
-arg2="$@"
-
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert.py "$arg2"
+    python3 ./convert.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
-    ./quantize "$arg2"
+    ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
-    ./main "$arg2"
+    ./main "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -27,7 +24,7 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
         fi
     done
 elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
-    ./server "$arg2"
+    ./server "$@"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "
diff --git a/.dockerignore b/.dockerignore
index 462fac23a..633bbc3a9 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,18 +1,14 @@
 *.o
 *.a
 .cache/
+.git/
+.github/
+.gitignore
 .vs/
 .vscode/
 .DS_Store
 
-build/
-build-em/
-build-debug/
-build-release/
-build-static/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build*/
 
 models/*
 
diff --git a/.editorconfig b/.editorconfig
index 135a7e4bc..f8245b85c 100644
--- a/.editorconfig
+++ b/.editorconfig
@@ -17,3 +17,6 @@ indent_style = tab
 
 [prompts/*.txt]
 insert_final_newline = unset
+
+[examples/server/public/*]
+indent_size = 2
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 84faad37a..e41be76db 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -10,15 +10,14 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
-  GGML_NITER: 1
   GGML_N_THREADS: 1
 
 jobs:
@@ -28,7 +27,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -39,7 +38,13 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          CC=gcc-8 make
+          CC=gcc-8 make -j $(nproc)
+
+      - name: Test
+        id: make_test
+        run: |
+          CC=gcc-8 make tests -j $(nproc)
+          make test -j $(nproc)
 
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
@@ -47,7 +52,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -61,7 +66,7 @@ jobs:
           mkdir build
           cd build
           cmake ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -82,7 +87,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -96,7 +101,7 @@ jobs:
           mkdir build
           cd build
           cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -116,7 +121,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -130,7 +135,7 @@ jobs:
           mkdir build
           cd build
           cmake -DLLAMA_MPI=ON ..
-          cmake --build . --config Release
+          cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
@@ -144,7 +149,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -155,11 +160,46 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          make
+          make -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: make_test
+        run: |
+          make tests -j $(sysctl -n hw.logicalcpu)
+          make test -j $(sysctl -n hw.logicalcpu)
 
   macOS-latest-cmake:
     runs-on: macos-latest
 
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest --verbose --timeout 900
+
+  macOS-latest-cmake-ios:
+    runs-on: macos-latest
+
     steps:
       - name: Clone
         id: checkout
@@ -177,14 +217,64 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF ..
-          cmake --build . --config Release
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: Test
-        id: cmake_test
+  macOS-latest-cmake-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
         run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
           cd build
-          ctest --verbose --timeout 900
+          cmake -G Xcode .. \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+  macOS-latest-swift:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme llama -destination "${{ matrix.destination }}"
 
   windows-latest-cmake:
     runs-on: windows-latest
@@ -198,22 +288,24 @@ jobs:
       matrix:
         include:
           - build: 'noavx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2'
-            defines: '-DLLAMA_BUILD_SERVER=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'clblast'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
-            defines: '-DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
       - name: Download OpenCL SDK
         id: get_opencl
@@ -255,7 +347,7 @@ jobs:
           mkdir build
           cd build
           cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add clblast.dll
         id: add_clblast_dll
@@ -291,93 +383,95 @@ jobs:
           cd build
           ctest -C Release --verbose --timeout 900
 
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
 
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v3
         with:
           path: |
-            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
+            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
 
   windows-latest-cmake-cublas:
     runs-on: windows-latest
 
     strategy:
       matrix:
-        cuda: ['12.1.0', '11.7.1']
+        cuda: ['12.2.0', '11.7.1']
         build: ['cublas']
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
 
-      - uses: Jimver/cuda-toolkit@v0.2.10
+      - uses: Jimver/cuda-toolkit@v0.2.11
         id: cuda-toolkit
         with:
           cuda: ${{ matrix.cuda }}
-          # TODO(green-sky): _dev seems to fail, and non dev are not enought
-          #sub-packages: '["nvcc", "cudart", "cublas", "cudart_dev", "cublas_dev"]'
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON
-          cmake --build . --config Release
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
 
       - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v3
         with:
           path: |
-            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '12.1.0' }}
-        # TODO(green-sky): paths are cuda 12 specific
         run: |
           echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_12.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_12.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
-
-      - name: Copy and pack Cuda runtime
-        if: ${{ matrix.cuda == '11.7.1' }}
-        # TODO(green-sky): paths are cuda 11 specific
-        run: |
-          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
-          mkdir '.\build\bin\cudart\'
-          ls "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin"
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cudart64_110.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublas64_11.dll" '.\build\bin\cudart\'
-          cp "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin\cublasLt64_11.dll" '.\build\bin\cudart\'
-          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip .\build\bin\cudart\*
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -386,6 +480,23 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
+#  freeBSD-latest:
+#    runs-on: macos-12
+#    steps:
+#    - name: Clone
+#      uses: actions/checkout@v3
+#
+#    - name: Build
+#      uses: cross-platform-actions/action@v0.19.0
+#      with:
+#        operating_system: freebsd
+#        version: '13.2'
+#        hypervisor: 'qemu'
+#        run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake automake autoconf pkgconf llvm15 clinfo clover opencl clblast openblas
+#            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu`
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -400,21 +511,36 @@ jobs:
       - windows-latest-cmake-cublas
 
     steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
       - name: Download artifacts
         id: download-artifact
         uses: actions/download-artifact@v3
 
-      - name: Get commit hash
-        id: commit
-        uses: pr-mpt/actions-commit-hash@v2
-
       - name: Create release
         id: create_release
         uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ steps.tag.outputs.name }}
 
       - name: Upload release
         id: upload_release
@@ -447,7 +573,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -471,7 +597,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -495,7 +621,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
@@ -525,7 +651,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -564,7 +690,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -610,7 +736,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v1
+#        uses: actions/checkout@v3
 #
 #      - name: Dependencies
 #        run: |
diff --git a/.github/workflows/code-coverage.yml b/.github/workflows/code-coverage.yml
new file mode 100644
index 000000000..392db8a08
--- /dev/null
+++ b/.github/workflows/code-coverage.yml
@@ -0,0 +1,36 @@
+name: Code Coverage
+on: [push, pull_request]
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+
+jobs:
+  run:
+    runs-on: ubuntu-20.04
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 lcov
+
+      - name: Build
+        run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests
+
+      - name: Run tests
+        run: CC=gcc-8 make test
+
+      - name: Generate coverage report
+        run: |
+          make coverage
+          make lcov-report
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v3
+        env:
+           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
+        with:
+          files: lcov-report/coverage.info
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 379fbd7ad..9c90c77ac 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -26,8 +26,15 @@ jobs:
     strategy:
       matrix:
         config:
-          - { tag: "light", dockerfile: ".devops/main.Dockerfile" }
-          - { tag: "full", dockerfile: ".devops/full.Dockerfile" }
+          - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
+          #                     have disabled them for now until the reason why
+          #                     is understood.
+          - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3
@@ -51,7 +58,7 @@ jobs:
         with:
           context: .
           push: true
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
@@ -60,6 +67,6 @@ jobs:
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.config.platforms }}
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}
diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml
new file mode 100644
index 000000000..57db17512
--- /dev/null
+++ b/.github/workflows/gguf-publish.yml
@@ -0,0 +1,44 @@
+# This workflow will upload a Python Package using Twine when a GGUF release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# See `gguf-py/README.md` for how to make a release.
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  workflow_dispatch:
+  push:
+    # Pattern matched against refs/tags
+    tags:
+      - 'gguf-v*'           # Push events to every version tag
+
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v3
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.9.x'
+    - name: Install dependencies
+      run: |
+        cd gguf-py
+        python -m pip install poetry
+        poetry install
+
+    - name: Build package
+      run: cd gguf-py && poetry build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        password: ${{ secrets.PYPI_API_TOKEN }}
+        packages-dir: gguf-py/dist
diff --git a/.gitignore b/.gitignore
index f3121794a..420e0d6d0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,6 +5,12 @@
 *.bin
 *.exe
 *.dll
+*.log
+*.gcov
+*.gcno
+*.gcda
+*.dot
+*.metallib
 .DS_Store
 .build/
 .cache/
@@ -16,50 +22,52 @@
 .vs/
 .vscode/
 
-build/
-build-em/
-build-debug/
-build-release/
-build-ci-debug/
-build-ci-release/
-build-static/
-build-cublas/
-build-opencl/
-build-metal/
-build-mpi/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+lcov-report/
+gcovr-report/
+
+build*/
 out/
 tmp/
 
 models/*
 models-mnt
 
+/Pipfile
+/baby-llama
+/beam-search
+/benchmark-matmult
+/convert-llama2c-to-ggml
+/embd-input-test
+/embedding
+/gguf
+/gguf-llama-simple
+/infill
+/libllama.so
+/llama-bench
 /main
+/metal
+/perplexity
+/q8dot
 /quantize
 /quantize-stats
 /result
-/perplexity
-/embedding
-/train-text-from-scratch
-/convert-llama2c-to-ggml
-/simple
-/benchmark-matmult
-/vdot
+/save-load-state
 /server
-/Pipfile
-/embd-input-test
-/gguf
-/gguf-llama-simple
-/libllama.so
-/llama-bench
+/simple
+/batched
+/export-lora
+/finetune
+/speculative
+/parallel
+/train-text-from-scratch
+/vdot
 build-info.h
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
 
 __pycache__
+dist
 
 zig-out/
 zig-cache/
@@ -70,16 +78,19 @@ perf-*.txt
 
 examples/jeopardy/results.txt
 
-pyproject.toml
 poetry.lock
 poetry.toml
 
 # Test binaries
 tests/test-grammar-parser
+tests/test-llama-grammar
 tests/test-double-float
 tests/test-grad0
 tests/test-opt
 tests/test-quantize-fns
 tests/test-quantize-perf
 tests/test-sampling
-tests/test-tokenizer-0
+tests/test-tokenizer-0-llama
+tests/test-tokenizer-0-falcon
+tests/test-tokenizer-1-llama
+tests/test-tokenizer-1-bpe
diff --git a/CMakeLists.txt b/CMakeLists.txt
index bb63ef98e..7c79ec486 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.12) # Don't bump this version for no reason
+cmake_minimum_required(VERSION 3.13)  # for add_link_options
 project("llama.cpp" C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -36,9 +36,15 @@ endif()
 # Option list
 #
 
+if (APPLE)
+    set(LLAMA_METAL_DEFAULT ON)
+else()
+    set(LLAMA_METAL_DEFAULT OFF)
+endif()
+
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
@@ -52,15 +58,21 @@ option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"
 option(LLAMA_SANITIZE_UNDEFINED         "llama: enable undefined sanitizer"                     OFF)
 
 # instruction set specific
-option(LLAMA_AVX                        "llama: enable AVX"                                     ON)
-option(LLAMA_AVX2                       "llama: enable AVX2"                                    ON)
-option(LLAMA_AVX512                     "llama: enable AVX512"                                  OFF)
-option(LLAMA_AVX512_VBMI                "llama: enable AVX512-VBMI"                             OFF)
-option(LLAMA_AVX512_VNNI                "llama: enable AVX512-VNNI"                             OFF)
-option(LLAMA_FMA                        "llama: enable FMA"                                     ON)
+if (LLAMA_NATIVE)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
+option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
+option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
+option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
+option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
+option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
-    option(LLAMA_F16C                   "llama: enable F16C"                                    ON)
+    option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
 # 3rd party libs
@@ -74,8 +86,12 @@ set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_MMV_Y        "1" CACHE STRING "llama: y block size for mmv CUDA kernels")
 option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some calculations"   OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
+set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                             "llama: max. batch size for using peer access")
+option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
-option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
+option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
+option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -108,7 +124,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
     add_custom_command(
         OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
         COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
+        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         DEPENDS "${GIT_DIR}/index"
         VERBATIM
@@ -127,6 +143,7 @@ set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
+include(CheckCXXCompilerFlag)
 
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
@@ -151,12 +168,40 @@ if (APPLE AND LLAMA_ACCELERATE)
         message(STATUS "Accelerate framework found")
 
         add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
     else()
         message(WARNING "Accelerate framework not found")
     endif()
 endif()
 
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+    message(STATUS "Metal framework found")
+    set(GGML_HEADERS_METAL ggml-metal.h)
+    set(GGML_SOURCES_METAL ggml-metal.m)
+
+    add_compile_definitions(GGML_USE_METAL)
+    if (LLAMA_METAL_NDEBUG)
+        add_compile_definitions(GGML_METAL_NDEBUG)
+    endif()
+
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        )
+endif()
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -233,7 +278,8 @@ if (LLAMA_BLAS)
 endif()
 
 if (LLAMA_K_QUANTS)
-    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    set(GGML_HEADERS_EXTRA k_quants.h)
+    set(GGML_SOURCES_EXTRA k_quants.c)
     add_compile_definitions(GGML_USE_K_QUANTS)
     if (LLAMA_QKK_64)
         add_compile_definitions(GGML_QKK_64)
@@ -249,7 +295,8 @@ if (LLAMA_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+        set(GGML_HEADERS_CUDA ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
 #        if (LLAMA_CUDA_CUBLAS)
@@ -267,6 +314,7 @@ if (LLAMA_CUBLAS)
             add_compile_definitions(GGML_CUDA_F16)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -292,39 +340,18 @@ if (LLAMA_CUBLAS)
     endif()
 endif()
 
-if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-
-    set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
-
-    add_compile_definitions(GGML_USE_METAL)
-    add_compile_definitions(GGML_METAL_NDEBUG)
-
-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
-    # copy ggml-metal.metal to bin directory
-    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
-        ${FOUNDATION_LIBRARY}
-        ${METAL_FRAMEWORK}
-        ${METALKIT_FRAMEWORK}
-        )
-endif()
-
 if (LLAMA_MPI)
     cmake_minimum_required(VERSION 3.10)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
+        set(GGML_HEADERS_MPI ggml-mpi.h)
         set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-        set(cxx_flags ${cxx_flags} -Wno-cast-qual)
-        set(c_flags   ${c_flags}   -Wno-cast-qual)
+        if (NOT MSVC)
+            add_compile_options(-Wno-cast-qual)
+        endif()
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
         # Even if you're only using the C header, C++ programs may bring in MPI
@@ -342,7 +369,8 @@ if (LLAMA_CLBLAST)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_HEADERS_OPENCL ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
@@ -352,39 +380,98 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
+        endif()
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    endif()
+endif()
+
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        set(c_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wdouble-promotion
-            -Wshadow
-            -Wstrict-prototypes
-            -Wpointer-arith
-            -Wmissing-prototypes
-        )
-        set(cxx_flags
-            -Wall
-            -Wextra
-            -Wpedantic
-            -Wcast-qual
-            -Wno-unused-function
-            -Wno-multichar
-        )
+        set(warning_flags -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(c_flags -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int
+            -Werror=implicit-function-declaration)
+        set(cxx_flags -Wmissing-declarations -Wmissing-noreturn)
+        set(host_cxx_flags "")
+
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
+            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
+
+            if (
+                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
+                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
+            )
+                set(c_flags ${c_flags} -Wdouble-promotion)
+            endif()
+        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+            set(c_flags ${c_flags} -Wdouble-promotion)
+            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
+
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
+            endif()
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
+            endif()
+        endif()
     else()
         # todo : msvc
     endif()
 
-    add_compile_options(
-            "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
-            "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
-    )
+    set(c_flags   ${c_flags}   ${warning_flags})
+    set(cxx_flags ${cxx_flags} ${warning_flags})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags} ${host_cxx_flags}>")
 
 endif()
 
-if (MSVC)
+if (NOT MSVC)
+    set(cuda_flags -Wno-pedantic)
+endif()
+set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+
+list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
+if (NOT cuda_host_flags STREQUAL "")
+    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
+
+if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
     if (BUILD_SHARED_LIBS)
@@ -406,6 +493,13 @@ endif()
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
 if (NOT MSVC)
     if (LLAMA_STATIC)
         add_link_options(-static)
@@ -416,30 +510,35 @@ if (NOT MSVC)
     if (LLAMA_GPROF)
         add_compile_options(-pg)
     endif()
-    if (LLAMA_NATIVE)
-        add_compile_options(-march=native)
-    endif()
 endif()
 
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
     if (MSVC)
-        # TODO: arm msvc?
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+        add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
     else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            add_compile_options(-mfp16-format=ieee)
+        endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
             # Raspberry Pi 2
-            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            add_compile_options(-mfp16-format=ieee -mno-unaligned-access)
+            add_compile_options(-mno-unaligned-access)
         endif()
     endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
         if (LLAMA_AVX512)
@@ -465,6 +564,9 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
+        if (LLAMA_NATIVE)
+            add_compile_options(-march=native)
+        endif()
         if (LLAMA_F16C)
             add_compile_options(-mf16c)
         endif()
@@ -496,27 +598,84 @@ else()
     message(STATUS "Unknown architecture")
 endif()
 
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
 #
 # libraries
 #
 
 # ggml
 
+if (GGML_USE_CPU_HBM)
+    add_definitions(-DGGML_USE_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+endif()
+
 add_library(ggml OBJECT
             ggml.c
             ggml.h
             ggml-alloc.c
             ggml-alloc.h
-            ${GGML_SOURCES_CUDA}
-            ${GGML_SOURCES_OPENCL}
-            ${GGML_SOURCES_METAL}
-            ${GGML_SOURCES_MPI}
-            ${GGML_SOURCES_EXTRA}
+            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
 target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+if (GGML_USE_CPU_HBM)
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
 if (BUILD_SHARED_LIBS)
@@ -546,14 +705,54 @@ if (BUILD_SHARED_LIBS)
     if (LLAMA_METAL)
         set_target_properties(llama PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
     endif()
-    install(TARGETS llama LIBRARY)
 endif()
 
+
 #
 # install
 #
 
 include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR}
+    CACHE PATH "Location of header files")
+set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR}
+    CACHE PATH "Location of library files")
+set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR}
+    CACHE PATH "Location of binary files")
+set(LLAMA_BUILD_NUMBER ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/scripts/LlamaConfig.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama
+    PATH_VARS LLAMA_INCLUDE_INSTALL_DIR
+              LLAMA_LIB_INSTALL_DIR
+              LLAMA_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+    VERSION ${LLAMA_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
+
+set(GGML_PUBLIC_HEADERS "ggml.h"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+install(TARGETS ggml PUBLIC_HEADER)
+
+set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/llama.h)
+install(TARGETS llama LIBRARY PUBLIC_HEADER)
+
 install(
     FILES convert.py
     PERMISSIONS
diff --git a/Makefile b/Makefile
index d31acc450..b8b0d4b56 100644
--- a/Makefile
+++ b/Makefile
@@ -1,10 +1,11 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test gguf llama-bench
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
-TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
+TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
 
-default: $(BUILD_TARGETS)
+# Code coverage output files
+COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -18,12 +19,27 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
-CXXV := $(shell $(CXX) --version | head -n 1)
+ifeq '' '$(findstring clang,$(shell $(CC) --version))'
+	CC_IS_GCC=1
+	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+else
+	CC_IS_CLANG=1
+	ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
+		CC_IS_LLVM_CLANG=1
+	else
+		CC_IS_APPLE_CLANG=1
+	endif
+	CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
+				| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+endif
 
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
+	ifndef LLAMA_NO_METAL
+		LLAMA_METAL := 1
+	endif
+
 	ifneq ($(UNAME_P),arm)
 		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
 		ifeq ($(SYSCTL_M),1)
@@ -34,64 +50,176 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
+ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
+BUILD_TARGETS += metal
+endif
+
+default: $(BUILD_TARGETS)
+
+test: $(TEST_TARGETS)
+	@failures=0; \
+	for test_target in $(TEST_TARGETS); do \
+		if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \
+			continue; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \
+			continue; \
+		else \
+			echo "Running test $$test_target..."; \
+			./$$test_target; \
+		fi; \
+		if [ $$? -ne 0 ]; then \
+			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			failures=$$(( failures + 1 )); \
+		else \
+			printf 'Test %s passed.\n\n' $$test_target; \
+		fi; \
+	done; \
+	if [ $$failures -gt 0 ]; then \
+		printf '\n%s tests failed.\n' $$failures; \
+		exit 1; \
+	fi
+	@echo 'All tests passed.'
+
+all: $(BUILD_TARGETS) $(TEST_TARGETS)
+
+coverage: ## Run code coverage
+	gcov -pb tests/*.cpp
+
+lcov-report: coverage ## Generate lcov report
+	mkdir -p lcov-report
+	lcov --capture --directory . --output-file lcov-report/coverage.info
+	genhtml lcov-report/coverage.info --output-directory lcov-report
+
+gcovr-report: coverage ## Generate gcovr report
+	mkdir -p gcovr-report
+	gcovr --root . --html --html-details --output gcovr-report/coverage.html
+
+ifdef RISCV_CROSS_COMPILE
+CC	:= riscv64-unknown-linux-gnu-gcc
+CXX	:= riscv64-unknown-linux-gnu-g++
+endif
+
 #
 # Compile flags
 #
 
 # keep standard at C11 and C++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
+
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-OPT = -Ofast
+MK_CFLAGS        += -Ofast
+MK_HOST_CXXFLAGS += -Ofast
+MK_CUDA_CXXFLAGS += -O3
 else
-OPT = -O3
+MK_CFLAGS        += -O3
+MK_CXXFLAGS      += -O3
+endif
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+MK_CPPFLAGS += -D_XOPEN_SOURCE=600
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+endif
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+ifeq ($(UNAME_S),Linux)
+	MK_CPPFLAGS += -D_GNU_SOURCE
+endif
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+ifeq ($(UNAME_S),Darwin)
+	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
+endif
+ifeq ($(UNAME_S),DragonFly)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+ifeq ($(UNAME_S),FreeBSD)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+ifeq ($(UNAME_S),NetBSD)
+	MK_CPPFLAGS += -D_NETBSD_SOURCE
+endif
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
-CFLAGS   = -I.            $(OPT) -std=c11   -fPIC
-CXXFLAGS = -I. -I./common $(OPT) -std=c++11 -fPIC
-LDFLAGS  =
 
 ifdef LLAMA_DEBUG
-	CFLAGS   += -O0 -g
-	CXXFLAGS += -O0 -g
-	LDFLAGS  += -g
+	MK_CFLAGS   += -O0 -g
+	MK_CXXFLAGS += -O0 -g
+	MK_LDFLAGS  += -g
 else
-	CFLAGS   += -DNDEBUG
-	CXXFLAGS += -DNDEBUG
+	MK_CPPFLAGS += -DNDEBUG
 endif
 
 ifdef LLAMA_SERVER_VERBOSE
-	CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
+	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
+
+ifdef LLAMA_CODE_COVERAGE
+	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
+endif
+
+ifdef LLAMA_DISABLE_LOGS
+	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
+endif # LLAMA_DISABLE_LOGS
+
 # warnings
-CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \
-			-Wmissing-prototypes
-CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
+WARN_FLAGS    = -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int \
+				-Werror=implicit-function-declaration
+MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
+
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
+endif
 
 # OS specific
 # TODO: support Windows
-ifeq ($(UNAME_S),Linux)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Darwin)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),FreeBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),NetBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Haiku)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
+ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
+	MK_CFLAGS   += -pthread
+	MK_CXXFLAGS += -pthread
 endif
 
 # detect Windows
@@ -117,104 +245,119 @@ ifeq ($(_WIN32),1)
 endif
 
 ifdef LLAMA_GPROF
-	CFLAGS   += -pg
-	CXXFLAGS += -pg
+	MK_CFLAGS   += -pg
+	MK_CXXFLAGS += -pg
 endif
 ifdef LLAMA_PERF
-	CFLAGS   += -DGGML_PERF
-	CXXFLAGS += -DGGML_PERF
+	MK_CPPFLAGS += -DGGML_PERF
 endif
 
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
+
+ifndef RISCV
+
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
-	CFLAGS   += -march=native -mtune=native
-	CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS   += -march=native -mtune=native
+	MK_HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
-	#CFLAGS   += -mfma -mf16c -mavx
-	#CXXFLAGS += -mfma -mf16c -mavx
+	#MK_CFLAGS   += -mfma -mf16c -mavx
+	#MK_CXXFLAGS += -mfma -mf16c -mavx
 
 	# Usage SSSE3-only (Not is SSE3!)
-	#CFLAGS   += -mssse3
-	#CXXFLAGS += -mssse3
+	#MK_CFLAGS   += -mssse3
+	#MK_CXXFLAGS += -mssse3
+endif
+
+# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+# https://github.com/ggerganov/llama.cpp/issues/2922
+ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
+	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
+	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	CFLAGS   += -mcpu=native
-	CXXFLAGS += -mcpu=native
+	MK_CFLAGS   += -mcpu=native
+	MK_CXXFLAGS += -mcpu=native
 endif
 
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, Zero
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
 
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# Raspberry Pi 2
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 3, 4, Zero 2 (32-bit)
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+	MK_CFLAGS   += -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		CFLAGS   += -mcpu=power9
-		CXXFLAGS += -mcpu=power9
-	endif
-	# Require c++23's std::byteswap for big-endian support.
-	ifeq ($(UNAME_M),ppc64)
-		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
+		MK_CFLAGS   += -mcpu=power9
+		MK_CXXFLAGS += -mcpu=power9
 	endif
 endif
 
+else
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
+endif
+
 ifndef LLAMA_NO_K_QUANTS
-	CFLAGS   += -DGGML_USE_K_QUANTS
-	CXXFLAGS += -DGGML_USE_K_QUANTS
+	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
 	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
-	CFLAGS   += -DGGML_QKK_64
-	CXXFLAGS += -DGGML_QKK_64
+	MK_CPPFLAGS += -DGGML_QKK_64
 endif
 endif
 
 ifndef LLAMA_NO_ACCELERATE
-	# Mac M1 - include Accelerate framework.
-	# `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time).
+	# Mac OS - include Accelerate framework.
+	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		CFLAGS  += -DGGML_USE_ACCELERATE
-		LDFLAGS += -framework Accelerate
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE
+		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
+		MK_LDFLAGS  += -framework Accelerate
 	endif
 endif # LLAMA_NO_ACCELERATE
 
 ifdef LLAMA_MPI
-	CFLAGS += -DGGML_USE_MPI -Wno-cast-qual
-	CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
+	MK_CPPFLAGS += -DGGML_USE_MPI
+	MK_CFLAGS   += -Wno-cast-qual
+	MK_CXXFLAGS += -Wno-cast-qual
 	OBJS     += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags openblas)
-	LDFLAGS += $(shell pkg-config --libs openblas)
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
 ifdef LLAMA_BLIS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
-	LDFLAGS += -lblis -L/usr/local/lib
+	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	CFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
+	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
 	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
@@ -253,6 +396,11 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
+ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+else
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
 #	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
@@ -260,19 +408,20 @@ ifdef LLAMA_CUDA_CCBIN
 	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(subst -Ofast,-O3,$(CXXFLAGS)) -Wno-pedantic -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
 
-	CFLAGS   += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
-	CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
+	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
 
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
-		LDFLAGS += -lclblast -framework OpenCL
+		MK_LDFLAGS += -lclblast -framework OpenCL
 	else
-		LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
+		MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
 	endif
 	OBJS    += ggml-opencl.o
 
@@ -280,11 +429,35 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
+ifdef LLAMA_HIPBLAS
+	ROCM_PATH	?= /opt/rocm
+	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	LLAMA_CUDA_DMMV_X       ?= 32
+	LLAMA_CUDA_MMV_Y        ?= 1
+	LLAMA_CUDA_KQUANTS_ITER ?= 2
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+ifdef LLAMA_CUDA_FORCE_DMMV
+	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
+endif # LLAMA_CUDA_FORCE_DMMV
+	OBJS        += ggml-cuda.o
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # LLAMA_HIPBLAS
+
 ifdef LLAMA_METAL
-	CFLAGS   += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
-	CXXFLAGS += -DGGML_USE_METAL
-	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJS     += ggml-metal.o
+	MK_CPPFLAGS += -DGGML_USE_METAL
+	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	OBJS		+= ggml-metal.o
+ifdef LLAMA_METAL_NDEBUG
+	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
@@ -297,24 +470,36 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifdef LLAMA_NO_K_QUANTS
+ifndef LLAMA_NO_K_QUANTS
 k_quants.o: k_quants.c k_quants.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_NO_K_QUANTS
 
+# combine build flags with cmdline overrides
+override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
+override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
+override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
+override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
+
+# save CXXFLAGS before we add host-only options
+NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
+override CXXFLAGS += $(HOST_CXXFLAGS)
+
 #
 # Print build information
 #
 
 $(info I llama.cpp build info: )
-$(info I UNAME_S:  $(UNAME_S))
-$(info I UNAME_P:  $(UNAME_P))
-$(info I UNAME_M:  $(UNAME_M))
-$(info I CFLAGS:   $(CFLAGS))
-$(info I CXXFLAGS: $(CXXFLAGS))
-$(info I LDFLAGS:  $(LDFLAGS))
-$(info I CC:       $(CCV))
-$(info I CXX:      $(CXXV))
+$(info I UNAME_S:   $(UNAME_S))
+$(info I UNAME_P:   $(UNAME_P))
+$(info I UNAME_M:   $(UNAME_M))
+$(info I CFLAGS:    $(CFLAGS))
+$(info I CXXFLAGS:  $(CXXFLAGS))
+$(info I NVCCFLAGS: $(NVCCFLAGS))
+$(info I LDFLAGS:   $(LDFLAGS))
+$(info I CC:        $(shell $(CC) --version | head -n 1))
+$(info I CXX:       $(shell $(CXX) --version | head -n 1))
 $(info )
 
 #
@@ -332,7 +517,7 @@ OBJS += ggml-alloc.o
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common.o: common/common.cpp common/common.h
+common.o: common/common.cpp common/common.h build-info.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 console.o: common/console.cpp common/console.h
@@ -341,11 +526,14 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+train.o: common/train.cpp common/train.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch convert-llama2c-to-ggml embd-input-test gguf llama-bench build-info.h $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
@@ -357,9 +545,15 @@ main: examples/main/main.cpp                                  build-info.h ggml.
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
+infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o common.o console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -385,20 +579,43 @@ $(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-in
 embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
-gguf: examples/gguf/gguf.cpp                                  build-info.h ggml.o llama.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o common.o $(OBJS)
+train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp    build-info.h ggml.o llama.o $(OBJS)
+convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o common.o train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+ifdef LLAMA_METAL
+metal: examples/metal/metal.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif
+
 build-info.h: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh > $@.tmp
+	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
 	else \
@@ -413,34 +630,53 @@ tests: $(TEST_TARGETS)
 
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+run-benchmark-matmult: benchmark-matmult
 	./$@
 
+.PHONY: run-benchmark-matmult
+
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0: tests/test-tokenizer-0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-c.o: tests/test-c.c llama.h
+	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
diff --git a/Package.swift b/Package.swift
index 73d027c70..1ea414cc1 100644
--- a/Package.swift
+++ b/Package.swift
@@ -2,8 +2,33 @@
 
 import PackageDescription
 
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v11),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let resources: [Resource] = [
+    .process("ggml-metal.metal")
+]
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let resources: [Resource] = []
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
     name: "llama",
+    platforms: platforms,
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -11,14 +36,29 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: ["ggml-metal.metal"],
-            sources: ["ggml.c", "llama.cpp"],
+            exclude: exclude,
+            sources: [
+                "ggml.c",
+                "llama.cpp",
+                "ggml-alloc.c",
+                "k_quants.c",
+            ] + additionalSources,
+            resources: resources,
             publicHeadersPath: "spm-headers",
-            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32"]),
+                .define("GGML_USE_K_QUANTS"),
+                .define("GGML_USE_ACCELERATE")
+                // NOTE: NEW_LAPACK will required iOS version 16.4+
+                // We should consider add this in the future when we drop support for iOS 14
+                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+                // .define("ACCELERATE_NEW_LAPACK"),
+                // .define("ACCELERATE_LAPACK_ILP64")
+            ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]
-        ),
+        )
     ],
     cxxLanguageStandard: .cxx11
 )
diff --git a/README.md b/README.md
index f746c49eb..056279562 100644
--- a/README.md
+++ b/README.md
@@ -5,21 +5,18 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398)
+- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
+- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
+  **Devs should become familiar with the new API**
+- Local Falcon 180B inference on Mac Studio
 
-Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa)
-
-### Current `master` should be considered in Beta - expect some issues for a few days!
-
-### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up!
-
-### Issues with non-GGUF models will be considered with low priority!
+  https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
 
 ----
 
@@ -66,12 +63,11 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 - Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
-- 4-bit, 5-bit and 8-bit integer quantization support
-- Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS
-- cuBLAS and CLBlast support
+- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
+- CUDA, Metal and OpenCL GPU backend support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
-Since then, the project has improved significantly thanks to many contributions. This project is for educational purposes and serves
+Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves
 as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
 
 **Supported platforms:**
@@ -85,6 +81,7 @@ as the main playground for developing new features for the [ggml](https://github
 
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [X] Falcon
 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
 - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
@@ -96,109 +93,109 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
 - [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
+- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
+- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
+- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
 
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
+- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
+- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
+- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
 
 **UI:**
 
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
+- [withcatai/catai](https://github.com/withcatai/catai)
 
 ---
 
-Here is a typical run using LLaMA-7B:
+Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
 ```java
-make -j && ./main -m ./models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512
+$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
 I UNAME_P:  arm
 I UNAME_M:  arm64
-I CFLAGS:   -I.              -O3 -DNDEBUG -std=c11   -fPIC -pthread -DGGML_USE_ACCELERATE
-I CXXFLAGS: -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -pthread
+I CFLAGS:   -I.            -O3 -std=c11   -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -pthread -DGGML_USE_K_QUANTS -DGGML_USE_ACCELERATE
+I CXXFLAGS: -I. -I./common -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -DGGML_USE_K_QUANTS
 I LDFLAGS:   -framework Accelerate
-I CC:       Apple clang version 14.0.0 (clang-1400.0.29.202)
-I CXX:      Apple clang version 14.0.0 (clang-1400.0.29.202)
+I CC:       Apple clang version 14.0.3 (clang-1403.0.22.14.1)
+I CXX:      Apple clang version 14.0.3 (clang-1403.0.22.14.1)
 
 make: Nothing to be done for `default'.
-main: seed = 1678486056
-llama_model_load: loading model from './models/7B/ggml-model-q4_0.bin' - please wait ...
-llama_model_load: n_vocab = 32000
-llama_model_load: n_ctx   = 512
-llama_model_load: n_embd  = 4096
-llama_model_load: n_mult  = 256
-llama_model_load: n_head  = 32
-llama_model_load: n_layer = 32
-llama_model_load: n_rot   = 128
-llama_model_load: f16     = 2
-llama_model_load: n_ff    = 11008
-llama_model_load: ggml ctx size = 4529.34 MB
-llama_model_load: memory_size =   512.00 MB, n_mem = 16384
-llama_model_load: .................................... done
-llama_model_load: model size =  4017.27 MB / num tensors = 291
+main: build = 1041 (cf658ad)
+main: seed  = 1692823051
+llama_model_loader: loaded meta data with 16 key-value pairs and 363 tensors from models/llama-13b-v2/ggml-model-q4_0.gguf (version GGUF V1 (latest))
+llama_model_loader: - type  f32:   81 tensors
+llama_model_loader: - type q4_0:  281 tensors
+llama_model_loader: - type q6_K:    1 tensors
+llm_load_print_meta: format         = GGUF V1 (latest)
+llm_load_print_meta: arch           = llama
+llm_load_print_meta: vocab type     = SPM
+llm_load_print_meta: n_vocab        = 32000
+llm_load_print_meta: n_merges       = 0
+llm_load_print_meta: n_ctx_train    = 4096
+llm_load_print_meta: n_ctx          = 512
+llm_load_print_meta: n_embd         = 5120
+llm_load_print_meta: n_head         = 40
+llm_load_print_meta: n_head_kv      = 40
+llm_load_print_meta: n_layer        = 40
+llm_load_print_meta: n_rot          = 128
+llm_load_print_meta: n_gqa          = 1
+llm_load_print_meta: f_norm_eps     = 1.0e-05
+llm_load_print_meta: f_norm_rms_eps = 1.0e-05
+llm_load_print_meta: n_ff           = 13824
+llm_load_print_meta: freq_base      = 10000.0
+llm_load_print_meta: freq_scale     = 1
+llm_load_print_meta: model type     = 13B
+llm_load_print_meta: model ftype    = mostly Q4_0
+llm_load_print_meta: model size     = 13.02 B
+llm_load_print_meta: general.name   = LLaMA v2
+llm_load_print_meta: BOS token = 1 '<s>'
+llm_load_print_meta: EOS token = 2 '</s>'
+llm_load_print_meta: UNK token = 0 '<unk>'
+llm_load_print_meta: LF token  = 13 '<0x0A>'
+llm_load_tensors: ggml ctx size =    0.11 MB
+llm_load_tensors: mem required  = 7024.01 MB (+  400.00 MB per state)
+...................................................................................................
+llama_new_context_with_model: kv self size  =  400.00 MB
+llama_new_context_with_model: compute buffer total size =   75.41 MB
 
-main: prompt: 'Building a website can be done in 10 simple steps:'
-main: number of tokens in prompt = 15
-     1 -> ''
-  8893 -> 'Build'
-   292 -> 'ing'
-   263 -> ' a'
-  4700 -> ' website'
-   508 -> ' can'
-   367 -> ' be'
-  2309 -> ' done'
-   297 -> ' in'
- 29871 -> ' '
- 29896 -> '1'
- 29900 -> '0'
-  2560 -> ' simple'
-  6576 -> ' steps'
- 29901 -> ':'
-
-sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000
+system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
+sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
+generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0
 
 
-Building a website can be done in 10 simple steps:
-1) Select a domain name and web hosting plan
-2) Complete a sitemap
-3) List your products
-4) Write product descriptions
-5) Create a user account
-6) Build the template
-7) Start building the website
-8) Advertise the website
-9) Provide email support
-10) Submit the website to search engines
-A website is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves.
-The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user's browser.
-The web pages are stored in a web server. The web server is also called a host. When the website is accessed, it is retrieved from the server and displayed on the user's computer.
-A website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server.
-A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user's screen.
-A website can also be viewed on different devices such as desktops, tablets and smartphones.
-Hence, to have a website displayed on a browser, the website must be hosted.
-A domain name is an address of a website. It is the name of the website.
-The website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server.
-A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user’s screen.
-A website can also be viewed on different devices such as desktops, tablets and smartphones. Hence, to have a website displayed on a browser, the website must be hosted.
-A domain name is an address of a website. It is the name of the website.
-A website is an address of a website. It is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves.
-The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user’s browser.
-A website is known as a website when it is hosted
-
-main: mem per token = 14434244 bytes
-main:     load time =  1332.48 ms
-main:   sample time =  1081.40 ms
-main:  predict time = 31378.77 ms / 61.41 ms per token
-main:    total time = 34036.74 ms
+ Building a website can be done in 10 simple steps:
+Step 1: Find the right website platform.
+Step 2: Choose your domain name and hosting plan.
+Step 3: Design your website layout.
+Step 4: Write your website content and add images.
+Step 5: Install security features to protect your site from hackers or spammers
+Step 6: Test your website on multiple browsers, mobile devices, operating systems etc…
+Step 7: Test it again with people who are not related to you personally – friends or family members will work just fine!
+Step 8: Start marketing and promoting the website via social media channels or paid ads
+Step 9: Analyze how many visitors have come to your site so far, what type of people visit more often than others (e.g., men vs women) etc…
+Step 10: Continue to improve upon all aspects mentioned above by following trends in web design and staying up-to-date on new technologies that can enhance user experience even further!
+How does a Website Work?
+A website works by having pages, which are made of HTML code. This code tells your computer how to display the content on each page you visit – whether it’s an image or text file (like PDFs). In order for someone else’s browser not only be able but also want those same results when accessing any given URL; some additional steps need taken by way of programming scripts that will add functionality such as making links clickable!
+The most common type is called static HTML pages because they remain unchanged over time unless modified manually (either through editing files directly or using an interface such as WordPress). They are usually served up via HTTP protocols – this means anyone can access them without having any special privileges like being part of a group who is allowed into restricted areas online; however, there may still exist some limitations depending upon where one lives geographically speaking.
+How to
+llama_print_timings:        load time =   576.45 ms
+llama_print_timings:      sample time =   283.10 ms /   400 runs   (    0.71 ms per token,  1412.91 tokens per second)
+llama_print_timings: prompt eval time =   599.83 ms /    19 tokens (   31.57 ms per token,    31.68 tokens per second)
+llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms per token,    16.28 tokens per second)
+llama_print_timings:       total time = 25431.49 ms
 ```
 
 And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
@@ -277,29 +274,11 @@ In order to build llama.cpp you have three different options.
 
 ### Metal Build
 
-Using Metal allows the computation to be executed on the GPU for Apple devices:
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or the `LLAMA_METAL=OFF` cmake option.
 
-- Using `make`:
-
-  ```bash
-  LLAMA_METAL=1 make
-  ```
-
-- Using `CMake`:
-
-    ```bash
-    mkdir build-metal
-    cd build-metal
-    cmake -DLLAMA_METAL=ON ..
-    cmake --build . --config Release
-    ```
-
-When built with Metal support, you can enable GPU inference with the `--gpu-layers|-ngl` command-line argument.
-Any value larger than 0 will offload the computation to the GPU. For example:
-
-```bash
-./main -m ./models/7B/ggml-model-q4_0.gguf -n 128 -ngl 1
-```
+When built with Metal support, you can explicitly disable GPU inference with the `--gpu-layers|-ngl 0` command-line
+argument.
 
 ### MPI Build
 
@@ -399,7 +378,7 @@ Building the program with BLAS support may lead to some performance improvements
 
 - #### cuBLAS
 
-  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -418,13 +397,43 @@ Building the program with BLAS support may lead to some performance improvements
 <!---
   | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
 --->
+  | Option                         | Legal values           | Default | Description |
+  |--------------------------------|------------------------|---------|-------------|
+  | LLAMA_CUDA_FORCE_DMMV          | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
+  | LLAMA_CUDA_F16                 | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       |     128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
+
+- #### hipBLAS
+
+  This provides BLAS acceleration on HIP-supported AMD GPUs.
+  Make sure to have ROCm installed.
+  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
+  Windows support is coming soon...
+
+  - Using `make`:
+    ```bash
+    make LLAMA_HIPBLAS=1
+    ```
+  - Using `CMake`:
+    ```bash
+    mkdir build
+    cd build
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
+    cmake --build .
+    ```
+
+  The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
+  If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
+  The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
-  | LLAMA_CUDA_F16          | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 
 - #### CLBlast
 
@@ -433,6 +442,8 @@ Building the program with BLAS support may lead to some performance improvements
   You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
     - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
 
+    - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
+
     - <details>
         <summary>Installing the OpenCL SDK from source</summary>
 
@@ -450,10 +461,27 @@ Building the program with BLAS support may lead to some performance improvements
         ```
       </details>
 
-  Installing CLBlast: it may be found in your operating system's packages.
+  ##### Installing CLBlast
+
+  Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
+
+  Alternatively, they may be built from source.
 
   - <details>
-    <summary>If not, then installing from source:</summary>
+    <summary>Windows:</summary>
+
+      ```cmd
+      set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64"
+      git clone https://github.com/CNugteren/CLBlast.git
+      mkdir CLBlast\build
+      cd CLBlast\build
+      cmake .. -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64
+      cmake --build . --config Release
+      cmake --install . --prefix C:/CLBlast
+      ```
+
+  - <details>
+    <summary>Unix:</summary>
 
       ```sh
       git clone https://github.com/CNugteren/CLBlast.git
@@ -467,21 +495,32 @@ Building the program with BLAS support may lead to some performance improvements
       Where `/some/path` is where the built library will be installed (default is `/usr/local`).
     </details>
 
-  Building:
+  ##### Building Llama with CLBlast
 
   - Build with make:
     ```sh
     make LLAMA_CLBLAST=1
     ```
-  - CMake:
+  - CMake (Unix):
     ```sh
     mkdir build
     cd build
-    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path
+    cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
     cmake --build . --config Release
     ```
+  - CMake (Windows):
+    ```cmd
+    set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast"
+    git clone https://github.com/ggerganov/llama.cpp
+    cd llama.cpp
+    mkdir build
+    cd build
+    cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64
+    cmake --build . --config Release
+    cmake --install . --prefix C:/LlamaCPP
+    ```
 
-  Running:
+  ##### Running Llama with CLBlast
 
   The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does.
 
@@ -522,6 +561,10 @@ python3 convert.py models/7B/
 # quantize the model to 4-bits (using q4_0 method)
 ./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0
 
+# update the gguf filetype to current if older version is unsupported by another application
+./quantize ./models/7B/ggml-model-q4_0.gguf ./models/7B/ggml-model-q4_0-v2.gguf COPY
+
+
 # run the inference
 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 ```
@@ -543,6 +586,8 @@ As the models are currently fully loaded into memory, you will need adequate dis
 
 Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
 
+*(outdated)*
+
 | Model | Measure      | F16    | Q4_0   | Q4_1   | Q5_0   | Q5_1   | Q8_0   |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
 |    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
@@ -556,6 +601,11 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
+- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- recent k-quants improvements
+  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+
 ### Perplexity (measuring model quality)
 
 You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
@@ -564,6 +614,18 @@ For more information, see [https://huggingface.co/docs/transformers/perplexity](
 The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
 The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
 
+#### How to run
+
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
+3. Output:
+```
+perplexity : calculating perplexity over 655 chunks
+24.43 seconds per pass - ETA 4.45 hours
+[1]4.5970,[2]5.1807,[3]6.0382,...
+```
+And after 4.45 hours, you will have the final perplexity.
+
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
@@ -615,6 +677,8 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
 
 The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
 
+For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
+
 ### Instruction mode with Alpaca
 
 1. First, download the `ggml` Alpaca model into the `./models` folder
@@ -691,14 +755,12 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
 
 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
 - Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
-  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGML)
-  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGML)
-  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGML)
-  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML)
-  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML)
-  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML)
-- Specify `-eps 1e-5` for best generation quality
-- Specify `-gqa 8` for 70B models to work
+  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
+  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
+  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
+  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
+  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
+  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
 
 ### Verifying the model files
 
@@ -726,18 +788,6 @@ If your issue is with model generation quality, then please at least scan the fo
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
 
-#### How to run
-
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
-2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
-3. Output:
-```
-perplexity : calculating perplexity over 655 chunks
-24.43 seconds per pass - ETA 4.45 hours
-[1]4.5970,[2]5.1807,[3]6.0382,...
-```
-And after 4.45 hours, you will have the final perplexity.
-
 ### Android
 
 #### Building the Project using Android NDK
@@ -812,8 +862,17 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 #### Images
 We have two Docker images available for this project:
 
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file.
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
 #### Usage
 
diff --git a/build.zig b/build.zig
index f2769ba8c..b95491e03 100644
--- a/build.zig
+++ b/build.zig
@@ -36,17 +36,20 @@ const Maker = struct {
     }
 
     fn init(builder: *std.build.Builder) !Maker {
-        const commit_hash = @embedFile(".git/refs/heads/master");
+        // const commit_hash = @embedFile(".git/refs/heads/master");
+        const target = builder.standardTargetOptions(.{});
         const config_header = builder.addConfigHeader(
             .{ .style = .blank, .include_path = "build-info.h" },
             .{
                 .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
+                .BUILD_COMMIT = "12345", // omit newline
+                .BUILD_COMPILER = "Zig 0.11.0",
+                .BUILD_TARGET = try target.allocDescription(builder.allocator),
             },
         );
         var m = Maker{
             .builder = builder,
-            .target = builder.standardTargetOptions(.{}),
+            .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
             .config_header = config_header,
             .enable_lto = false,
@@ -58,7 +61,7 @@ const Maker = struct {
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"examples"});
+        try m.addProjectInclude(&.{"common"});
         return m;
     }
 
@@ -71,6 +74,7 @@ const Maker = struct {
             o.addCSourceFiles(&.{src}, m.cxxflags.items);
             o.linkLibCpp();
         }
+        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -104,15 +108,17 @@ pub fn build(b: *std.build.Builder) !void {
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const llama = make.obj("llama", "llama.cpp");
-    const common = make.obj("common", "examples/common.cpp");
-    const console = make.obj("common", "examples/console.cpp");
-    const grammar_parser = make.obj("grammar-parser", "examples/grammar-parser.cpp");
+    const common = make.obj("common", "common/common.cpp");
+    const console = make.obj("common", "common/console.cpp");
+    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const train = make.obj("train", "common/train.cpp");
 
     _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, llama, common, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, llama, common });
     _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, llama, common });
     _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, llama, common });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, llama, common, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, llama, common, train });
 
     const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, llama, common, grammar_parser });
     if (server.target.isWindows()) {
diff --git a/ci/run.sh b/ci/run.sh
index 54ba6d710..942b2e00c 100755
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -196,17 +196,17 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     function check_ppl {
         qnt="$1"
@@ -233,6 +233,48 @@ function gg_run_open_llama_3b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/3B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+
     set +e
 }
 
@@ -242,6 +284,7 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf 'OpenLLaMA 3B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -253,6 +296,11 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # open_llama_7b_v2
@@ -310,17 +358,17 @@ function gg_run_open_llama_7b_v2 {
     ./bin/quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/main --model ${model_f16}  -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -359,6 +407,48 @@ function gg_run_open_llama_7b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
+    # lora
+    function compare_ppl {
+        qnt="$1"
+        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
+            return 20
+        fi
+
+        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
+        return 0
+    }
+
+    path_lora="../models-mnt/open-llama/7B-v2/lora"
+    path_shakespeare="../models-mnt/shakespeare"
+
+    shakespeare="${path_shakespeare}/shakespeare.txt"
+    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
+
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
+    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
+    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
+
+    python3 ../convert-lora-to-ggml.py ${path_lora}
+
+    # f16
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # currently not supported by the CUDA backend
+    # q8_0
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
+    # q8_0 + f16 lora-base
+    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
+
     set +e
 }
 
@@ -368,6 +458,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -379,6 +470,11 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
+    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
+    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
+    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 ## main
@@ -391,6 +487,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
     python3 -m pip install -r ${SRC}/requirements.txt
+    python3 -m pip install --editable gguf-py
 fi
 
 ret=0
diff --git a/codecov.yml b/codecov.yml
new file mode 100644
index 000000000..a301c5b2c
--- /dev/null
+++ b/codecov.yml
@@ -0,0 +1,14 @@
+comment: off
+
+coverage:
+  status:
+    project:
+      default:
+        target: auto
+        threshold: 0
+        base: auto
+    patch:
+      default:
+        target: auto
+        threshold: 0
+        base: auto
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index dead56118..951aa8340 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -9,6 +9,8 @@ add_library(${TARGET} OBJECT
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    train.h
+    train.cpp
     )
 
 if (BUILD_SHARED_LIBS)
diff --git a/common/common.cpp b/common/common.cpp
index 88a962ae3..0f55c33a7 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -1,15 +1,21 @@
 #include "common.h"
+#include "build-info.h"
+#include "llama.h"
 
-#include <cassert>
-#include <iostream>
-#include <cstring>
-#include <fstream>
-#include <string>
-#include <iterator>
 #include <algorithm>
-#include <sstream>
-#include <unordered_set>
+#include <cassert>
+#include <cmath>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iterator>
+#include <iostream>
 #include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_set>
+#include <vector>
+#include <cinttypes>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -18,12 +24,17 @@
 
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
-#define NOMINMAX
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <codecvt>
+#include <locale>
 #include <windows.h>
 #include <fcntl.h>
 #include <io.h>
 #else
 #include <sys/ioctl.h>
+#include <sys/stat.h>
 #include <unistd.h>
 #endif
 
@@ -46,7 +57,7 @@ int32_t get_num_physical_cores() {
             siblings.insert(line);
         }
     }
-    if (siblings.size() > 0) {
+    if (!siblings.empty()) {
         return static_cast<int32_t>(siblings.size());
     }
 #elif defined(__APPLE__) && defined(__MACH__)
@@ -93,7 +104,6 @@ void process_escapes(std::string& input) {
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
-    bool escape_prompt = false;
     std::string arg;
     gpt_params default_params;
     const std::string arg_prefix = "--";
@@ -119,14 +129,23 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads <= 0) {
                 params.n_threads = std::thread::hardware_concurrency();
             }
+        } else if (arg == "-tb" || arg == "--threads-batch") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+            if (params.n_threads_batch <= 0) {
+                params.n_threads_batch = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.prompt = argv[i];
-        } else if (arg == "-e") {
-            escape_prompt = true;
+        } else if (arg == "-e" || arg == "--escape") {
+            params.escape = true;
         } else if (arg == "--prompt-cache") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -148,8 +167,10 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+            // store the external file name in params
+            params.prompt_file = argv[i];
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
-            if (params.prompt.back() == '\n') {
+            if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
         } else if (arg == "-n" || arg == "--n-predict") {
@@ -274,7 +295,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.cfg_negative_prompt));
-            if (params.cfg_negative_prompt.back() == '\n') {
+            if (!params.cfg_negative_prompt.empty() && params.cfg_negative_prompt.back() == '\n') {
                 params.cfg_negative_prompt.pop_back();
             }
         } else if (arg == "--cfg-scale") {
@@ -295,18 +316,42 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_keep = std::stoi(argv[i]);
+        } else if (arg == "--draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_draft = std::stoi(argv[i]);
         } else if (arg == "--chunks") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.n_chunks = std::stoi(argv[i]);
+        } else if (arg == "-np" || arg == "--parallel") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_parallel = std::stoi(argv[i]);
+        } else if (arg == "-ns" || arg == "--sequences") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_sequences = std::stoi(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "-md" || arg == "--model-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_draft = argv[i];
         } else if (arg == "-a" || arg == "--alias") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -318,7 +363,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f));
+            params.use_mmap = false;
+        } else if (arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i])));
             params.use_mmap = false;
         } else if (arg == "--lora-base") {
             if (++i >= argc) {
@@ -334,10 +391,14 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
+        } else if (arg == "--infill") {
+            params.infill = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
             params.simple_io = true;
+        } else if (arg == "-cb" || arg == "--cont-batching") {
+            params.cont_batching = true;
         } else if (arg == "--color") {
             params.use_color = true;
         } else if (arg == "--mlock") {
@@ -352,6 +413,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+#endif
+        } else if (arg == "--gpu-layers-draft" || arg == "-ngld" || arg == "--n-gpu-layers-draft") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
+            params.n_gpu_layers_draft = std::stoi(argv[i]);
+#else
+            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
+            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
         } else if (arg == "--main-gpu" || arg == "-mg") {
             if (++i >= argc) {
@@ -392,21 +464,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.mul_mat_q = false;
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
-#endif // GGML_USE_CUBLAS
-        } else if (arg == "--low-vram" || arg == "-lv") {
-#ifdef GGML_USE_CUBLAS
-            params.low_vram = true;
-#else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
-        } else if (arg == "--mtest") {
-            params.mem_test = true;
         } else if (arg == "--numa") {
             params.numa = true;
-        } else if (arg == "--export") {
-            params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
@@ -415,8 +477,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.antiprompt.push_back(argv[i]);
-        } else if (arg == "--perplexity") {
-            params.perplexity = true;
+        } else if (arg == "-ld" || arg == "--logdir") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.logdir = argv[i];
+
+            if (params.logdir.back() != DIRECTORY_SEPARATOR) {
+                params.logdir += DIRECTORY_SEPARATOR;
+            }
+        } else if (arg == "--perplexity" || arg == "--all-logits") {
+            params.logits_all = true;
         } else if (arg == "--ppl-stride") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -462,6 +534,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, default_params);
+#ifndef LOG_DISABLE_LOGS
+            log_print_usage();
+#endif // LOG_DISABLE_LOGS
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -501,6 +576,25 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(params.grammar)
             );
+#ifndef LOG_DISABLE_LOGS
+        // Parse args for logging parameters
+        } else if ( log_param_single_parse( argv[i] ) ) {
+            // Do nothing, log_param_single_parse automatically does it's thing
+            //  and returns if a match was found and parsed.
+        } else if ( log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i] ) ) {
+            // We have a matching known parameter requiring an argument,
+            //  now we need to check if there is anything after this argv
+            //  and flag invalid_param or parse it.
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            if( !log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i-1], argv[i]) ) {
+                invalid_param = true;
+                break;
+            }
+        // End of Parse args for logging parameters
+#endif // LOG_DISABLE_LOGS
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
@@ -520,112 +614,139 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
         exit(1);
     }
 
-    if (escape_prompt) {
+    if (params.escape) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
         process_escapes(params.input_suffix);
+        for (auto & antiprompt : params.antiprompt) {
+            process_escapes(antiprompt);
+        }
     }
 
     return true;
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -i, --interactive     run in interactive mode\n");
-    fprintf(stdout, "  --interactive-first   run in interactive mode and wait for input right away\n");
-    fprintf(stdout, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    fprintf(stdout, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    fprintf(stdout, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stdout, "                        halt generation at PROMPT, return control in interactive mode\n");
-    fprintf(stdout, "                        (can be specified more than once for multiple prompts).\n");
-    fprintf(stdout, "  --color               colorise output to distinguish prompt and user input from generations\n");
-    fprintf(stdout, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -p PROMPT, --prompt PROMPT\n");
-    fprintf(stdout, "                        prompt to start generation with (default: empty)\n");
-    fprintf(stdout, "  -e                    process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stdout, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    fprintf(stdout, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    fprintf(stdout, "                        not supported with --interactive or other interactive options\n");
-    fprintf(stdout, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    fprintf(stdout, "  --random-prompt       start with a randomized prompt.\n");
-    fprintf(stdout, "  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
-    fprintf(stdout, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    fprintf(stdout, "  -f FNAME, --file FNAME\n");
-    fprintf(stdout, "                        prompt file to start generation.\n");
-    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    fprintf(stdout, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    fprintf(stdout, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    fprintf(stdout, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    fprintf(stdout, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    fprintf(stdout, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    fprintf(stdout, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    fprintf(stdout, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
-    fprintf(stdout, "  --mirostat N          use Mirostat sampling.\n");
-    fprintf(stdout, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    fprintf(stdout, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    fprintf(stdout, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    fprintf(stdout, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
-    fprintf(stdout, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    fprintf(stdout, "                        modifies the likelihood of token appearing in the completion,\n");
-    fprintf(stdout, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
-    fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
-    fprintf(stdout, "  --cfg-negative-prompt PROMPT\n");
-    fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-negative-prompt-file FNAME\n");
-    fprintf(stdout, "                        negative prompt file to use for guidance. (default: empty)\n");
-    fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stdout, "  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale (default: %g)\n", 1.0f/params.rope_freq_scale);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency linear scaling factor, inverse of --rope-scale (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
-    fprintf(stdout, "  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
-    fprintf(stdout, "  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
-    fprintf(stdout, "  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
-    fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --interactive-first   run in interactive mode and wait for input right away\n");
+    printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    printf("  -r PROMPT, --reverse-prompt PROMPT\n");
+    printf("                        halt generation at PROMPT, return control in interactive mode\n");
+    printf("                        (can be specified more than once for multiple prompts).\n");
+    printf("  --color               colorise output to distinguish prompt and user input from generations\n");
+    printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N\n");
+    printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
+    printf("  -p PROMPT, --prompt PROMPT\n");
+    printf("                        prompt to start generation with (default: empty)\n");
+    printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    printf("  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    printf("  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    printf("                        not supported with --interactive or other interactive options\n");
+    printf("  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    printf("  --random-prompt       start with a randomized prompt.\n");
+    printf("  --in-prefix-bos       prefix BOS to user inputs, preceding the `--in-prefix` string\n");
+    printf("  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    printf("  -f FNAME, --file FNAME\n");
+    printf("                        prompt file to start generation.\n");
+    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    printf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    printf("  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    printf("  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    printf("  --mirostat N          use Mirostat sampling.\n");
+    printf("                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    printf("                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    printf("  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    printf("  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    printf("  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    printf("                        modifies the likelihood of token appearing in the completion,\n");
+    printf("                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
+    printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  --cfg-negative-prompt PROMPT\n");
+    printf("                        negative prompt to use for guidance. (default: empty)\n");
+    printf("  --cfg-negative-prompt-file FNAME\n");
+    printf("                        negative prompt file to use for guidance. (default: empty)\n");
+    printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
+    printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
+    printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
+    printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
+    printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
+    printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
+    printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
+    printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     if (llama_mlock_supported()) {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported()) {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
-    fprintf(stdout, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    fprintf(stdout, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ngld N, --n-gpu-layers-draft N\n");
+    printf("                        number of layers to store in VRAM for the draft model\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+#ifdef GGML_USE_CUBLAS
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif // GGML_USE_CUBLAS
 #endif
-    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
-    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
-    fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
+    printf("  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -md FNAME, --model-draft FNAME\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("  -ld LOGDIR, --logdir LOGDIR\n");
+    printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("\n");
+}
+
+std::string get_system_info(const gpt_params & params) {
+    std::ostringstream os;
+
+    os << "system_info: n_threads = " << params.n_threads;
+    if (params.n_threads_batch != -1) {
+        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+    }
+    os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+
+    return os.str();
 }
 
 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -641,58 +762,74 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
         case 7: return "He";
         case 8: return "She";
         case 9: return "They";
-        default: return "To";
     }
 
-    return "The";
+    GGML_UNREACHABLE();
 }
 
 //
 // Model utils
 //
 
+struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+    auto mparams = llama_model_default_params();
+
+    if (params.n_gpu_layers != -1) {
+        mparams.n_gpu_layers = params.n_gpu_layers;
+    }
+    mparams.main_gpu        = params.main_gpu;
+    mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_mlock       = params.use_mlock;
+
+    return mparams;
+}
+
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
-    auto lparams = llama_context_default_params();
+    auto cparams = llama_context_default_params();
 
-    lparams.n_ctx           = params.n_ctx;
-    lparams.n_batch         = params.n_batch;
-    lparams.n_gpu_layers    = params.n_gpu_layers;
-    lparams.main_gpu        = params.main_gpu;
-    lparams.tensor_split    = params.tensor_split;
-    lparams.low_vram        = params.low_vram;
-    lparams.mul_mat_q       = params.mul_mat_q;
-    lparams.seed            = params.seed;
-    lparams.f16_kv          = params.memory_f16;
-    lparams.use_mmap        = params.use_mmap;
-    lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
-    lparams.embedding       = params.embedding;
-    lparams.rope_freq_base  = params.rope_freq_base;
-    lparams.rope_freq_scale = params.rope_freq_scale;
+    cparams.n_ctx           = params.n_ctx;
+    cparams.n_batch         = params.n_batch;
+    cparams.n_threads       = params.n_threads;
+    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q       = params.mul_mat_q;
+    cparams.seed            = params.seed;
+    cparams.f16_kv          = params.memory_f16;
+    cparams.logits_all      = params.logits_all;
+    cparams.embedding       = params.embedding;
+    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale;
 
-    return lparams;
+    return cparams;
 }
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
-    auto lparams = llama_context_params_from_gpt_params(params);
+    auto mparams = llama_model_params_from_gpt_params(params);
 
-    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), mparams);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return std::make_tuple(nullptr, nullptr);
     }
 
-    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    auto cparams = llama_context_params_from_gpt_params(params);
+
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
         return std::make_tuple(nullptr, nullptr);
     }
 
-    if (!params.lora_adapter.empty()) {
+    for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+        const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
+        float lora_scale = std::get<1>(params.lora_adapter[i]);
         int err = llama_model_apply_lora_from_file(model,
-                                             params.lora_adapter.c_str(),
-                                             params.lora_base.empty() ? NULL : params.lora_base.c_str(),
+                                             lora_adapter.c_str(),
+                                             lora_scale,
+                                             ((i > 0) || params.lora_base.empty())
+                                                ? NULL
+                                                : params.lora_base.c_str(),
                                              params.n_threads);
         if (err != 0) {
             fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
@@ -706,6 +843,15 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
     }
 
+    {
+        LOG("warming up the model with an empty run\n");
+
+        std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
+        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
+        llama_reset_timings(lctx);
+    }
+
     return std::make_tuple(model, lctx);
 }
 
@@ -714,16 +860,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 //
 
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
+           const std::string & text,
+                        bool   add_bos) {
+    return llama_tokenize(llama_get_model(ctx), text, add_bos);
+}
+
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_bos) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + add_bos;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(ctx, text.c_str(), result.data(), result.size(), add_bos);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -731,12 +884,12 @@ std::vector<llama_token> llama_tokenize(
     return result;
 }
 
-std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_str(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -745,34 +898,459 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok
     return std::string(result.data(), result.size());
 }
 
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
-           const std::string & text,
-                        bool   add_bos) {
-    int n_tokens = text.length() + add_bos;
-    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
+std::string llama_detokenize_spm(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    const llama_token bos_id = llama_token_bos(ctx);
+
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        // remove the leading space of the first non-BOS token
+        if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') {
+            piece = piece.substr(1);
+        }
+
+        result += piece;
     }
+
     return result;
 }
 
-std::string llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token) {
-    std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        const int check = llama_token_to_str_bpe(ctx, token, result.data(), result.size());
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
+std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_token> & tokens) {
+    std::string piece;
+    std::string result;
+
+    for (size_t i = 0; i < tokens.size(); ++i) {
+        piece = llama_token_to_piece(ctx, tokens[i]);
+
+        result += piece;
     }
 
-    return std::string(result.data(), result.size());
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return result;
 }
 
+//
+// Sampling utils
+//
+
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx) {
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+    const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
+    const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
+    const float   repeat_penalty  = params.repeat_penalty;
+    const float   alpha_presence  = params.presence_penalty;
+    const float   alpha_frequency = params.frequency_penalty;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+    const bool    penalize_nl     = params.penalize_nl;
+
+    llama_token id = 0;
+
+    float * logits = llama_get_logits_ith(ctx, idx);
+
+    // Apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    candidates.clear();
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+    if (ctx_guidance) {
+        llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
+    }
+
+    // apply penalties
+    if (!last_tokens.empty()) {
+        const float nl_logit = logits[llama_token_nl(ctx)];
+        const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
+
+        llama_sample_repetition_penalty(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, repeat_penalty);
+        llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
+                last_tokens.data() + last_tokens.size() - last_n_repeat,
+                last_n_repeat, alpha_frequency, alpha_presence);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(ctx)) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
+                }
+            }
+        }
+    }
+
+    if (grammar != NULL) {
+        llama_sample_grammar(ctx, &cur_p, grammar);
+    }
+
+    if (temp <= 0) {
+        // Greedy sampling
+        id = llama_sample_token_greedy(ctx, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
+        } else if (mirostat == 2) {
+            static float mirostat_mu = 2.0f * mirostat_tau;
+            llama_sample_temp(ctx, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu);
+        } else {
+            // Temperature sampling
+            size_t min_keep = std::max(1, params.n_probs);
+            llama_sample_top_k      (ctx, &cur_p, top_k, min_keep);
+            llama_sample_tail_free  (ctx, &cur_p, tfs_z, min_keep);
+            llama_sample_typical    (ctx, &cur_p, typical_p, min_keep);
+            llama_sample_top_p      (ctx, &cur_p, top_p, min_keep);
+            llama_sample_temp(ctx, &cur_p, temp);
+
+            {
+                const int n_top = 10;
+                LOG("top %d candidates:\n", n_top);
+
+                for (int i = 0; i < n_top; i++) {
+                    const llama_token id = cur_p.data[i].id;
+                    LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
+                }
+            }
+
+            id = llama_sample_token(ctx, &cur_p);
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
+        }
+    }
+    // printf("`%d`", candidates_p.size);
+
+    if (grammar != NULL) {
+        llama_grammar_accept_token(ctx, grammar, id);
+    }
+
+    return id;
+}
+
+//
+// YAML utils
+//
+
+// returns true if successful, false otherwise
+bool create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%e, ", data[i]);
+    }
+    fprintf(stream, "%e]\n", data.back());
+}
+
+void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%d, ", data[i]);
+    }
+    fprintf(stream, "%d]\n", data.back());
+}
+
+void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data) {
+    std::string data_str(data == NULL ? "" : data);
+
+    if (data_str.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    size_t pos_start = 0;
+    size_t pos_found = 0;
+
+    if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) {
+        data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
+        data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
+        data_str = "\"" + data_str + "\"";
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    if (data_str.find('\n') == std::string::npos) {
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    fprintf(stream, "%s: |\n", prop_name);
+    while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
+        fprintf(stream, "  %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
+        pos_start = pos_found + 1;
+    }
+}
+
+std::string get_sortable_timestamp() {
+    using clock = std::chrono::system_clock;
+
+    const clock::time_point current_time = clock::now();
+    const time_t as_time_t = clock::to_time_t(current_time);
+    char timestamp_no_ns[100];
+    std::strftime(timestamp_no_ns, 100, "%Y_%m_%d-%H_%M_%S", std::localtime(&as_time_t));
+
+    const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+        current_time.time_since_epoch() % 1000000000).count();
+    char timestamp_ns[11];
+    snprintf(timestamp_ns, 11, "%09" PRId64, ns);
+
+    return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
+}
+
+void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
+                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
+    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
+    fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
+
+#ifdef NDEBUG
+    fprintf(stream, "debug: false\n");
+#else
+    fprintf(stream, "debug: true\n");
+#endif // NDEBUG
+
+    fprintf(stream, "model_desc: %s\n", model_desc);
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
+
+#ifdef __OPTIMIZE__
+    fprintf(stream, "optimize: true\n");
+#else
+    fprintf(stream, "optimize: false\n");
+#endif // __OPTIMIZE__
+
+    fprintf(stream, "time: %s\n", timestamp.c_str());
+
+    fprintf(stream, "\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "# User Inputs #\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
+    fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
+    dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str());
+    fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale);
+    fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
+    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
+    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
+    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
+    dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
+    fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
+    fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
+    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
+
+    const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
+    const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
+    fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
+
+    dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str());
+    fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
+    dump_string_yaml_multiline(stream, "in_suffix", params.input_prefix.c_str());
+    fprintf(stream, "instruct: %s # default: false\n", params.instruct ? "true" : "false");
+    fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
+    fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
+    fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
+    fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
+
+    fprintf(stream, "logit_bias:\n");
+    for (std::pair<llama_token, float> lb : params.logit_bias) {
+        if (ignore_eos && lb.first == logit_bias_eos->first) {
+            continue;
+        }
+        fprintf(stream, "  %d: %f", lb.first, lb.second);
+    }
+
+    fprintf(stream, "lora:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) != 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s\n", std::get<0>(la).c_str());
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (std::tuple<std::string, float> la : params.lora_adapter) {
+        if (std::get<1>(la) == 1.0f) {
+            continue;
+        }
+        fprintf(stream, "  - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
+    }
+    fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
+    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
+    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
+    fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
+    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
+    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
+    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs);
+    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
+    fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false");
+    fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
+    fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
+    fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty);
+    dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
+    fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
+    fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
+    fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
+    dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
+    fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty);
+
+    fprintf(stream, "reverse_prompt:\n");
+    for (std::string ap : params.antiprompt) {
+        size_t pos = 0;
+        while ((pos = ap.find('\n', pos)) != std::string::npos) {
+            ap.replace(pos, 1, "\\n");
+            pos += 1;
+        }
+
+        fprintf(stream, "  - %s\n", ap.c_str());
+    }
+
+    fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
+    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
+    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
+    fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
+
+    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
+    dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
+
+    fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z);
+    fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "top_k: %d # default: 40\n", params.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p);
+    fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p);
+    fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+}
diff --git a/common/common.h b/common/common.h
index d68a8ef88..c80215279 100644
--- a/common/common.h
+++ b/common/common.h
@@ -4,6 +4,9 @@
 
 #include "llama.h"
 
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
 #include <string>
 #include <vector>
 #include <random>
@@ -11,6 +14,20 @@
 #include <unordered_map>
 #include <tuple>
 
+#ifdef _WIN32
+#define DIRECTORY_SEPARATOR '\\'
+#else
+#define DIRECTORY_SEPARATOR '/'
+#endif // _WIN32
+
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define print_build_info() do {                                                             \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+} while(0)
+
 //
 // CLI argument parsing
 //
@@ -19,17 +36,23 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_predict                       = -1;   // new tokens to predict
     int32_t n_ctx                           = 512;  // context size
     int32_t n_batch                         = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;    // number of tokens to keep from initial prompt
+    int32_t n_draft                         = 16;   // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;   // max number of chunks to process (-1 = unlimited)
-    int32_t n_gpu_layers                    = 0;    // number of layers to store in VRAM
+    int32_t n_parallel                      = 1;    // number of parallel sequences to decode
+    int32_t n_sequences                     = 1;    // number of sequences to decode
+    int32_t n_gpu_layers                    = -1;   // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft              = -1;   // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
     int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
-    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
 
     // sampling parameters
     int32_t top_k             = 40;    // <= 0 to use vocab size
@@ -53,16 +76,19 @@ struct gpt_params {
     float       cfg_scale         = 1.f;   // How strong is guidance
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft       = "";                              // draft model for speculative decoding
     std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
+    std::string prompt_file       = "";  // store the external prompt file name
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
     std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
+    std::string logdir            = "";  // directory in which to save YAML log files
 
-    std::string lora_adapter = "";  // lora adapter path
-    std::string lora_base    = "";  // base model path for the lora adapter
+    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    std::string lora_base  = "";                              // base model path for the lora adapter
 
     int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -71,7 +97,6 @@ struct gpt_params {
     bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
@@ -81,54 +106,120 @@ struct gpt_params {
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
     bool embedding         = false; // get only sentence embedding
+    bool escape            = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
+    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
-    bool perplexity        = false; // compute perplexity over the prompt
+    bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool mem_test          = false; // compute maximum memory usage
     bool numa              = false; // attempt optimizations that help on some NUMA systems
-    bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool infill            = false; // use infill mode
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
 
+std::string get_system_info(const gpt_params & params);
+
 std::string gpt_random_prompt(std::mt19937 & rng);
 
+void process_escapes(std::string& input);
+
 //
 // Model utils
 //
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
 //
 // Vocab utils
 //
 
+// tokenizes a string into a vector of tokens
+// should work similar to Python's `tokenizer.encode`
 std::vector<llama_token> llama_tokenize(
-        struct llama_context * ctx,
+  const struct llama_context * ctx,
            const std::string & text,
                         bool   add_bos);
 
-std::vector<llama_token> llama_tokenize_bpe(
-        struct llama_context * ctx,
+std::vector<llama_token> llama_tokenize(
+    const struct llama_model * model,
            const std::string & text,
                         bool   add_bos);
 
-std::string llama_token_to_str(
+// tokenizes a token into a piece
+// should work similar to Python's `tokenizer.id_to_piece`
+std::string llama_token_to_piece(
         const struct llama_context * ctx,
                        llama_token   token);
 
-std::string llama_token_to_str_bpe(
-    const struct llama_context * ctx,
-                   llama_token   token);
+// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
+//       that takes into account the tokenizer type and decides how to handle the leading space
+//
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+// removes the leading space from the first non-BOS token
+std::string llama_detokenize_spm(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+std::string llama_detokenize_bpe(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+//
+// Sampling utils
+//
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+//
+// required:
+//  - ctx:    context to use for sampling
+//  - params: sampling parameters
+//
+// optional:
+//  - ctx_guidance:  context to use for classifier-free guidance, ignore if NULL
+//  - grammar:       grammar to use for sampling, ignore if NULL
+//  - last_tokens:   needed for repetition penalty, ignore if empty
+//  - idx:           sample from llama_get_logits_ith(ctx, idx)
+//
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
+//
+llama_token llama_sample_token(
+                  struct llama_context * ctx,
+                  struct llama_context * ctx_guidance,
+                  struct llama_grammar * grammar,
+               const struct gpt_params & params,
+        const std::vector<llama_token> & last_tokens,
+         std::vector<llama_token_data> & candidates,
+                                   int   idx = 0);
+
+//
+// YAML utils
+//
+
+bool create_directory_with_parents(const std::string & path);
+void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
+void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
+void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
+std::string get_sortable_timestamp();
+
+void dump_non_result_info_yaml(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
diff --git a/common/console.cpp b/common/console.cpp
index 8efa2a674..f65cbc6ed 100644
--- a/common/console.cpp
+++ b/common/console.cpp
@@ -158,7 +158,7 @@ namespace console {
         }
     }
 
-    char32_t getchar32() {
+    static char32_t getchar32() {
 #if defined(_WIN32)
         HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
         wchar_t high_surrogate = 0;
@@ -212,7 +212,7 @@ namespace console {
 #endif
     }
 
-    void pop_cursor() {
+    static void pop_cursor() {
 #if defined(_WIN32)
         if (hConsole != NULL) {
             CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
@@ -233,15 +233,16 @@ namespace console {
         putc('\b', out);
     }
 
-    int estimateWidth(char32_t codepoint) {
+    static int estimateWidth(char32_t codepoint) {
 #if defined(_WIN32)
+        (void)codepoint;
         return 1;
 #else
         return wcwidth(codepoint);
 #endif
     }
 
-    int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
+    static int put_codepoint(const char* utf8_codepoint, size_t length, int expectedWidth) {
 #if defined(_WIN32)
         CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
         if (!GetConsoleScreenBufferInfo(hConsole, &bufferInfo)) {
@@ -302,7 +303,7 @@ namespace console {
 #endif
     }
 
-    void replace_last(char ch) {
+    static void replace_last(char ch) {
 #if defined(_WIN32)
         pop_cursor();
         put_codepoint(&ch, 1, 1);
@@ -311,7 +312,7 @@ namespace console {
 #endif
     }
 
-    void append_utf8(char32_t ch, std::string & out) {
+    static void append_utf8(char32_t ch, std::string & out) {
         if (ch <= 0x7F) {
             out.push_back(static_cast<unsigned char>(ch));
         } else if (ch <= 0x7FF) {
@@ -332,7 +333,7 @@ namespace console {
     }
 
     // Helper function to remove the last UTF-8 character from a string
-    void pop_back_utf8_char(std::string & line) {
+    static void pop_back_utf8_char(std::string & line) {
         if (line.empty()) {
             return;
         }
@@ -348,7 +349,7 @@ namespace console {
         line.erase(pos);
     }
 
-    bool readline_advanced(std::string & line, bool multiline_input) {
+    static bool readline_advanced(std::string & line, bool multiline_input) {
         if (out != stdout) {
             fflush(stdout);
         }
@@ -451,7 +452,7 @@ namespace console {
         return has_more;
     }
 
-    bool readline_simple(std::string & line, bool multiline_input) {
+    static bool readline_simple(std::string & line, bool multiline_input) {
 #if defined(_WIN32)
         std::wstring wline;
         if (!std::getline(std::wcin, wline)) {
diff --git a/common/grammar-parser.cpp b/common/grammar-parser.cpp
index e76bd11c3..5a545a807 100644
--- a/common/grammar-parser.cpp
+++ b/common/grammar-parser.cpp
@@ -9,7 +9,7 @@
 namespace grammar_parser {
     // NOTE: assumes valid utf8 (but checks for overrun)
     // copied from llama.cpp
-    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
         static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
         uint8_t  first_byte = static_cast<uint8_t>(*src);
         uint8_t  highbits   = first_byte >> 4;
@@ -24,19 +24,19 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
         return result.first->second;
     }
 
-    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
         return next_id;
     }
 
-    void add_rule(
+    static void add_rule(
             parse_state & state,
             uint32_t      rule_id,
             const std::vector<llama_grammar_element> & rule) {
@@ -46,11 +46,11 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
-    bool is_word_char(char c) {
+    static bool is_word_char(char c) {
         return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
     }
 
-    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
         const char * pos   = src;
         const char * end   = src + size;
         uint32_t     value = 0;
@@ -73,7 +73,7 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    const char * parse_space(const char * src, bool newline_ok) {
+    static const char * parse_space(const char * src, bool newline_ok) {
         const char * pos = src;
         while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                 (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@@ -88,7 +88,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_name(const char * src) {
+    static const char * parse_name(const char * src) {
         const char * pos = src;
         while (is_word_char(*pos)) {
             pos++;
@@ -99,7 +99,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    std::pair<uint32_t, const char *> parse_char(const char * src) {
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
                 case 'x': return parse_hex(src + 2, 2);
@@ -129,7 +129,7 @@ namespace grammar_parser {
             uint32_t            rule_id,
             bool                is_nested);
 
-    const char * parse_sequence(
+    static const char * parse_sequence(
             parse_state                        & state,
             const char                         * src,
             const std::string                  & rule_name,
@@ -247,7 +247,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_rule(parse_state & state, const char * src) {
+    static const char * parse_rule(parse_state & state, const char * src) {
         const char * name_end = parse_name(src);
         const char * pos      = parse_space(name_end, false);
         size_t       name_len = name_end - src;
@@ -285,7 +285,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_grammar_char(FILE * file, uint32_t c) {
+    static void print_grammar_char(FILE * file, uint32_t c) {
         if (0x20 <= c && c <= 0x7f) {
             fprintf(file, "%c", static_cast<char>(c));
         } else {
@@ -294,7 +294,7 @@ namespace grammar_parser {
         }
     }
 
-    bool is_char_element(llama_grammar_element elem) {
+    static bool is_char_element(llama_grammar_element elem) {
         switch (elem.type) {
             case LLAMA_GRETYPE_CHAR:           return true;
             case LLAMA_GRETYPE_CHAR_NOT:       return true;
@@ -304,7 +304,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
         for (auto elem : rule) {
             switch (elem.type) {
                 case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
@@ -334,7 +334,7 @@ namespace grammar_parser {
         fprintf(file, "\n");
     }
 
-    void print_rule(
+    static void print_rule(
             FILE     * file,
             uint32_t   rule_id,
             const std::vector<llama_grammar_element> & rule,
@@ -415,6 +415,7 @@ namespace grammar_parser {
 
     std::vector<const llama_grammar_element *> parse_state::c_rules() {
         std::vector<const llama_grammar_element *> ret;
+        ret.reserve(rules.size());
         for (const auto & rule : rules) {
             ret.push_back(rule.data());
         }
diff --git a/common/log.h b/common/log.h
new file mode 100644
index 000000000..b8953fdca
--- /dev/null
+++ b/common/log.h
@@ -0,0 +1,643 @@
+#pragma once
+
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
+
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a diffrent function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_diffrent()
+//  #include "log.h"
+//
+//  FILE* log_handler_diffrent()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
+#endif
+
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
+
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+    static std::string pid;
+    if (pid.empty())
+    {
+        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+        //  it's not the same as "pid" but is unique enough to solve multiple instances
+        //  trying to write to the same log.
+        std::stringstream ss;
+        ss << std::this_thread::get_id();
+        pid = ss.str();
+    }
+
+    return pid;
+}
+
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
+
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
+
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    std::stringstream buf;
+
+    buf << log_file_basename;
+    buf << ".";
+    buf << log_get_pid();
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_IMPL(str, ...)                                                                                      \
+    do {                                                                                                            \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
+    } while (0)
+#else
+    #define LOG_IMPL(str, ...)                                                                                           \
+    do {                                                                                                                 \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    } while (0)
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#ifndef _MSC_VER
+    #define LOG_TEE_IMPL(str, ...)                                                                                                      \
+    do {                                                                                                                                \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    } while (0)
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                           \
+    do {                                                                                                                                     \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    } while (0)
+#endif
+
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
+
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#ifndef _MSC_VER
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#ifndef _MSC_VER
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#ifndef _MSC_VER
+    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
+    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
+#else
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized{false};
+    static bool _disabled{(filename.empty() && target == nullptr)};
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
+    if (_initialized)
+    {
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  untill enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n");
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n");
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n");
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n");
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n");
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n");
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n");
+#ifdef _MSC_VER
+    LOG_TEE("15 Hello msvc TEE without arguments\n");
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n");
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
+    LOG("19 Hello msvc LOG without arguments\n");
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
+    LOGLN("21 Hello msvc LOGLN without arguments\n");
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    printf("log options:\n");
+    /* format
+    printf("  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str());
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens)                                 \
+    [&tokens, &ctx]()                                                        \
+    {                                                                        \
+        std::stringstream buf;                                               \
+        buf << "[ ";                                                         \
+                                                                             \
+        bool first = true;                                                   \
+        for (const auto &token : tokens)                                     \
+        {                                                                    \
+            if (!first)                                                      \
+                buf << ", ";                                                 \
+            else                                                             \
+                first = false;                                               \
+                                                                             \
+            auto detokenized = llama_token_to_piece(ctx, token);             \
+                                                                             \
+            detokenized.erase(                                               \
+                std::remove_if(                                              \
+                    detokenized.begin(),                                     \
+                    detokenized.end(),                                       \
+                    [](const unsigned char c) { return !std::isprint(c); }), \
+                detokenized.end());                                          \
+                                                                             \
+            buf                                                              \
+                << "'" << detokenized << "'"                                 \
+                << ":" << std::to_string(token);                             \
+        }                                                                    \
+        buf << " ]";                                                         \
+                                                                             \
+        return buf.str();                                                    \
+    }()                                                                      \
+        .c_str()
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS
diff --git a/common/train.cpp b/common/train.cpp
new file mode 100644
index 000000000..35a4cf9e6
--- /dev/null
+++ b/common/train.cpp
@@ -0,0 +1,1496 @@
+#include "train.h"
+#include "common.h"
+
+#include <random>
+#include <sstream>
+#include <functional>
+
+struct random_normal_distribution {
+    std::mt19937 gen;
+    std::normal_distribution<float> rd;
+    float min;
+    float max;
+};
+
+struct random_uniform_distribution {
+    std::mt19937 gen;
+    std::uniform_real_distribution<float> rd;
+};
+
+struct train_state  * init_train_state() {
+    struct train_state * state = new struct train_state;
+    state->train_its     = 0;
+    state->train_samples = 0;
+    state->train_tokens  = 0;
+    state->train_epochs  = 0;
+    state->shuffle_samples_hash  = 0;
+    state->shuffle_sample_count  = 0;
+    state->shuffle_next_sample   = 0;
+    state->shuffle_rng_state_current = "";
+    state->shuffle_rng_state_next    = "";
+
+    state->opt = new struct ggml_opt_context;
+    state->opt->ctx = NULL;
+    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    state->opt->loss_after = 0.0f;
+
+    return state;
+}
+
+void free_train_state(struct train_state  * state) {
+    delete state->opt;
+    delete state;
+}
+
+struct random_normal_distribution * init_random_normal_distribution(
+    int seed, float mean, float std, float min, float max
+) {
+    struct random_normal_distribution * rnd = (struct random_normal_distribution *) malloc(sizeof(struct random_normal_distribution));
+    rnd->gen = std::mt19937(seed);
+    rnd->rd = std::normal_distribution<float>{mean, std};
+    rnd->min = min;
+    rnd->max = max;
+    return rnd;
+}
+
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max) {
+    struct random_uniform_distribution * rnd = (struct random_uniform_distribution *) malloc(sizeof(struct random_uniform_distribution));
+    rnd->gen = std::mt19937(seed);
+    rnd->rd = std::uniform_real_distribution<float>{min, max};
+    return rnd;
+}
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd) {
+    free(rnd);
+}
+
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd) {
+    free(rnd);
+}
+
+struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
+    float scale = 1.0f; // xavier
+    switch (tensor->n_dims) {
+        case 1:
+            scale /= sqrtf((float) tensor->ne[0]);
+            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
+                *dst = scale * frand_normal(rnd);
+            }
+            break;
+        case 2:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+                    *dst = scale * frand_normal(rnd);
+                }
+            }
+            break;
+        case 3:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
+                        *dst = scale * frand_normal(rnd);
+                    }
+                }
+            }
+            break;
+        case 4:
+            scale /= sqrtf((float) tensor->ne[0]+tensor->ne[1]);
+            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
+                            *dst = scale * frand_normal(rnd);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            die("Unsupported tensor->n_dims");
+    };
+    return tensor;
+}
+
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
+    switch (tensor->n_dims) {
+        case 1:
+            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
+                *dst = frand_uniform(rnd);
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
+                    *dst = frand_uniform(rnd);
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
+                        *dst = frand_uniform(rnd);
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
+                            *dst = frand_uniform(rnd);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            die("Unsupported tensor->n_dims");
+    };
+    return tensor;
+}
+
+float frand() {
+    return (float)rand()/((float)(RAND_MAX) + 1.0f);
+}
+
+float frand_normal(struct random_normal_distribution * rnd) {
+    return fclamp(rnd->rd(rnd->gen), rnd->min, rnd->max);
+}
+
+float frand_uniform(struct random_uniform_distribution * rnd) {
+    return rnd->rd(rnd->gen);
+}
+
+int clamp(const int v, const int min, const int max) {
+    return ((v < min) ? (min) : (v > max) ? (max) : v);
+}
+
+float fclamp(const float v, const float min, const float max) {
+    return ((v < min) ? (min) : (v > max) ? (max) : v);
+}
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+    GGML_ASSERT(tensor->n_dims == 1);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+}
+
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
+    GGML_ASSERT(tensor->n_dims == 2);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+}
+
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
+    GGML_ASSERT(tensor->n_dims == 3);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == ne2);
+}
+
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+    GGML_ASSERT(tensor->n_dims == 4);
+    GGML_ASSERT(tensor->ne[0] == ne0);
+    GGML_ASSERT(tensor->ne[1] == ne1);
+    GGML_ASSERT(tensor->ne[2] == ne2);
+    GGML_ASSERT(tensor->ne[3] == ne3);
+}
+
+int64_t get_example_targets_batch(
+    struct llama_context * lctx,
+    struct ggml_tensor   * tokens_input,
+    struct ggml_tensor   * target_probs,
+    int64_t                example_id,
+    const size_t         * samples_offs,
+    const size_t         * samples_begin,
+    const size_t         * samples_size,
+          size_t           samples_count,
+    const llama_token    * train_data,
+    size_t                 n_train_data,
+    bool                   separate_with_eos,
+    bool                   separate_with_bos,
+    bool                   fill_with_next_samples,
+    bool                   sample_random_offsets
+) {
+    GGML_ASSERT(samples_count > 0);
+    GGML_ASSERT(tokens_input->n_dims  == 2);
+    GGML_ASSERT(target_probs->n_dims  == 3);
+    int64_t n_vocab  = target_probs->ne[0];
+    int64_t n_tokens = tokens_input->ne[0];
+    int64_t n_batch  = tokens_input->ne[1];
+    GGML_ASSERT(n_vocab  == target_probs->ne[0]);
+    GGML_ASSERT(n_tokens == target_probs->ne[1]);
+    GGML_ASSERT(n_batch  == target_probs->ne[2]);
+
+    int64_t used_samples = 0;
+
+    ggml_set_f32(target_probs, 0.0f);
+    llama_token bos = llama_token_bos(lctx);
+    llama_token eos = llama_token_eos(lctx);
+    // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples);
+    for (int k=0; k<n_batch; ++k) {
+        // printf("%s: batch %d\n", __func__, k);
+        size_t sample_idx   = (example_id + used_samples) % samples_count;
+        size_t sample_offs  = sample_random_offsets ? samples_offs[sample_idx] : 0;
+        size_t sample_begin = samples_begin[sample_idx];
+        size_t sample_size  = samples_size[sample_idx];
+        ++used_samples;
+
+        // printf("%s: sample_idx=%zu sample=%zu\n", __func__, sample_idx, sample);
+        GGML_ASSERT(sample_begin+sample_size-1 < n_train_data);
+
+        ggml_set_i32_nd(tokens_input, 0, k, 0, 0, bos);
+        bool sample_separation_eos = !separate_with_eos;
+        bool sample_separation_bos = !separate_with_bos;
+        for (int64_t i=0; i<n_tokens; ++i) {
+            llama_token token = eos;
+            if (sample_offs >= sample_size && fill_with_next_samples) {
+                if (!sample_separation_eos) {
+                    // insert eos token to separate samples
+                    sample_separation_eos = true;
+                } else if (!sample_separation_bos) {
+                    // insert bos token to separate samples
+                    sample_separation_bos = true;
+                    token = bos;
+                } else {
+                    // sample separation is done, continue with next sample
+                    sample_separation_eos = !separate_with_eos;
+                    sample_separation_bos = !separate_with_bos;
+                    sample_offs  = 0;
+                    sample_idx   = (example_id + used_samples) % samples_count;
+                    sample_begin = samples_begin[sample_idx];
+                    sample_size  = samples_size[sample_idx];
+                    ++used_samples;
+                }
+            }
+            // note: no else-if here
+            if (sample_offs < sample_size) {
+                token = clamp(train_data[sample_begin+sample_offs], 0, (llama_token) (n_vocab - 1));
+                ++sample_offs;
+            }
+            ggml_set_f32_nd(target_probs,  token, (int) i, (int) k, 0, +1.0f);
+            if (i+1<n_tokens) {
+                ggml_set_i32_nd(tokens_input, (int) (i + 1), (int) k, 0, 0, token);
+            }
+        }
+    }
+
+    return used_samples;
+}
+
+void mt19937_set_state(std::mt19937& rng, const std::string& rng_state) {
+    std::stringstream s_rng_state;
+    s_rng_state.imbue(std::locale::classic());
+    s_rng_state.exceptions(std::stringstream::failbit);
+    s_rng_state.str(rng_state);
+    s_rng_state >> rng;
+}
+
+std::string mt19937_get_state(const std::mt19937& rng) {
+    std::stringstream s_rng_state;
+    s_rng_state.imbue(std::locale::classic());
+    s_rng_state << rng;
+    return s_rng_state.str();
+}
+
+std::string mt19937_seed_to_state(unsigned seed) {
+    std::mt19937 rng(seed);
+    return mt19937_get_state(rng);
+}
+
+std::string shuffle_samples(
+        const std::string & rng_state,
+        size_t            * shuffled_offs,
+        size_t            * shuffled_begins,
+        size_t            * shuffled_sizes,
+        const size_t      * begins,
+        const size_t      * sizes,
+        size_t              count) {
+    if (count == 0) return rng_state;
+
+    std::mt19937 rng;
+    mt19937_set_state(rng, rng_state);
+
+    // sort indices by random value for each index
+    std::vector<size_t> idcs;
+    {
+        std::vector<unsigned> rnd;
+        idcs.resize(count);
+        rnd.resize(count);
+        for (unsigned i=0; i<count; ++i) {
+            idcs[i] = i;
+            rnd[i]  = rng();
+        }
+
+        std::sort(idcs.begin(), idcs.end(), [&rnd](size_t a, size_t b){
+            // stable sort for reproducibility
+            return (rnd[a] == rnd[b]) ? (a < b) : (rnd[a] < rnd[b]);
+        });
+    }
+
+    // create random offsets
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_offs[i] = (size_t) ((sizes[idcs[i]] - 1) * ((double) rng() / (double) (rng.max()-1)));
+    }
+
+    // reorder begins and sizes by sorted indices
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_begins[i] = begins[idcs[i]];
+    }
+
+    for (unsigned i=0; i<count; ++i) {
+        shuffled_sizes[i] = sizes[idcs[i]];
+    }
+
+    return mt19937_get_state(rng);
+}
+
+size_t hash_combine(size_t h1, size_t h2) {
+    return h1 ^ (h2 << 1);
+}
+
+size_t compute_samples_hash(const char* fn, const size_t* samples_begin, const size_t* samples_size, size_t sample_count) {
+    std::hash<std::string> h_string;
+    std::hash<unsigned long long> h_ull;
+    size_t h = h_string(std::string(fn));
+    h = hash_combine(h, h_ull((unsigned long long) sample_count));
+    for (size_t i=0; i< sample_count; ++i) {
+        h = hash_combine(h, h_ull((unsigned long long) samples_begin[i]));
+        h = hash_combine(h, h_ull((unsigned long long) samples_size[i]));
+    }
+    return h;
+}
+
+std::string replace_str(const char * s, const char * needle, const char * replacement) {
+    std::string str = s;
+    size_t pos = str.find(needle);
+    if (pos != std::string::npos) {
+        str.replace(pos, strlen(needle), replacement);
+    }
+    return str;
+}
+
+void print_duration(double fmillis) {
+    if (fmillis < 1000.0f) {
+        printf("%.1fms", (float) fmillis);
+        return;
+    }
+    const int64_t one_sec  = 1000;
+    const int64_t one_min  = one_sec  * 60;
+    const int64_t one_hour = one_min  * 60;
+    const int64_t one_day  = one_hour * 24;
+
+    int64_t millis  = (int64_t) fmillis;
+    int64_t days    = millis/one_day;
+    int64_t hours   = (millis - days*one_day)/one_hour;
+    int64_t minutes = (millis - days*one_day - hours*one_hour)/one_min;
+    int64_t seconds = (millis - days*one_day - hours*one_hour - minutes*one_min)/one_sec;
+
+    // to print int64_t either cast to (long long int) or use macro PRId64 from <inttypes.h>
+    if (days > 0) {
+        printf("%lldd ", (long long int) days);
+    }
+    printf("%02lld:%02lld:%02lld", (long long int) hours, (long long int) minutes, (long long int) seconds);
+}
+
+float cosine_decay(int64_t step, int64_t decay_steps, float minimum) {
+    if (step > decay_steps) {
+        step = decay_steps;
+    }
+    const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps));
+    const float decay = (1 - minimum)*cosine_decay + minimum;
+    return decay;
+}
+
+float cosine_decay_restart(int64_t step, int64_t decay_steps, float minimum, float restart_step_mult) {
+    while (step > decay_steps) {
+        step -= decay_steps;
+        decay_steps = (int64_t) (restart_step_mult * decay_steps);
+    }
+    return cosine_decay(step, decay_steps, minimum);
+}
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t cos_decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart) {
+
+    float result =
+        (step < warmup_steps)
+            ? (float) step / (float) warmup_steps
+            : enable_restart
+                ? cosine_decay_restart(
+                    step - warmup_steps,
+                    cos_decay_steps,
+                    cos_decay_minimum,
+                    cos_decay_restart_step_mult)
+                : cosine_decay(
+                    step,
+                    cos_decay_steps,
+                    cos_decay_minimum);
+
+    float min = overall_minimum / learning_rate;
+    result = min + result * (1.0f - min);
+    return result;
+}
+
+static bool are_same_layout(struct ggml_tensor * a, struct ggml_tensor * b) {
+    GGML_ASSERT(a != NULL);
+    GGML_ASSERT(b != NULL);
+    GGML_ASSERT(a->type == b->type);
+    GGML_ASSERT(ggml_are_same_shape(a, b));
+    GGML_ASSERT(ggml_is_contiguous(a) && ggml_is_contiguous(b));
+
+    return true;
+}
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name) {
+    if (dst == NULL) {
+        return;
+    }
+    struct ggml_tensor * t  = ggml_get_tensor(ctx, name);
+    GGML_ASSERT(are_same_layout(dst, t));
+    memcpy(dst->data, t->data, ggml_nbytes(t));
+
+    if (strlen(ggml_get_name(dst)) == 0) {
+        ggml_set_name(dst, name);
+    }
+}
+
+// gguf constants
+static const char * LLM_KV_OPTIMIZER_TYPE = "optimizer.type";
+static const char * LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam";
+static const char * LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs";
+static const char * LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version";
+static const char * LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count";
+static const char * LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count";
+static const char * LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count";
+static const char * LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized";
+static const char * LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss";
+static const char * LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss";
+static const char * LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count";
+static const char * LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count";
+static const char * LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k";
+static const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end";
+static const char * LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count";
+
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments";
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments";
+static const char * LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values";
+
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s";
+static const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y";
+
+static const char * LLM_KV_TRAINING_FILE_VERSION         = "training.file_version";
+static const char * LLM_KV_TRAINING_ITERATION_COUNT      = "training.iteration_count";
+static const char * LLM_KV_TRAINING_SAMPLE_COUNT         = "training.sample_count";
+static const char * LLM_KV_TRAINING_TOKEN_COUNT          = "training.token_count";
+static const char * LLM_KV_TRAINING_EPOCH_COUNT          = "training.epoch_count";
+static const char * LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH = "training.shuffle.samples_hash";
+static const char * LLM_KV_TRAINING_SHUFFLE_RNG_STATE    = "training.shuffle.rng_state";
+static const char * LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT = "training.shuffle.sample_count";
+static const char * LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE  = "training.shuffle.next_sample";
+
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+{ \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        die_fmt("key not found in model: %s", skey.c_str()); \
+    } \
+}
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt) {
+    // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
+
+    uint32_t file_version;
+    GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_FILE_VERSION);
+    GGML_ASSERT(file_version == 0);
+
+    GGUF_GET_KEY(fctx, opt->params.past, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT);
+    GGUF_GET_KEY(fctx, opt->iter, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ITERATION_COUNT);
+    GGUF_GET_KEY(fctx, opt->just_initialized, gguf_get_val_bool, GGUF_TYPE_BOOL, true, LLM_KV_OPTIMIZER_JUST_INITIALIZED);
+
+    uint64_t nx;
+    GGUF_GET_KEY(fctx, nx, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_OPTIMIZER_PARAMETER_COUNT);
+    opt->nx = (size_t) nx;
+
+    // don't call ggml_opt_init until optimizer type and optimizer specific parameters are know
+
+    std::string opt_type;
+    GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
+    if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
+        opt->params.type = GGML_OPT_ADAM;
+
+        GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
+        GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
+        GGUF_GET_KEY(fctx, opt->adam.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT);
+
+        ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
+
+        copy_tensor_by_name(opt->adam.m,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
+        copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
+        copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
+    } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
+        opt->params.type = GGML_OPT_LBFGS;
+
+        GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
+        GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
+        GGUF_GET_KEY(fctx, opt->lbfgs.step,             gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP);
+        GGUF_GET_KEY(fctx, opt->lbfgs.j,                gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J);
+        GGUF_GET_KEY(fctx, opt->lbfgs.k,                gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K);
+        GGUF_GET_KEY(fctx, opt->lbfgs.end,              gguf_get_val_i32, GGUF_TYPE_INT32,   true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END);
+        GGUF_GET_KEY(fctx, opt->lbfgs.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT);
+
+        ggml_opt_init(opt->ctx, opt, opt->params, opt->nx);
+
+        copy_tensor_by_name(opt->lbfgs.x,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
+        copy_tensor_by_name(opt->lbfgs.xp,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
+        copy_tensor_by_name(opt->lbfgs.g,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
+        copy_tensor_by_name(opt->lbfgs.gp,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
+        copy_tensor_by_name(opt->lbfgs.d,    f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
+        copy_tensor_by_name(opt->lbfgs.pf,   f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
+        copy_tensor_by_name(opt->lbfgs.lmal, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
+        copy_tensor_by_name(opt->lbfgs.lmys, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
+        copy_tensor_by_name(opt->lbfgs.lms,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
+        copy_tensor_by_name(opt->lbfgs.lmy,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
+    } else {
+        die("unknown optimizer type\n");
+    }
+}
+
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt) {
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_FILE_VERSION, 0);
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, opt->params.past);
+    gguf_set_val_u64(fctx, LLM_KV_OPTIMIZER_PARAMETER_COUNT, (uint64_t) opt->nx);
+    gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ITERATION_COUNT, opt->iter);
+    gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
+
+    switch (opt->params.type) {
+        case GGML_OPT_ADAM:
+            {
+                gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS,        opt->adam.fx_prev);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, opt->adam.n_no_improvement);
+
+                ggml_set_name(opt->adam.m, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS);
+                ggml_set_name(opt->adam.v, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
+                if (opt->adam.pf) {
+                    ggml_set_name(opt->adam.pf, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
+                }
+
+                gguf_add_tensor(fctx, opt->adam.m);
+                gguf_add_tensor(fctx, opt->adam.v);
+                if (opt->adam.pf) {
+                    gguf_add_tensor(fctx, opt->adam.pf);
+                }
+            } break;
+        case GGML_OPT_LBFGS:
+            {
+                gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS,            opt->lbfgs.fx_best);
+                gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP,     opt->lbfgs.step);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J,        opt->lbfgs.j);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K,        opt->lbfgs.k);
+                gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END,      opt->lbfgs.end);
+                gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, opt->lbfgs.n_no_improvement);
+
+                ggml_set_name(opt->lbfgs.x,    LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS);
+                ggml_set_name(opt->lbfgs.xp,   LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS);
+                ggml_set_name(opt->lbfgs.g,    LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS);
+                ggml_set_name(opt->lbfgs.gp,   LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS);
+                ggml_set_name(opt->lbfgs.d,    LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION);
+                if (opt->lbfgs.pf) {
+                    ggml_set_name(opt->lbfgs.pf, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES);
+                }
+                ggml_set_name(opt->lbfgs.lmal, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA);
+                ggml_set_name(opt->lbfgs.lmys, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS);
+                ggml_set_name(opt->lbfgs.lms,  LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S);
+                ggml_set_name(opt->lbfgs.lmy,  LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y);
+
+                gguf_add_tensor(fctx, opt->lbfgs.x);
+                gguf_add_tensor(fctx, opt->lbfgs.xp);
+                gguf_add_tensor(fctx, opt->lbfgs.g);
+                gguf_add_tensor(fctx, opt->lbfgs.gp);
+                gguf_add_tensor(fctx, opt->lbfgs.d);
+                if (opt->lbfgs.pf) {
+                    gguf_add_tensor(fctx, opt->lbfgs.pf);
+                }
+                gguf_add_tensor(fctx, opt->lbfgs.lmal);
+                gguf_add_tensor(fctx, opt->lbfgs.lmys);
+                gguf_add_tensor(fctx, opt->lbfgs.lms);
+                gguf_add_tensor(fctx, opt->lbfgs.lmy);
+            } break;
+    }
+}
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train) {
+    if (gguf_find_key(fctx, LLM_KV_TRAINING_FILE_VERSION) < 0) {
+        return false;
+    }
+
+    uint32_t file_version;
+    GGUF_GET_KEY(fctx, file_version,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
+    GGML_ASSERT(file_version <= 1);
+
+    if (file_version == 0) {
+
+        GGUF_GET_KEY(fctx, train->train_its,     gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT);
+        GGUF_GET_KEY(fctx, train->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->train_tokens,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
+
+    } else if (file_version == 1) {
+
+        GGUF_GET_KEY(fctx, train->train_its,     gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_ITERATION_COUNT);
+        GGUF_GET_KEY(fctx, train->train_samples, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->train_tokens,  gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_TOKEN_COUNT);
+        GGUF_GET_KEY(fctx, train->train_epochs,  gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_TRAINING_EPOCH_COUNT);
+
+        GGUF_GET_KEY(fctx, train->shuffle_samples_hash,      gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH);
+        GGUF_GET_KEY(fctx, train->shuffle_rng_state_current, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_SHUFFLE_RNG_STATE);
+        GGUF_GET_KEY(fctx, train->shuffle_sample_count,      gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT);
+        GGUF_GET_KEY(fctx, train->shuffle_next_sample,       gguf_get_val_u64, GGUF_TYPE_UINT64, false, LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE);
+    }
+
+    load_opt_context_gguf(fctx, f_ggml_ctx, train->opt);
+    return true;
+}
+
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train) {
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_FILE_VERSION,    1);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_ITERATION_COUNT, train->train_its);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SAMPLE_COUNT,    train->train_samples);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_TOKEN_COUNT,     train->train_tokens);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_EPOCH_COUNT,     train->train_epochs);
+
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_SAMPLES_HASH, (uint64_t) train->shuffle_samples_hash);
+    gguf_set_val_str(fctx, LLM_KV_TRAINING_SHUFFLE_RNG_STATE,    train->shuffle_rng_state_current.c_str());
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_SAMPLE_COUNT, (uint64_t) train->shuffle_sample_count);
+    gguf_set_val_u64(fctx, LLM_KV_TRAINING_SHUFFLE_NEXT_SAMPLE,  (uint64_t) train->shuffle_next_sample);
+
+    save_opt_context_gguf(fctx, train->opt);
+}
+
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static size_t utf8_len(char src) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
+    return lookup[highbits];
+}
+
+// mark each byte with its utf8 unit number.
+// returns the number of utf8 characters.
+// e.g. when bytes == '\x61\xD0\xB0\x62',
+// then utf8_units will become [0,0,1,0]
+// utf8_nunits will become [1,2,2,1] and 3 is returned.
+// bytes where utf8_units is zero, are the begin of an utf8 character.
+static size_t mark_utf8_units(const char* bytes, int * utf8_units, int * utf8_nunits, size_t count) {
+    size_t offs = 0;
+    size_t count_utf8 = 0;
+    while(offs < count) {
+        int len = (int) utf8_len(bytes[offs]);
+        for (int i=0; i<len; ++i) {
+            utf8_units[offs+i]  = i;
+            utf8_nunits[offs+i] = len;
+        }
+        offs += len;
+        ++count_utf8;
+    }
+    return count_utf8;
+}
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size) {
+    struct llama_file f(filename, "rb");
+
+    if (f.size == 0) {
+        out_tokens.clear();
+        out_samples_begin.clear();
+        out_samples_size.clear();
+        printf("%s: warning: empty or not existing training data file '%s'\n",
+            __func__, filename);
+        return out_tokens.size();
+    }
+
+    // account for possible leading whitespace that will be added by tokenizer
+    // e.g. '\t' will be tokenized by llama spm tokenizer to [29871, 12]
+    const int n_max_tokens_overhead = 1;
+
+    std::vector<char> buf;
+    buf.resize(f.size);
+
+    f.read_raw(buf.data(), f.size);
+
+    std::vector<int> utf8_units;
+    std::vector<int> utf8_nunits;
+    utf8_units.resize(buf.size());
+    utf8_nunits.resize(buf.size());
+    mark_utf8_units(buf.data(), utf8_units.data(), utf8_nunits.data(), buf.size());
+
+    if (sample_start.size() == 0) {
+        // tokenize all data at once
+        out_tokens.resize(buf.size() + n_max_tokens_overhead);
+
+        int n_tokens = llama_tokenize(
+            llama_get_model(lctx),
+            buf.data(),
+            (int) buf.size(),
+            out_tokens.data(),
+            (int) out_tokens.size(),
+            false);
+        if (n_tokens < 0) {
+            out_tokens.resize(-n_tokens);
+            n_tokens = llama_tokenize(
+                llama_get_model(lctx),
+                buf.data(),
+                (int) buf.size(),
+                out_tokens.data(),
+                (int) out_tokens.size(),
+                false);
+        }
+        if (n_tokens >= 0) {
+            out_tokens.resize(n_tokens);
+        }
+
+        // generate sample starts at all token positions
+        out_samples_begin.clear();
+        out_samples_begin.push_back(0);
+        out_samples_size.push_back(std::min((size_t) context_length, out_tokens.size()));
+        size_t end = (out_tokens.size() >= context_length) ? (out_tokens.size() - context_length) : 0;
+        for (size_t sample_begin = 1; sample_begin < end; ++sample_begin) {
+            out_samples_begin.push_back(sample_begin);
+            out_samples_size.push_back(context_length);
+        }
+    } else {
+        // split data into samples and tokenize each sample
+        std::string data_str(buf.data(), buf.size());
+        out_samples_begin.clear();
+        out_samples_size.clear();
+        out_tokens.clear();
+
+        // find all positions of pattern sample_start
+        size_t sample_begin = data_str.find(sample_start, 0);
+        while (sample_begin != std::string::npos) {
+            out_samples_begin.push_back(sample_begin);
+            const size_t search_start = sample_begin + sample_start.size();
+            sample_begin = data_str.find(sample_start, search_start);
+        }
+        if (out_samples_begin.size() == 0) {
+            printf("%s: warning: sample start pattern '%s' not found. inserting single sample at data begin\n",
+                __func__, sample_start.c_str());
+            out_samples_begin.push_back(0);
+        }
+
+        out_samples_size.resize(out_samples_begin.size(), 0);
+
+        std::vector<char>        buf_sample;
+        std::vector<llama_token> tok_sample;
+
+        const size_t sample_begin_offset = (include_sample_start ? 0 : sample_start.size());
+        size_t found_too_big_sample   = 0;
+        size_t found_too_small_sample = 0;
+        size_t found_empty_sample     = 0;
+        size_t found_min_sample_size  = SIZE_MAX;
+        size_t found_max_sample_size  = 0;
+
+        size_t max_token_text_size = 0;
+        int n_vocab = llama_n_vocab(llama_get_model(lctx));
+        for (llama_token token=0; token < n_vocab; ++token) {
+            max_token_text_size = std::max(
+                max_token_text_size,
+                strlen(llama_token_get_text(lctx, token)));
+        }
+
+        // upper bound of context byte length.
+        // strings with this byte length should always tokenize to at least context_length tokens.
+        size_t context_byte_len = max_token_text_size*context_length;
+
+        for (unsigned i=0; i<out_samples_begin.size(); ++i) {
+            // determine sample begin and end from pattern positions
+            size_t sample_begin = out_samples_begin[i] + sample_begin_offset;
+            size_t sample_end   = overlapping_samples
+                                    ? std::min(
+                                        data_str.size(),
+                                        sample_begin + context_byte_len)
+                                    : (i+1 < out_samples_begin.size()
+                                        ? out_samples_begin[i+1]
+                                        : data_str.size());
+            if (sample_end < utf8_units.size() && utf8_units[sample_end] > 0) {
+                // sample end is in the middle of an utf8 character.
+                // advance sample_end to the begin of the next utf8 character.
+                sample_end += utf8_nunits[sample_end] - utf8_units[sample_end];
+            }
+            size_t sample_size = sample_end - sample_begin;
+            if (sample_size == 0) {
+                ++found_empty_sample;
+            }
+
+            if (sample_size > 0) {
+                // llama_tokenize expects zero terminated string,
+                // copy sample into buffer and zero terminate it.
+                buf_sample.resize(sample_size);
+                memcpy(buf_sample.data(), data_str.data() + sample_begin, sample_size);
+
+                // printf("sample: '%s'\n", buf_sample.data());
+
+                // tokenize the sample
+                tok_sample.resize(buf_sample.size() + n_max_tokens_overhead);
+                int n_tokens = llama_tokenize(llama_get_model(lctx),
+                    buf_sample.data(),
+                    (int) buf_sample.size(),
+                    tok_sample.data(),
+                    (int) tok_sample.size(),
+                    false);
+                if (n_tokens < 0) {
+                    tok_sample.resize(-n_tokens);
+                    n_tokens = llama_tokenize(llama_get_model(lctx),
+                        buf_sample.data(),
+                        (int) buf_sample.size(),
+                        tok_sample.data(),
+                        (int) tok_sample.size(),
+                        false);
+                    GGML_ASSERT(n_tokens >= 0);
+                }
+                GGML_ASSERT(n_tokens <= (int) tok_sample.size());
+
+                if ((size_t) n_tokens > context_length) {
+                    ++found_too_big_sample;
+                } else if ((size_t) n_tokens < context_length) {
+                    ++found_too_small_sample;
+                }
+                found_max_sample_size = std::max(found_max_sample_size, (size_t) n_tokens);
+                found_min_sample_size = std::min(found_min_sample_size, (size_t) n_tokens);
+
+                // write out tokens, start and size of sample
+                // overwrite the string start position with the token start position
+                out_samples_begin[i] = out_tokens.size();
+                out_samples_size[i] = (size_t) n_tokens;
+                out_tokens.insert(out_tokens.end(), tok_sample.begin(), tok_sample.begin() + n_tokens);
+            } else {
+                out_samples_begin[i] = out_tokens.size();
+                out_samples_size[i] = 0;
+            }
+
+        }
+        if (found_too_big_sample > 0) {
+            printf("%s: warning: found %zu samples (max length %zu) that exceed context length of %u. samples will be cut off.\n",
+                __func__, found_too_big_sample, found_max_sample_size, context_length);
+        }
+
+        if (found_too_small_sample > 0) {
+            printf("%s: warning: found %zu samples (min length %zu) that are shorter than context length of %u.\n",
+                __func__, found_too_small_sample, found_min_sample_size, context_length);
+        }
+
+        if (found_empty_sample) {
+            printf("%s: warning: found %zu empty samples.\n",
+                __func__, found_empty_sample);
+        }
+    }
+    printf("%s: total number of samples: %zu\n",
+        __func__, out_samples_begin.size());
+
+    GGML_ASSERT(out_samples_begin.size() == out_samples_size.size());
+
+    return out_tokens.size();
+}
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration) {
+    std::string sit = (iteration >= 0) ? std::to_string(iteration) : std::string(latest);
+    return replace_str(filename, pattern_it, sit.c_str());
+}
+
+struct train_params_common get_default_train_params_common() {
+    struct train_params_common params;
+    params.fn_train_data     = "shakespeare.txt";
+    params.fn_checkpoint_in  = "checkpoint.gguf";
+    params.fn_checkpoint_out = "checkpoint-ITERATION.gguf";
+    params.pattern_fn_it     = "ITERATION";
+    params.fn_latest         = "LATEST";
+
+    params.print_usage = false;
+
+    params.save_every = 10;
+
+    params.seed       =   -1;
+
+    params.n_ctx      =  128;
+    params.n_threads  =    6;
+    params.n_batch    =    8;
+    params.n_gradient_accumulation = 1;
+    params.n_epochs   = -1;
+
+    params.custom_n_ctx = false;
+
+    params.use_flash              = true;
+    params.use_checkpointing      = true;
+
+    params.sample_start           = "";
+    params.include_sample_start   = false;
+    params.escape                 = false;
+    params.overlapping_samples    = false;
+    params.fill_with_next_samples = false;
+    params.separate_with_eos      = false;
+    params.separate_with_bos      = true;
+    params.sample_random_offsets  = false;
+    params.force_reshuffle        = false;
+
+    params.opt_past               = 0;
+    params.opt_delta              = 1e-5f;
+    params.opt_max_no_improvement = 0;
+
+    params.warmup            =  100;
+    params.cos_decay_steps   = 1000;
+    params.cos_decay_restart = 1.1f;
+    params.cos_decay_min     = 0.1f;
+    params.enable_restart    = false;
+
+    params.adam_n_iter         = 256;
+    params.adam_alpha          = 1e-3f;
+    params.adam_min_alpha      = 0;
+    params.adam_decay          = 1e-1f;
+    params.adam_decay_min_ndim = 2;
+    params.adam_beta1          = 0.9f;
+    params.adam_beta2          = 0.999f;
+    params.adam_gclip          = 1.0f;
+    params.adam_eps_f          = 0.0f;
+    return params;
+}
+
+void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train_params_common * params) {
+    // fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    // fprintf(stderr, "\n");
+    // fprintf(stderr, "options:\n");
+    // fprintf(stderr, "  -h, --help                 show this help message and exit\n");
+    fprintf(stderr, "  --train-data FNAME         path from which to load training data (default '%s')\n", params->fn_train_data);
+    fprintf(stderr, "  --checkpoint-in FNAME      path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
+    fprintf(stderr, "  --checkpoint-out FNAME     path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
+    fprintf(stderr, "  --pattern-fn-it STR        pattern in output filenames to be replaced by iteration number (default '%s')\n", params->pattern_fn_it);
+    fprintf(stderr, "  --fn-latest STR            string to use instead of iteration number for saving latest output (default '%s')\n", params->fn_latest);
+    fprintf(stderr, "  --save-every N             save checkpoint and lora every N iterations. Disabled when N <= 0. (default '%d')\n", params->save_every);
+    fprintf(stderr, "  -s SEED, --seed SEED       RNG seed (default: -1, use random seed for -1)\n");
+    fprintf(stderr, "  -c N, --ctx N              Context size used during training (default %d)\n", params->n_ctx);
+    fprintf(stderr, "  -t N, --threads N          Number of threads (default %d)\n", params->n_threads);
+    fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
+    fprintf(stderr, "  --grad-acc N               Number of gradient accumulation steps (simulates larger batch size of batch*gradacc) (default %d)\n", params->n_gradient_accumulation);
+    fprintf(stderr, "  --sample-start STR         Sets the starting point for samples after the specified pattern. If empty use every token position as sample start. (default '%s')\n", params->sample_start.c_str());
+    fprintf(stderr, "  --include-sample-start     Include the sample start in the samples. (default off)\n");
+    fprintf(stderr, "  --escape                   process sample start escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    fprintf(stderr, "  --overlapping-samples      Samples my overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
+    fprintf(stderr, "  --fill-with-next-samples   Samples shorter than context length will be followed by the next (shuffled) samples. (default off)\n");
+    fprintf(stderr, "  --separate-with-eos        When fill-with-next-samples, insert end-of-sequence token between samples.%s\n", params->separate_with_eos ? " (default)" : "");
+    fprintf(stderr, "  --separate-with-bos        When fill-with-next-samples, insert begin-of-sequence token between samples.%s\n", params->separate_with_bos ? " (default)" : "");
+    fprintf(stderr, "  --no-separate-with-eos     When fill-with-next-samples, don't insert end-of-sequence token between samples.%s\n", !params->separate_with_eos ? " (default)" : "");
+    fprintf(stderr, "  --no-separate-with-bos     When fill-with-next-samples, don't insert begin-of-sequence token between samples.%s\n", !params->separate_with_bos ? " (default)" : "");
+    fprintf(stderr, "  --sample-random-offsets    Use samples beginning at random offsets. Together with fill-with-next-samples this may help for training endless text generation.%s\n", params->sample_random_offsets ? " (default)" : "");
+    fprintf(stderr, "  --force-reshuffle          Force a reshuffling of data at program start, otherwise the shuffling of loaded checkpoint is resumed.\n");
+    fprintf(stderr, "  --no-flash                 Don't use flash attention \n");
+    fprintf(stderr, "  --use-flash                Use flash attention (default)\n");
+    fprintf(stderr, "  --no-checkpointing         Don't use gradient checkpointing\n");
+    fprintf(stderr, "  --use-checkpointing        Use gradient checkpointing (default)\n");
+    fprintf(stderr, "  --warmup N                 Only for Adam optimizer. Number of warmup steps (default %d)\n", params->warmup);
+    fprintf(stderr, "  --cos-decay-steps N        Only for Adam optimizer. Number of cosine decay steps (default %d)\n", params->cos_decay_steps);
+    fprintf(stderr, "  --cos-decay-restart N      Only for Adam optimizer. Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart);
+    fprintf(stderr, "  --cos-decay-min N          Only for Adam optimizer. Cosine decay minimum (default %f)\n", params->cos_decay_min);
+    fprintf(stderr, "  --enable-restart N         Only for Adam optimizer. Enable restarts of cos-decay %s\n", params->enable_restart ? "(default)" : "");
+    fprintf(stderr, "  --disable-restart N        Only for Adam optimizer. Disable restarts of cos-decay %s\n", !params->enable_restart ? "(default)" : "");
+    fprintf(stderr, "  --opt-past N               Number of optimization iterations to track for delta convergence test. Disabled when zero. (default %d)\n", params->opt_past);
+    fprintf(stderr, "  --opt-delta N              Maximum delta for delta convergence test. Disabled when <= zero. (default %f)\n", params->opt_delta);
+    fprintf(stderr, "  --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. (default %d)\n", params->opt_max_no_improvement);
+    fprintf(stderr, "  --epochs N                 Maximum number epochs to process. (default %d)\n", params->n_epochs);
+    fprintf(stderr, "  --adam-iter N              Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter);
+    fprintf(stderr, "  --adam-alpha N             Adam learning rate alpha (default %f)\n", params->adam_alpha);
+    fprintf(stderr, "  --adam-min-alpha N         Adam minimum learning rate alpha - including warmup phase (default %f)\n", params->adam_min_alpha);
+    fprintf(stderr, "  --adam-decay N             AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
+    fprintf(stderr, "  --adam-decay-min-ndim N    Minimum number of tensor dimensions to apply AdamW weight decay. Weight decay is not applied to tensors with less n_dims. (default %d)\n", params->adam_decay_min_ndim);
+    fprintf(stderr, "  --adam-beta1 N             AdamW beta1 in interval [0,1). How much to smooth the first moment of gradients. (default %f)\n", params->adam_beta1);
+    fprintf(stderr, "  --adam-beta2 N             AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
+    fprintf(stderr, "  --adam-gclip N             AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
+    fprintf(stderr, "  --adam-epsf N              AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f);
+    fprintf(stderr, "\n");
+}
+
+bool consume_common_train_arg(
+    int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param
+) {
+    int& i = *idx;
+    std::string arg = argv[i];
+    const std::string arg_prefix = "--";
+    if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+        std::replace(arg.begin(), arg.end(), '_', '-');
+    }
+    if (arg == "--train-data") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_train_data = argv[i];
+    } else if (arg == "--checkpoint-in") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_checkpoint_in = argv[i];
+    } else if (arg == "--checkpoint-out") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_checkpoint_out = argv[i];
+    } else if (arg == "--pattern-fn-it") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->pattern_fn_it = argv[i];
+    } else if (arg == "--fn-latest") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->fn_latest = argv[i];
+    } else if (arg == "--save-every") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->save_every = std::stoi(argv[i]);
+    } else if (arg == "-s" || arg == "--seed") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->seed = std::stoi(argv[i]);
+    } else if (arg == "-c" || arg == "--ctx") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_ctx = std::stoi(argv[i]);
+        params->custom_n_ctx = true;
+    } else if (arg == "-t" || arg == "--threads") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_threads = std::stoi(argv[i]);
+    } else if (arg == "-b" || arg == "--batch") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_batch = std::stoi(argv[i]);
+    } else if (arg == "--grad-acc") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_gradient_accumulation = std::max(1, std::stoi(argv[i]));
+    } else if (arg == "--sample-start") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->sample_start = std::string(argv[i]);
+    } else if (arg == "--escape") {
+        params->escape = true;
+    } else if (arg == "--include-sample-start") {
+        params->include_sample_start = true;
+    } else if (arg == "--overlapping-samples") {
+        params->overlapping_samples = true;
+    } else if (arg == "--fill-with-next-samples") {
+        params->fill_with_next_samples = true;
+    } else if (arg == "--separate-with-eos") {
+        params->separate_with_eos = true;
+    } else if (arg == "--separate-with-bos") {
+        params->separate_with_bos = true;
+    } else if (arg == "--no-separate-with-eos") {
+        params->separate_with_eos = false;
+    } else if (arg == "--no-separate-with-bos") {
+        params->separate_with_bos = false;
+    } else if (arg == "--sample-random-offsets") {
+        params->sample_random_offsets = true;
+    } else if (arg == "--force-reshuffle") {
+        params->force_reshuffle = true;
+    } else if (arg == "--no-flash") {
+        params->use_flash = false;
+    } else if (arg == "--use-flash") {
+        params->use_flash = true;
+    } else if (arg == "--no-checkpointing") {
+        params->use_checkpointing = false;
+    } else if (arg == "--use-checkpointing") {
+        params->use_checkpointing = true;
+    } else if (arg == "--warmup") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->warmup = std::stoi(argv[i]);
+    } else if (arg == "--cos-decay-steps") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_steps = std::stoi(argv[i]);
+    } else if (arg == "--cos-decay-restart") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_restart = std::stof(argv[i]);
+    } else if (arg == "--cos-decay-min") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->cos_decay_min = std::stof(argv[i]);
+    } else if (arg == "--enable-restart") {
+        params->enable_restart = true;
+    } else if (arg == "--disable-restart") {
+        params->enable_restart = false;
+    } else if (arg == "--opt-past") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_past = std::stoi(argv[i]);
+    } else if (arg == "--opt-delta") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_delta = std::stof(argv[i]);
+    } else if (arg == "--opt-max-no-improvement") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->opt_max_no_improvement = std::stoi(argv[i]);
+    } else if (arg == "--adam-epsf") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_eps_f = std::stof(argv[i]);
+    } else if (arg == "--epochs") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->n_epochs = std::stoi(argv[i]);
+    } else if (arg == "--adam-iter") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_n_iter = std::stoi(argv[i]);
+    } else if (arg == "--adam-alpha") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_alpha = std::stof(argv[i]);
+    } else if (arg == "--adam-min-alpha") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_min_alpha = std::stof(argv[i]);
+    } else if (arg == "--adam-decay") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_decay = std::stof(argv[i]);
+    } else if (arg == "--adam-decay-min-ndim") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_decay_min_ndim = std::stoi(argv[i]);
+    } else if (arg == "--adam-beta1") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_beta1 = std::stof(argv[i]);
+    } else if (arg == "--adam-beta2") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_beta2 = std::stof(argv[i]);
+    } else if (arg == "--adam-gclip") {
+        if (++i >= argc) {
+            *invalid_param = true;
+            return true;
+        }
+        params->adam_gclip = std::stof(argv[i]);
+    } else if (arg == "-h" || arg == "--help") {
+        params->print_usage = true;
+        return true;
+    } else {
+        return false;
+    }
+    return true;
+}
+
+void finish_processing_train_args(struct train_params_common * params) {
+    if (params->escape) {
+        process_escapes(params->sample_start);
+    }
+}
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel) {
+    struct train_opt_callback_data * data   = (struct train_opt_callback_data *) vdata;
+    struct train_params_common     * params = data->params;
+    struct train_state             * train  = data->train;
+    struct ggml_opt_context        * opt    = train->opt;
+    int n_batch = params->n_batch;
+    int n_ctx = params->n_ctx;
+
+    if (accum_step == 0) {
+        // time measurement
+        int64_t now = ggml_time_ms();
+        if (now > data->last_time && opt->iter > data->first_iter) {
+            double dt = (double) (now - data->last_time);
+            if (data->millis_per_iter == 0.0) {
+                data->millis_per_iter = dt;
+            } else {
+                const double gain = 0.7;
+                data->millis_per_iter = data->millis_per_iter*(1.0-gain) + dt*gain;
+            }
+        }
+
+        double remaining_millis = 0.0;
+        if (data->millis_per_iter > 0.0) {
+            const int n_iter = params->adam_n_iter;
+            const int done_iter = opt->iter - data->first_iter;
+            const int remaining_iter = n_iter - done_iter;
+            remaining_millis = remaining_iter * data->millis_per_iter;
+        }
+
+        // file saving
+        const bool save_now = (params->save_every > 0) && (opt->iter - data->last_save_iter >= params->save_every);
+        if (save_now) {
+            int new_iters = opt->iter - data->last_save_iter;
+            train->train_its    += new_iters;
+            train->train_tokens += new_iters * opt->params.n_gradient_accumulation * n_batch * n_ctx;
+
+            if (data->save_cb) {
+                data->save_cb(data->save_data, train);
+            }
+
+            data->last_save_iter = opt->iter;
+        }
+
+        // exclude file saving from time measurement, by measuring last_time after saving
+        data->last_time = ggml_time_ms();
+
+        *sched = learning_schedule(
+            opt->iter,
+            params->warmup,
+            params->cos_decay_steps,
+            params->adam_alpha,
+            params->adam_min_alpha,
+            params->cos_decay_min,
+            params->cos_decay_restart,
+            params->enable_restart);
+
+        int impr_plot = -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
+        if (impr_plot > 0) impr_plot = 0;
+        if (std::isnan(opt->loss_before) || std::isnan(opt->loss_before)) impr_plot = 0;
+        printf("%s: iter=%6d sample=%zu/%zu sched=%f loss=%f",
+            __func__, opt->iter, std::min(1+train->shuffle_next_sample, train->shuffle_sample_count), train->shuffle_sample_count,
+            *sched, opt->loss_after);
+
+
+        if (data->millis_per_iter > 0) {
+            printf(" dt=");
+            print_duration(data->millis_per_iter);
+            printf(" eta=");
+            print_duration(remaining_millis);
+        }
+
+        float improvement = opt->loss_before - opt->loss_after;
+        const float plot_scale = 10.0f;
+        int bar_len = (int)(1 + improvement*plot_scale + 0.5);
+        printf(" |");
+        for (int i=0; i<bar_len; ++i) {
+            printf("-");
+        }
+        printf(">");
+        printf("\n");
+    }
+
+    int64_t used_samples = get_example_targets_batch(
+        data->lctx,
+        data->tokens_input,
+        data->target_probs,
+        train->shuffle_next_sample,
+        data->shuffled_samples_offs,
+        data->shuffled_samples_begin,
+        data->shuffled_samples_size,
+        data->samples_count,
+        data->tokens_data,
+        data->tokens_size,
+        params->separate_with_eos,
+        params->separate_with_bos,
+        params->fill_with_next_samples,
+        params->sample_random_offsets);
+
+    train->train_samples += used_samples;
+    train->shuffle_next_sample += used_samples;
+
+    if (train->shuffle_next_sample >= train->shuffle_sample_count) {
+        ++train->train_epochs;
+        printf("%s: reshuffle samples. completed epochs: %llu\n", __func__, (long long unsigned) train->train_epochs);
+        // note: we may have used some samples from the current shuffling more than once
+        train->shuffle_rng_state_current = train->shuffle_rng_state_next;
+        train->shuffle_rng_state_next = shuffle_samples(
+            train->shuffle_rng_state_current,
+            data->shuffled_samples_offs,
+            data->shuffled_samples_begin,
+            data->shuffled_samples_size,
+            data->samples_begin,
+            data->samples_size,
+            data->samples_count);
+        train->shuffle_next_sample = 0;
+    }
+
+    const bool last_epoch_reached = (params->n_epochs > 0 && (int64_t) train->train_epochs - data->first_epoch >= params->n_epochs);
+    if (last_epoch_reached) {
+        // allow optimization iteration at last epoch to be completed before canceling
+        if (data->iter_at_last_epoch < 0) {
+            data->iter_at_last_epoch = opt->iter;
+        } else if (opt->iter > data->iter_at_last_epoch) {
+            *cancel = true;
+        }
+    }
+}
diff --git a/common/train.h b/common/train.h
new file mode 100644
index 000000000..42fa704b8
--- /dev/null
+++ b/common/train.h
@@ -0,0 +1,230 @@
+// Various helper functions and utilities for training
+
+#pragma once
+
+#include <string>
+#include <random>
+#include <vector>
+
+#include "ggml.h"
+#include "llama.h"
+
+typedef std::string mt19937_state;
+
+struct train_state {
+    struct ggml_opt_context * opt;
+
+    uint64_t train_its;
+    uint64_t train_samples;
+    uint64_t train_tokens;
+    uint64_t train_epochs;
+
+    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
+    mt19937_state shuffle_rng_state_current;
+    mt19937_state shuffle_rng_state_next;
+    size_t        shuffle_sample_count;
+    size_t        shuffle_next_sample;
+};
+
+struct train_params_common {
+    const char * fn_train_data;
+    const char * fn_checkpoint_in;
+    const char * fn_checkpoint_out;
+    const char * pattern_fn_it;
+    const char * fn_latest;
+
+    bool print_usage;
+
+    int save_every;
+
+    uint32_t seed;
+
+    int n_ctx;
+    int n_threads;
+    int n_batch;
+    int n_gradient_accumulation;
+    int n_epochs;
+
+    bool custom_n_ctx;
+
+    bool use_flash;
+    bool use_checkpointing;
+
+    std::string sample_start;
+    bool include_sample_start;
+    bool escape;
+    bool overlapping_samples;
+    bool fill_with_next_samples;
+    bool separate_with_eos;
+    bool separate_with_bos;
+    bool sample_random_offsets;
+
+    bool force_reshuffle;
+
+    int   warmup;
+    int   cos_decay_steps;
+    float cos_decay_restart;
+    float cos_decay_min;
+    bool  enable_restart;
+
+    int   opt_past;
+    float opt_delta;
+    int   opt_max_no_improvement;
+
+    int   adam_n_iter;
+    float adam_alpha;
+    float adam_min_alpha;
+    float adam_decay;
+    int   adam_decay_min_ndim;
+    float adam_beta1;
+    float adam_beta2;
+    float adam_gclip;
+    float adam_eps_f;
+};
+
+typedef void (*save_train_files_callback)(void * data, struct train_state * train);
+
+struct train_opt_callback_data {
+    struct train_params_common * params;
+    struct train_state         * train;
+    save_train_files_callback    save_cb;
+    void                       * save_data;
+    struct llama_context       * lctx;
+    int                          last_save_iter;
+    llama_token                * tokens_data;
+    size_t                       tokens_size;
+    size_t                     * samples_begin;
+    size_t                     * samples_size;
+    size_t                     * shuffled_samples_offs;
+    size_t                     * shuffled_samples_begin;
+    size_t                     * shuffled_samples_size;
+    size_t                       samples_count;
+    struct ggml_tensor         * tokens_input;
+    struct ggml_tensor         * target_probs;
+    int                          first_iter;
+    int                          first_epoch;
+    int                          iter_at_last_epoch;
+    int64_t                      last_time;
+    double                       millis_per_iter;
+};
+
+struct train_state * init_train_state();
+void free_train_state(struct train_state  * state);
+
+struct train_params_common get_default_train_params_common();
+void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
+
+bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
+void finish_processing_train_args(struct train_params_common * params);
+
+struct random_normal_distribution;
+struct random_uniform_distribution;
+
+struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
+struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
+
+void free_random_normal_distribution (struct random_normal_distribution  * rnd);
+void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
+
+struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
+struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
+
+// generate random float in interval [0,1)
+float frand();
+float frand_normal (struct random_normal_distribution * rnd);
+float frand_uniform(struct random_uniform_distribution * rnd);
+
+int   clamp (const int v, const int min, const int max);
+float fclamp(const float v, const float min, const float max);
+
+void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
+void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
+void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
+void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
+
+size_t tokenize_file(
+        struct llama_context     * lctx,
+        const char               * filename,
+        const std::string        & sample_start,
+        bool                       include_sample_start,
+        bool                       overlapping_samples,
+        unsigned                   context_length,
+        std::vector<llama_token> & out_tokens,
+        std::vector<size_t>      & out_samples_begin,
+        std::vector<size_t>      & out_samples_size);
+
+int64_t get_example_targets_batch(
+        struct llama_context * lctx,
+        struct ggml_tensor   * tokens_input,
+        struct ggml_tensor   * target_probs,
+        int64_t                example_id,
+        const size_t         * samples_offs,
+        const size_t         * samples_begin,
+        const size_t         * samples_size,
+              size_t           samples_count,
+        const llama_token    * train_data,
+        size_t                 n_train_data,
+        bool                   separate_with_eos,
+        bool                   separate_with_bos,
+        bool                   fill_with_next_samples,
+        bool                   sample_random_offsets);
+
+
+void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
+mt19937_state mt19937_get_state(const std::mt19937& rng);
+mt19937_state mt19937_seed_to_state(unsigned seed);
+
+mt19937_state shuffle_samples(
+        const mt19937_state & rng_state,
+        size_t              * shuffled_offs,
+        size_t              * shuffled_begins,
+        size_t              * shuffled_sizes,
+        const size_t        * begins,
+        const size_t        * sizes,
+        size_t                count);
+
+size_t hash_combine(size_t h1, size_t h2);
+
+size_t compute_samples_hash(
+    const char* fn,
+    const size_t* samples_begin,
+    const size_t* samples_size,
+    size_t sample_count);
+
+
+std::string replace_str(const char * s, const char * needle, const char * replacement);
+
+void print_duration(double milliseconds);
+
+float cosine_decay(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum);
+
+float cosine_decay_restart(
+    int64_t step,
+    int64_t decay_steps,
+    float   minimum,
+    float   restart_step_mult);
+
+float learning_schedule(
+    int64_t step,
+    int64_t warmup_steps,
+    int64_t decay_steps,
+    float   learning_rate,
+    float   overall_minimum,
+    float   cos_decay_minimum,
+    float   cos_decay_restart_step_mult,
+    bool    enable_restart);
+
+void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
+
+void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
+void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
+
+bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
+void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
+
+std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
+
+void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);
diff --git a/convert-baichuan-hf-to-gguf.py b/convert-baichuan-hf-to-gguf.py
new file mode 100755
index 000000000..513a7516a
--- /dev/null
+++ b/convert-baichuan-hf-to-gguf.py
@@ -0,0 +1,310 @@
+#!/usr/bin/env python3
+# HF baichuan --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+import itertools
+import numpy as np
+import torch
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
+
+# reverse HF permute back to original pth layout
+
+
+def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
+
+    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape))
+
+def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray:
+        r = weights.shape[0] // 3
+        return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
+
+def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray:
+        r = weights.shape[0] // 3
+        return weights[r * n_part : r * n_part + r, ...]
+
+def count_model_parts(dir_model: str) -> int:
+    num_parts = 0
+
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+
+    return num_parts
+
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+print("hello print: ",hparams["architectures"][0])
+if hparams["architectures"][0] != "BaichuanForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+print(f"num_parts:{num_parts}\n")
+ARCH=gguf.MODEL_ARCH.BAICHUAN
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+head_count = hparams["num_attention_heads"]
+
+if "num_key_value_heads" in hparams:
+    head_count_kv = hparams["num_key_value_heads"]
+else:
+    head_count_kv = head_count
+
+if "_name_or_path" in hparams:
+    hf_repo = hparams["_name_or_path"]
+else:
+    hf_repo = ""
+
+if "max_sequence_length" in hparams:
+    ctx_length = hparams["max_sequence_length"]
+elif "max_position_embeddings" in hparams:
+    ctx_length = hparams["max_position_embeddings"]
+elif "model_max_length" in hparams:
+    ctx_length = hparams["model_max_length"]
+else:
+    print("gguf: can not find ctx length parameter.")
+
+    sys.exit()
+
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_source_hf_repo(hf_repo)
+gguf_writer.add_tensor_data_layout("Meta AI original pth")
+gguf_writer.add_context_length(ctx_length)
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+gguf_writer.add_head_count(head_count)
+gguf_writer.add_head_count_kv(head_count_kv)
+gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+
+if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
+    if "type" in hparams["rope_scaling"]:
+        if hparams["rope_scaling"]["type"] == "linear":
+            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytes] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
+
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab, scores and token types")
+
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
+vocab_size = hparams.get('vocab_size')
+if vocab_size is None:
+    vocab_size = tokenizer.vocab_size()
+
+for i in range(vocab_size):
+    text: bytes
+    score: float
+
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
+
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
+
+    # toktype = 4 is user-defined = tokens from added_tokens.json
+
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
+
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
+
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
+
+        print("gguf: get added tokens")
+
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
+
+
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    tmp=model_part
+    for i in range(block_count):
+        if f"model.layers.{i}.self_attn.W_pack.weight" in model_part:
+            print(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+        # we don't need these
+        if name.endswith(".rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")
diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py
index 50069db56..9252e1c46 100755
--- a/convert-falcon-hf-to-gguf.py
+++ b/convert-falcon-hf-to-gguf.py
@@ -1,45 +1,30 @@
 #!/usr/bin/env python3
 # HF falcon--> gguf conversion
 
-import gguf
-import os
-import sys
-import struct
+from __future__ import annotations
+
+import argparse
+import contextlib
 import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
 import numpy as np
 import torch
+from transformers import AutoTokenizer  # type: ignore[import]
 
-from typing import Any, List
-from pathlib import Path
-from transformers import AutoTokenizer
-
-def bytes_to_unicode():
-    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path, prefix: str) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
+        if filename.startswith(prefix):
             num_parts += 1
 
     if num_parts > 0:
@@ -47,17 +32,34 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -65,219 +67,184 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
-if hparams["architectures"][0] != "RWForCausalLM":
+if hparams["architectures"][0] != "FalconForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
-    sys.exit()
+    sys.exit(1)
 
 # get number of model parts
-num_parts = count_model_parts(dir_model)
+num_parts = count_model_parts(dir_model, "model-00")
+if num_parts:
+    is_safetensors = True
+    from safetensors import safe_open
+else:
+    is_safetensors = False
+    num_parts = count_model_parts(dir_model, "pytorch_model-")
 
 ARCH=gguf.MODEL_ARCH.FALCON
 gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-block_count = hparams["n_layer"]
+block_count = hparams["num_hidden_layers"]
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
 gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
-if "n_head_kv" in hparams: gguf_writer.add_head_count_kv(hparams["n_head_kv"])
+gguf_writer.add_head_count(hparams["num_attention_heads"])
+if "num_kv_heads" in hparams:
+    gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
+else:
+    gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
 
 # TOKENIZATION
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
-merges: List[str] = []
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+print("gguf: get gpt2 tokenizer vocab")
 
-    print("gguf: get gpt2 tokenizer merges")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
 
-    gguf_writer.add_token_merges(merges)
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
-    print("gguf: get gpt2 tokenizer vocab")
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i])
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
-
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
-
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-
-    gguf_writer.add_token_list(tokens)
-
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        # find special token ids
-
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
 # params for qkv transform
-n_head = hparams["n_head"]
-n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+n_head    = hparams["num_attention_heads"]
+n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
+
 head_dim = hparams["hidden_size"] // n_head
 
 # tensor info
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
+elif is_safetensors:
+    part_names = (
+        f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
+    )
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+    if is_safetensors:
+        ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
+    else:
+        ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
 
-    for name in model_part.keys():
-        data = model_part[name]
+    with ctx as model_part:
+        for name in model_part.keys():
+            data = model_part.get_tensor(name) if is_safetensors else model_part[name]
 
-        old_dtype = data.dtype
+            old_dtype = data.dtype
 
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
+            # convert any unsupported data types to float32
+            if data.dtype != torch.float16 and data.dtype != torch.float32:
+                data = data.to(torch.float32)
 
-        # QKV tensor transform
-        # The original query_key_value tensor contains n_head_kv "kv groups",
-        # each consisting of n_head/n_head_kv query weights followed by one key
-        # and one value weight (shared by all query heads in the kv group).
-        # This layout makes it a big pain to work with in GGML.
-        # So we rearrange them here,, so that we have n_head query weights
-        # followed by n_head_kv key weights followed by n_head_kv value weights,
-        # in contiguous fashion.
-        # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+            # QKV tensor transform
+            # The original query_key_value tensor contains n_head_kv "kv groups",
+            # each consisting of n_head/n_head_kv query weights followed by one key
+            # and one value weight (shared by all query heads in the kv group).
+            # This layout makes it a big pain to work with in GGML.
+            # So we rearrange them here,, so that we have n_head query weights
+            # followed by n_head_kv key weights followed by n_head_kv value weights,
+            # in contiguous fashion.
+            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
 
-        if "query_key_value" in name:
-            qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
-            q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
-            k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
-            v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
-            data = torch.cat((q,k,v)).reshape_as(data)
+            if "query_key_value" in name:
+                qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+                q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
+                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                data = torch.cat((q,k,v)).reshape_as(data)
 
-        data = data.squeeze().numpy()
+            data = data.squeeze().numpy()
 
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+            if new_name is None:
+                print("Can not map tensor '" + name + "'")
+                sys.exit()
 
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
 
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
+            # if f32 desired, convert any float16 to float32
+            if ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
 
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
 
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+            print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+            gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py
index 6eeff5bb1..d4e85f518 100755
--- a/convert-gptneox-hf-to-gguf.py
+++ b/convert-gptneox-hf-to-gguf.py
@@ -1,44 +1,26 @@
 #!/usr/bin/env python3
 # HF gptneox--> gguf conversion
 
-import gguf
-import os
-import sys
-import struct
+from __future__ import annotations
+
+import argparse
 import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
 import numpy as np
 import torch
+from transformers import AutoTokenizer  # type: ignore[import]
 
-from typing import Any, List
-from pathlib import Path
-from transformers import AutoTokenizer
-
-# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-
-def count_model_parts(dir_model: str) -> int:
+def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
     for filename in os.listdir(dir_model):
         if filename.startswith("pytorch_model-"):
@@ -49,17 +31,34 @@ def count_model_parts(dir_model: str) -> int:
     return num_parts
 
 
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
     sys.exit(1)
 
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
 # possible tensor data types
 #   ftype == 0 -> float32
 #   ftype == 1 -> float16
@@ -67,19 +66,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
 # map from ftype to string
 ftype_str = ["f32", "f16"]
 
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
 
-        sys.exit(1)
+print("gguf: loading model "+dir_model.name)
 
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
 if hparams["architectures"][0] != "GPTNeoXForCausalLM":
@@ -97,7 +92,7 @@ print("gguf: get model metadata")
 
 block_count = hparams["num_hidden_layers"]
 
-gguf_writer.add_name(last_dir)
+gguf_writer.add_name(dir_model.name)
 gguf_writer.add_context_length(hparams["max_position_embeddings"])
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
@@ -111,86 +106,36 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
 
 print("gguf: get tokenizer metadata")
 
-tokens: List[str] = []
-merges: List[str] = []
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
 
-if Path(dir_model + "/tokenizer.json").is_file():
-    # gpt2 tokenizer
-    gguf_writer.add_tokenizer_model("gpt2")
+print("gguf: get gpt2 tokenizer vocab")
 
-    print("gguf: get gpt2 tokenizer merges")
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
 
-    gguf_writer.add_token_merges(merges)
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
-    print("gguf: get gpt2 tokenizer vocab")
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
 
-    vocab_size = len(tokenizer_json["model"]["vocab"])
-
-    # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
-    tokenizer = AutoTokenizer.from_pretrained(dir_model)
-
-    reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-    byte_encoder = bytes_to_unicode()
-    byte_decoder = {v: k for k, v in byte_encoder.items()}
-
-    for i in range(vocab_size):
-        if i in reverse_vocab:
-            try:
-                text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
-            except KeyError:
-                text = bytearray()
-                for c in reverse_vocab[i]:
-                    if ord(c) < 256:  # single byte character
-                        text.append(byte_decoder[ord(c)])
-                    else:  # multibyte special token character
-                        text.extend(c.encode('utf-8'))
-        else:
-            print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
-            pad_token = f"[PAD{i}]".encode("utf8")
-            text = bytearray(pad_token)
-
-        tokens.append(text)
-
-    gguf_writer.add_token_list(tokens)
-
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
-        print("gguf: get special token ids")
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        # find special token ids
-
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS
 
@@ -200,13 +145,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 print("gguf: get tensor metadata")
 
 if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
+    part_names = iter(("pytorch_model.bin",))
 else:
     part_names = (
         f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
     )
 
 for part_name in part_names:
+    if args.vocab_only:
+        break
     print("gguf: loading model part '" + part_name + "'")
     model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
 
@@ -226,11 +173,8 @@ for part_name in part_names:
         data = data.squeeze().numpy()
 
         # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
             print("Can not map tensor '" + name + "'")
             sys.exit()
 
@@ -249,19 +193,20 @@ for part_name in part_names:
         if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
             data = data.astype(np.float16)
 
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-        gguf_writer.add_tensor(name, data)
+        gguf_writer.add_tensor(new_name, data)
 
 
 print("gguf: write header")
 gguf_writer.write_header_to_file()
 print("gguf: write metadata")
 gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
 
 gguf_writer.close()
 
-print("gguf: model successfully exported to '" + fname_out + "'")
+print(f"gguf: model successfully exported to '{fname_out}'")
 print("")
diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py
deleted file mode 100755
index f103f5f61..000000000
--- a/convert-llama-7b-pth-to-gguf.py
+++ /dev/null
@@ -1,308 +0,0 @@
-#!/usr/bin/env python3
-# 7b pth llama --> gguf conversion
-# Only models with a single datafile are supported, like 7B
-# HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model
-
-import gguf
-import os
-import sys
-import struct
-import json
-import numpy as np
-import torch
-
-from typing import Any, List
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
-
-#NDArray = np.ndarray[Any, Any]
-# compatible with python < 3.9
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
-
-
-def count_model_parts(dir_model: str) -> int:
-    num_parts = 0
-    for filename in os.listdir(dir_model):
-        if filename.startswith("consolidated."):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-    return num_parts
-
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-
-    sys.exit(1)
-
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-
-        sys.exit(1)
-
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "LlamaForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-if num_parts > 1:
-    print("gguf: Only models with a single datafile are supported.")
-
-    sys.exit()
-
-ARCH=gguf.MODEL_ARCH.LLAMA
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-head_count = hparams["num_attention_heads"]
-
-if "num_key_value_heads" in hparams:
-    head_count_kv = hparams["num_key_value_heads"]
-else:
-    head_count_kv = head_count
-
-if "_name_or_path" in hparams:
-    hf_repo = hparams["_name_or_path"]
-else:
-    hf_repo = ""
-
-if "max_sequence_length" in hparams:
-    ctx_length = hparams["max_sequence_length"]
-elif "max_position_embeddings" in hparams:
-    ctx_length = hparams["max_position_embeddings"]
-else:
-    print("gguf: can not find ctx length parameter.")
-
-    sys.exit()
-
-
-gguf_writer.add_name(last_dir)
-gguf_writer.add_source_hf_repo(hf_repo)
-gguf_writer.add_tensor_data_layout("Meta AI original pth")
-gguf_writer.add_context_length(ctx_length)
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-gguf_writer.add_head_count(head_count)
-gguf_writer.add_head_count_kv(head_count_kv)
-gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
-
-if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
-    if "type" in hparams["rope_scaling"]:
-        if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
-
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: List[bytes] = []
-scores: List[float] = []
-toktypes: List[int] = []
-
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab and scores")
-
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-
-        # toktype = 4 is user-defined = tokens from added_tokens.json
-
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
-
-            print("gguf: get added tokens")
-
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
-
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts))
-
-for part_name in part_names:
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name == "rope.freqs":
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-
-print("gguf: model successfully exported to '" + fname_out + "'")
-print("")
diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggml-to-gguf.py
similarity index 63%
rename from convert-llama-ggmlv3-to-gguf.py
rename to convert-llama-ggml-to-gguf.py
index 3bf93627d..b5d3e0b3c 100755
--- a/convert-llama-ggmlv3-to-gguf.py
+++ b/convert-llama-ggml-to-gguf.py
@@ -1,9 +1,18 @@
 #!/usr/bin/env python3
-import sys, struct, math, argparse
+from __future__ import annotations
+
+import argparse
+import math
+import struct
+import sys
+from enum import IntEnum
 from pathlib import Path
 
 import numpy as np
 
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
 # Note: Does not support GGML_QKK_64
@@ -26,10 +35,35 @@ GGML_QUANT_SIZES = {
     gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
 }
 
+class GGMLFormat(IntEnum):
+    GGML = 0
+    GGMF = 1
+    GGJT = 2
+
+class GGMLFType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1
+    MOSTLY_Q4_0          = 2
+    MOSTLY_Q4_1          = 3
+    MOSTLY_Q4_1_SOME_F16 = 4
+    MOSTLY_Q8_0          = 7
+    MOSTLY_Q5_0          = 8
+    MOSTLY_Q5_1          = 9
+    MOSTLY_Q2_K          = 10
+    MOSTLY_Q3_K_S        = 11
+    MOSTLY_Q3_K_M        = 12
+    MOSTLY_Q3_K_L        = 13
+    MOSTLY_Q4_K_S        = 14
+    MOSTLY_Q4_K_M        = 15
+    MOSTLY_Q5_K_S        = 16
+    MOSTLY_Q5_K_M        = 17
+    MOSTLY_Q6_K          = 18
+
 class Hyperparameters:
     def __init__(self):
-        self.n_vocab = self.n_embd = self.n_mult = self.n_head = self.n_layer = self.n_rot = self.ftype = 0
-        self.n_ff = 0
+        self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
+        self.n_layer = self.n_rot = self.n_ff = 0
+        self.ftype = GGMLFType.ALL_F32
 
     def set_n_ff(self, model):
         ff_tensor_idx = model.tensor_map.get(b'layers.0.feed_forward.w1.weight')
@@ -45,16 +79,21 @@ class Hyperparameters:
             self.n_head,
             self.n_layer,
             self.n_rot,
-            self.ftype,
+            ftype,
         ) = struct.unpack('<7I', data[offset:offset + (4 * 7)])
+        try:
+            self.ftype = GGMLFType(ftype)
+        except ValueError:
+            raise ValueError(f'Invalid ftype {ftype}')
         return 4 * 7
 
     def __str__(self):
-        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype}>'
+        return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
 
 class Vocab:
-    def __init__(self):
+    def __init__(self, load_scores = True):
         self.items = []
+        self.load_scores = load_scores
 
     def load(self, data, offset, n_vocab):
         orig_offset = offset
@@ -62,20 +101,24 @@ class Vocab:
             itemlen = struct.unpack('<I', data[offset:offset + 4])[0]
             assert itemlen < 4096, 'Absurd vocab item length'
             offset += 4
-            vocab = bytes(data[offset:offset + itemlen])
+            item_text = bytes(data[offset:offset + itemlen])
             offset += itemlen
-            score = struct.unpack('<f', data[offset:offset + 4])[0]
-            offset += 4
-            self.items.append((vocab, score))
+            if self.load_scores:
+                item_score = struct.unpack('<f', data[offset:offset + 4])[0]
+                offset += 4
+            else:
+                item_score = 0.0
+            self.items.append((item_text, item_score))
         return offset - orig_offset
 
 class Tensor:
-    def __init__(self):
+    def __init__(self, use_padding = True):
         self.name = None
-        self.dims = ()
+        self.dims: tuple[int, ...] = ()
         self.dtype = None
         self.start_offset = 0
-        self.len_bytes = 0
+        self.len_bytes = np.int64(0)
+        self.use_padding = use_padding
 
     def load(self, data, offset):
         orig_offset = offset
@@ -91,7 +134,7 @@ class Tensor:
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])
         offset += name_len
-        pad = ((offset + 31) & ~31) - offset
+        pad = ((offset + 31) & ~31) - offset if self.use_padding else 0
         offset += pad
         n_elems = np.prod(self.dims)
         n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize)
@@ -101,7 +144,7 @@ class Tensor:
         # print(n_dims, name_len, dtype, self.dims, self.name, pad)
         return offset - orig_offset
 
-class GGMLV3Model:
+class GGMLModel:
     def __init__(self):
         self.hyperparameters = None
         self.vocab = None
@@ -109,20 +152,52 @@ class GGMLV3Model:
         self.tensors = []
 
     def validate_header(self, data, offset):
-        if bytes(data[offset:offset + 4]) != b'tjgg' or struct.unpack('<I', data[offset + 4:offset + 8])[0] != 3:
-            raise ValueError('Only GGJTv3 supported')
-        return 8
+        magic = bytes(data[offset:offset + 4])
+        if magic == b'GGUF':
+            raise ValueError('File is already in GGUF format.')
+        if magic == b'lmgg':
+            self.file_format = GGMLFormat.GGML
+            self.format_version = 1
+            return 4
+        version = struct.unpack('<I', data[offset + 4:offset + 8])[0]
+        if magic == b'fmgg':
+            if version != 1:
+                raise ValueError(f'Cannot handle unexpected GGMF file version {version}')
+            self.file_format = GGMLFormat.GGMF
+            self.format_version = version
+            return 8
+        if magic == b'tjgg':
+            if version < 1 or version > 3:
+                raise ValueError(f'Cannot handle unexpected GGJT file version {version}')
+            self.file_format = GGMLFormat.GGJT
+            self.format_version = version
+            return 8
+        raise ValueError(f"Unexpected file magic {magic!r}! This doesn't look like a GGML format file.")
+
+    def validate_conversion(self, ftype):
+        err = ''
+        if (self.file_format < GGMLFormat.GGJT or self.format_version < 2):
+            if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
+                err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
+        elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
+            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+                err = 'Q4 and Q8 quantizations changed in GGJTv3.'
+        if len(err) > 0:
+            raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
 
     def load(self, data, offset):
         offset += self.validate_header(data, offset)
         hp = Hyperparameters()
         offset += hp.load(data, offset)
-        vocab = Vocab()
+        print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
+        self.validate_conversion(hp.ftype)
+        vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
         offset += vocab.load(data, offset, hp.n_vocab)
-        tensors = []
+        tensors: list[Tensor] = []
         tensor_map = {}
         while offset < len(data):
-            tensor = Tensor()
+            tensor = Tensor(use_padding = self.file_format > GGMLFormat.GGMF)
             offset += tensor.load(data, offset)
             tensor_map[tensor.name] = len(tensors)
             tensors.append(tensor)
@@ -134,13 +209,14 @@ class GGMLV3Model:
         return offset
 
 class GGMLToGGUF:
-    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None):
+    def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
         hp = ggml_model.hyperparameters
         self.model = ggml_model
         self.data = data
         self.cfg = cfg
         self.params_override = params_override
         self.vocab_override = vocab_override
+        self.special_vocab = special_vocab
         if params_override is not None:
             n_kv_head = params_override.n_head_kv
         else:
@@ -159,9 +235,14 @@ class GGMLToGGUF:
 
     def save(self):
         print('* Preparing to save GGUF file')
-        gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+        gguf_writer = gguf.GGUFWriter(
+            self.cfg.output,
+            gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
+            use_temp_file = False )
         self.add_params(gguf_writer)
         self.add_vocab(gguf_writer)
+        if self.special_vocab is not None:
+            self.special_vocab.add_to_gguf(gguf_writer)
         self.add_tensors(gguf_writer)
         print("    gguf: write header")
         gguf_writer.write_header_to_file()
@@ -174,7 +255,10 @@ class GGMLToGGUF:
     def add_params(self, gguf_writer):
         hp = self.model.hyperparameters
         cfg = self.cfg
-        desc = cfg.desc if cfg.desc is not None else 'converted from legacy GGJTv3 format'
+        if cfg.desc is not None:
+            desc = cfg.desc
+        else:
+            desc = f'converted from legacy {self.model.file_format.name}v{self.model.format_version} {hp.ftype.name} format'
         try:
             # Filenames aren't necessarily valid UTF8.
             name = cfg.name if cfg.name is not None else cfg.input.name
@@ -184,6 +268,7 @@ class GGMLToGGUF:
         if name is not None:
             gguf_writer.add_name(name)
         gguf_writer.add_description(desc)
+        gguf_writer.add_file_type(int(hp.ftype))
         if self.params_override is not None:
             po = self.params_override
             assert po.n_embd == hp.n_embd, 'Model hyperparams mismatch'
@@ -220,7 +305,8 @@ class GGMLToGGUF:
                 tokens.append(vbytes)
                 scores.append(score)
                 toktypes.append(ttype)
-            assert len(tokens) == hp.n_vocab, f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
+            assert len(tokens) == hp.n_vocab, \
+                f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}'
             gguf_writer.add_token_list(tokens)
             gguf_writer.add_token_scores(scores)
             if len(toktypes) > 0:
@@ -259,27 +345,24 @@ class GGMLToGGUF:
         gguf_writer.add_eos_token_id(2)
 
     def add_tensors(self, gguf_writer):
-        nm = self.name_map
+        tensor_map = self.name_map
         data = self.data
         print(f'* Adding {len(self.model.tensors)} tensor(s)')
         for tensor in self.model.tensors:
             name = str(tensor.name, 'UTF-8')
-            if name.endswith('.weight'):
-                name = name[:-7]
-                suffix = '.weight'
-            elif name.endswith('.bias'):
-                name = name[:-5]
-                suffix = '.bias'
-            mapped_name = nm.get(name)
+            mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
             assert mapped_name is not None, f'Bad name {name}'
-            mapped_name += suffix
             tempdims = list(tensor.dims[:])
             if len(tempdims) > 1:
                 temp = tempdims[1]
                 tempdims[1] = tempdims[0]
                 tempdims[0] = temp
             # print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}')
-            gguf_writer.add_tensor(mapped_name, data[tensor.start_offset:tensor.start_offset + tensor.len_bytes], raw_shape = tempdims, raw_dtype = tensor.dtype)
+            gguf_writer.add_tensor(
+                mapped_name,
+                data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
+                raw_shape = tempdims,
+                raw_dtype = tensor.dtype )
 
 def handle_metadata(cfg, hp):
     import convert
@@ -301,43 +384,66 @@ def handle_metadata(cfg, hp):
         params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
     else:
         raise ValueError('Unable to load metadata')
-    vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype)
+    vocab = convert.load_vocab(
+        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+        cfg.vocabtype )
+    # FIXME: Respect cfg.vocab_dir?
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
-    return (params, vocab)
+    return (params, vocab, svocab)
 
 def handle_args():
-    parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF')
-    parser.add_argument('--input', '-i', type = Path, help = 'Input GGMLv3 filename')
-    parser.add_argument('--output', '-o', type = Path, help ='Output GGUF filename')
-    parser.add_argument('--name', help = 'Set model name')
-    parser.add_argument('--desc', help = 'Set model description')
-    parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
-    parser.add_argument('--eps', default = '5.0e-06', help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
-    parser.add_argument('--context-length', '-c', type=int, default = 2048, help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
-    parser.add_argument('--model-metadata-dir', '-m', type = Path, help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
-    parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
-    parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)", default="spm")
+    parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
+    parser.add_argument('--input', '-i', type = Path, required = True,
+        help = 'Input GGMLv3 filename')
+    parser.add_argument('--output', '-o', type = Path, required = True,
+        help ='Output GGUF filename')
+    parser.add_argument('--name',
+        help = 'Set model name')
+    parser.add_argument('--desc',
+        help = 'Set model description')
+    parser.add_argument('--gqa', type = int, default = 1,
+        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+    parser.add_argument('--eps', default = '5.0e-06',
+        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+    parser.add_argument('--context-length', '-c', type=int, default = 2048,
+        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+    parser.add_argument('--model-metadata-dir', '-m', type = Path,
+        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+    parser.add_argument("--vocab-dir", type=Path,
+        help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
 def main():
     cfg = handle_args()
     print(f'* Using config: {cfg}')
     print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n')
+    if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
+        print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
     data = np.memmap(cfg.input, mode = 'r')
-    model = GGMLV3Model()
+    model = GGMLModel()
     print('* Scanning GGML input file')
     offset = model.load(data, 0)
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
+    special_vocab = None
     if cfg.model_metadata_dir is not None:
-        (params_override, vocab_override) = handle_metadata(cfg, model.hyperparameters)
+        (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
         print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
         print(f'* Overriding params: {params_override}')
         print(f'* Overriding vocab: {vocab_override}')
+        print(f'* Special vocab: {special_vocab}')
     else:
         print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
-    converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override)
+        if model.file_format == GGMLFormat.GGML:
+            print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
+    converter = GGMLToGGUF(model, data, cfg,
+        params_override = params_override,
+        vocab_override = vocab_override,
+        special_vocab = special_vocab )
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py
deleted file mode 100755
index 08fde238b..000000000
--- a/convert-llama-hf-to-gguf.py
+++ /dev/null
@@ -1,328 +0,0 @@
-#!/usr/bin/env python3
-# HF llama --> gguf conversion
-
-import gguf
-import os
-import sys
-import struct
-import json
-import numpy as np
-import torch
-
-from typing import Any, List, Optional
-from pathlib import Path
-from sentencepiece import SentencePieceProcessor
-
-#NDArray = np.ndarray[Any, Any]
-# compatible with python < 3.9
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
-
-# reverse HF permute back to original pth layout
-# https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py
-
-
-def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray:
-    if n_kv_head is not None and n_head != n_kv_head:
-        n_head //= n_kv_head
-
-    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-            .swapaxes(1, 2)
-            .reshape(weights.shape))
-
-
-def count_model_parts(dir_model: str) -> int:
-    num_parts = 0
-
-    for filename in os.listdir(dir_model):
-        if filename.startswith("pytorch_model-"):
-            num_parts += 1
-
-    if num_parts > 0:
-        print("gguf: found " + str(num_parts) + " model parts")
-
-    return num_parts
-
-
-if len(sys.argv) < 3:
-    print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
-    print("  ftype == 0 -> float32")
-    print("  ftype == 1 -> float16")
-
-    sys.exit(1)
-
-
-# output in the same directory as the model
-dir_model = sys.argv[1]
-last_dir = os.path.basename(os.path.normpath(dir_model))
-
-
-# possible tensor data types
-#   ftype == 0 -> float32
-#   ftype == 1 -> float16
-
-
-# map from ftype to string
-ftype_str = ["f32", "f16"]
-
-ftype = 1
-if len(sys.argv) > 2:
-    ftype = int(sys.argv[2])
-    if ftype < 0 or ftype > 1:
-        print("Invalid ftype: " + str(ftype))
-
-        sys.exit(1)
-
-fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
-
-print("gguf: loading model "+last_dir)
-
-with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
-    hparams = json.load(f)
-
-if hparams["architectures"][0] != "LlamaForCausalLM":
-    print("Model architecture not supported: " + hparams["architectures"][0])
-
-    sys.exit()
-
-# get number of model parts
-num_parts = count_model_parts(dir_model)
-
-ARCH=gguf.MODEL_ARCH.LLAMA
-gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
-
-print("gguf: get model metadata")
-
-block_count = hparams["num_hidden_layers"]
-head_count = hparams["num_attention_heads"]
-
-if "num_key_value_heads" in hparams:
-    head_count_kv = hparams["num_key_value_heads"]
-else:
-    head_count_kv = head_count
-
-if "_name_or_path" in hparams:
-    hf_repo = hparams["_name_or_path"]
-else:
-    hf_repo = ""
-
-if "max_sequence_length" in hparams:
-    ctx_length = hparams["max_sequence_length"]
-elif "max_position_embeddings" in hparams:
-    ctx_length = hparams["max_position_embeddings"]
-else:
-    print("gguf: can not find ctx length parameter.")
-
-    sys.exit()
-
-
-gguf_writer.add_name(last_dir)
-gguf_writer.add_source_hf_repo(hf_repo)
-gguf_writer.add_tensor_data_layout("Meta AI original pth")
-gguf_writer.add_context_length(ctx_length)
-gguf_writer.add_embedding_length(hparams["hidden_size"])
-gguf_writer.add_block_count(block_count)
-gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-gguf_writer.add_head_count(head_count)
-gguf_writer.add_head_count_kv(head_count_kv)
-gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
-
-if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
-    if "type" in hparams["rope_scaling"]:
-        if hparams["rope_scaling"]["type"] == "linear":
-            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
-
-
-# TOKENIZATION
-
-print("gguf: get tokenizer metadata")
-
-tokens: List[bytes] = []
-scores: List[float] = []
-toktypes: List[int] = []
-
-if Path(dir_model + "/tokenizer.model").is_file():
-    # vocab type sentencepiece
-    print("gguf: get sentencepiece tokenizer vocab, scores and token types")
-
-    tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model")
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1  # defualt to normal token type
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-
-        # toktype = 4 is user-defined = tokens from added_tokens.json
-
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-
-    if Path(dir_model + "/added_tokens.json").is_file():
-        with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
-            addtokens_json = json.load(f)
-
-            print("gguf: get added tokens")
-
-            for key in addtokens_json:
-                tokens.append( key.encode("utf-8") )
-                scores.append(-1000.0)
-                toktypes.append(4) # user-defined token type
-
-
-    gguf_writer.add_tokenizer_model("llama")
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-
-
-print("gguf: get special token ids")
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
-
-# TENSORS
-
-tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
-
-# tensor info
-print("gguf: get tensor metadata")
-
-if num_parts == 0:
-    part_names = ("pytorch_model.bin",)
-else:
-    part_names = (
-        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
-    )
-
-for part_name in part_names:
-    print("gguf: loading model part '" + part_name + "'")
-    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
-
-    for name in model_part.keys():
-        data = model_part[name]
-
-        # we don't need these
-        if name.endswith(".rotary_emb.inv_freq"):
-            continue
-
-        old_dtype = data.dtype
-
-        # convert any unsupported data types to float32
-        if data.dtype != torch.float16 and data.dtype != torch.float32:
-            data = data.to(torch.float32)
-
-        data = data.squeeze().numpy()
-
-        # reverse permute these
-        if name.endswith(".q_proj.weight"):
-            data = reverse_hf_permute(data, head_count)
-        if name.endswith(".k_proj.weight"):
-            data = reverse_hf_permute(data, head_count, head_count_kv)
-
-        # map tensor names
-        if name.endswith(".weight") and name[:-7] in tensor_map:
-            name = tensor_map[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tensor_map:
-            name = tensor_map[name[:-5]] + ".bias"
-        else:
-            print("Can not map tensor '" + name + "'")
-            sys.exit()
-
-        n_dims = len(data.shape)
-        data_dtype = data.dtype
-
-        # if f32 desired, convert any float16 to float32
-        if ftype == 0 and data_dtype == np.float16:
-            data = data.astype(np.float32)
-
-        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-            data = data.astype(np.float32)
-
-        # if f16 desired, convert any float32 2-dim weight tensors to float16
-        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-            data = data.astype(np.float16)
-
-        print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
-
-        gguf_writer.add_tensor(name, data)
-
-
-print("gguf: write header")
-gguf_writer.write_header_to_file()
-print("gguf: write metadata")
-gguf_writer.write_kv_data_to_file()
-print("gguf: write tensors")
-gguf_writer.write_tensors_to_file()
-
-gguf_writer.close()
-
-
-print("gguf: model successfully exported to '" + fname_out + "'")
-print("")
diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py
index a94a7d0af..a937410dd 100755
--- a/convert-lora-to-ggml.py
+++ b/convert-lora-to-ggml.py
@@ -1,15 +1,17 @@
 #!/usr/bin/env python3
+from __future__ import annotations
+
 import json
 import os
 import re
 import struct
 import sys
-from typing import Any, Dict, Sequence, TextIO
+from typing import Any, BinaryIO, Sequence
 
 import numpy as np
 import torch
 
-NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1}
+NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
 HF_SUBLAYER_TO_GGML = {
@@ -46,7 +48,7 @@ def translate_tensor_name(t: str) -> str:
         sys.exit(1)
 
 
-def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
+def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
     fout.write(struct.pack("i", params["r"]))
@@ -60,7 +62,7 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
 
 
 def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
 ) -> None:
     sname = name.encode("utf-8")
     fout.write(
diff --git a/convert-persimmon-to-gguf.py b/convert-persimmon-to-gguf.py
new file mode 100644
index 000000000..e022ffe46
--- /dev/null
+++ b/convert-persimmon-to-gguf.py
@@ -0,0 +1,130 @@
+import torch
+import os
+from pprint import pprint
+import sys
+import argparse
+from pathlib import Path
+from sentencepiece import SentencePieceProcessor
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+def _flatten_dict(dct, tensors, prefix=None):
+    assert isinstance(dct, dict)
+    for key in dct.keys():
+        new_prefix = prefix + '.' + key if prefix is not None else key
+        if isinstance(dct[key], torch.Tensor):
+            tensors[new_prefix] = dct[key]
+        elif isinstance(dct[key], dict):
+            _flatten_dict(dct[key], tensors, new_prefix)
+        else:
+            raise ValueError(type(dct[key]))
+    return None
+
+def _get_sentencepiece_tokenizer_info(dir_model: Path):
+    tokenizer_path = dir_model / 'adept_vocab.model'
+    print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
+    tokenizer = SentencePieceProcessor(str(tokenizer_path))
+    print('gguf: adding tokens')
+    tokens: list[bytes] = []
+    scores: list[float] = []
+    toktypes: list[int] = []
+
+    for i in range(tokenizer.vocab_size()):
+        text: bytes
+        score: float
+
+        piece = tokenizer.id_to_piece(i)
+        text = piece.encode("utf-8")
+        score = tokenizer.get_score(i)
+
+        toktype = 1
+        if tokenizer.is_unknown(i):
+            toktype = 2
+        if tokenizer.is_control(i):
+            toktype = 3
+        if tokenizer.is_unused(i):
+            toktype = 5
+        if tokenizer.is_byte(i):
+            toktype = 6
+
+        tokens.append(text)
+        scores.append(score)
+        toktypes.append(toktype)
+        pass
+    return tokens, scores, toktypes
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
+    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
+    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
+    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
+    parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory")
+    args = parser.parse_args()
+    sys.path.append(str(args.adept_inference_dir))
+    persimmon_model = torch.load(args.ckpt_path)
+    hparams = persimmon_model['args']
+    pprint(hparams)
+    tensors = {}
+    _flatten_dict(persimmon_model['model'], tensors, None)
+
+    arch = gguf.MODEL_ARCH.PERSIMMON
+    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
+
+    block_count = hparams.num_layers
+    head_count = hparams.num_attention_heads
+    head_count_kv = head_count
+    ctx_length = hparams.seq_length
+    hidden_size = hparams.hidden_size
+
+    gguf_writer.add_name('persimmon-8b-chat')
+    gguf_writer.add_context_length(ctx_length)
+    gguf_writer.add_embedding_length(hidden_size)
+    gguf_writer.add_block_count(block_count)
+    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+    gguf_writer.add_head_count(head_count)
+    gguf_writer.add_head_count_kv(head_count_kv)
+    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
+    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
+
+    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
+    gguf_writer.add_tokenizer_model('llama')
+    gguf_writer.add_token_list(tokens)
+    gguf_writer.add_token_scores(scores)
+    gguf_writer.add_token_types(toktypes)
+    gguf_writer.add_bos_token_id(71013)
+    gguf_writer.add_eos_token_id(71013)
+
+    tensor_map = gguf.get_tensor_name_map(arch, block_count)
+    print(tensor_map)
+    for name in tensors.keys():
+        data = tensors[name]
+        if name.endswith(".self_attention.rotary_emb.inv_freq"):
+            continue
+        old_dtype = data.dtype
+        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
+        data = data.to(torch.float32).squeeze().numpy()
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+        n_dims = len(data.shape)
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+    print("gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+    gguf_writer.close()
+
+    print(f"gguf: model successfully exported to '{args.outfile}'")
+    print("")
+
+
+
+if __name__ == '__main__':
+    main()
diff --git a/convert-refact-hf-to-gguf.py b/convert-refact-hf-to-gguf.py
new file mode 100755
index 000000000..e0cd417db
--- /dev/null
+++ b/convert-refact-hf-to-gguf.py
@@ -0,0 +1,318 @@
+#!/usr/bin/env python3
+# HF refact--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if "NO_LOCAL_GGUF" not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
+import gguf
+
+
+def bytes_to_unicode():
+    # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    return dict(zip(bs, (chr(n) for n in cs)))
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert a Refact model to a GGML compatible file"
+    )
+    parser.add_argument(
+        "--vocab-only",
+        action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile",
+        type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model",
+        type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype",
+        type=int,
+        choices=[0, 1],
+        default=1,
+        nargs="?",
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f"Error: {args.model} is not a directory", file=sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf"
+
+print("gguf: loading model " + dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTRefactForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH = gguf.MODEL_ARCH.REFACT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+# Get refact feed forward dimension
+hidden_dim = hparams["n_embd"]
+inner_dim = 4 * hidden_dim
+hidden_dim = int(2 * inner_dim / 3)
+multiple_of = 256
+ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Refact")
+# refact uses Alibi. So this is from config.json which might be used by training.
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+
+gguf_writer.add_feed_forward_length(ff_dim)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_json_file = dir_model / "tokenizer.json"
+if not tokenizer_json_file.is_file():
+    print(f"Error: Missing {tokenizer_json_file}", file=sys.stderr)
+    sys.exit(1)
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+with open(tokenizer_json_file, "r", encoding="utf-8") as f:
+    tokenizer_json = json.load(f)
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = (
+    hparams["vocab_size"]
+    if "vocab_size" in hparams
+    else len(tokenizer_json["model"]["vocab"])
+)
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+for i in range(vocab_size):
+    if i in reverse_vocab:
+        text = reverse_vocab[i]
+        try:
+            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+        except KeyError:
+            text = bytearray()
+            for c in reverse_vocab[i]:
+                if ord(c) < 256:  # single byte character
+                    text.append(byte_decoder[ord(c)])
+                else:  # multibyte special token character
+                    text.extend(c.encode("utf-8"))
+    else:
+        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
+        pad_token = f"[PAD{i}]".encode("utf8")
+        text = bytearray(pad_token)
+
+    tokens.append(text)
+    scores.append(0.0)  # dymmy
+    toktypes.append(gguf.TokenType.NORMAL)  # dummy
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+# params for qkv transform
+n_head = hparams["n_head"]
+n_head_kv = 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for i in range(block_count):
+        if f"transformer.h.{i}.attn.kv.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.attn.kv.weight"]
+            model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[
+                : n_head_kv * head_dim
+            ]
+            model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[
+                n_head_kv * head_dim :
+            ]
+            del model_part[f"transformer.h.{i}.attn.kv.weight"]
+        if f"transformer.h.{i}.attn.q.weight" in model_part:
+            model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[
+                f"transformer.h.{i}.attn.q.weight"
+            ]
+            del model_part[f"transformer.h.{i}.attn.q.weight"]
+        if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+            model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim]
+            model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:]
+            del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if (
+            ftype == 1
+            and data_dtype == np.float32
+            and name.endswith(".weight")
+            and n_dims == 2
+        ):
+            data = data.astype(np.float16)
+
+        print(
+            new_name
+            + ", n_dims = "
+            + str(n_dims)
+            + ", "
+            + str(old_dtype)
+            + " --> "
+            + str(data.dtype)
+        )
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")
diff --git a/convert-starcoder-hf-to-gguf.py b/convert-starcoder-hf-to-gguf.py
new file mode 100755
index 000000000..90fa0c32f
--- /dev/null
+++ b/convert-starcoder-hf-to-gguf.py
@@ -0,0 +1,202 @@
+#!/usr/bin/env python3
+# HF starcoder --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    parser.add_argument("--outfile",    type=Path,           help="path to write to; default: based on input")
+    parser.add_argument("model",        type=Path,           help="directory containing model file, or model file itself (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.STARCODER
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("StarCoder")
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# params for qkv transform
+n_head    = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")
diff --git a/convert.py b/convert.py
index a701ab41b..e9b08d344 100755
--- a/convert.py
+++ b/convert.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
+from __future__ import annotations
 
-import gguf
 import argparse
 import concurrent.futures
 import copy
@@ -17,52 +17,98 @@ import re
 import signal
 import struct
 import sys
+import time
 import zipfile
-import numpy as np
-
 from abc import ABCMeta, abstractmethod
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List, Literal, Optional, Sequence, Tuple, TypeVar, Union)
-from sentencepiece import SentencePieceProcessor  # type: ignore
+from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
+
+import numpy as np
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+import os
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
 
 if TYPE_CHECKING:
-    from typing_extensions import TypeAlias
+    from typing import TypeAlias
 
 if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
     faulthandler.register(signal.SIGUSR1)
 
-NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 
-ARCH=gguf.MODEL_ARCH.LLAMA
-NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
+ARCH = gguf.MODEL_ARCH.LLAMA
 
+DEFAULT_CONCURRENCY = 8
 #
 # data types
 #
 
 @dataclass(frozen=True)
-class UnquantizedDataType:
+class DataType:
     name: str
+    dtype: np.dtype[Any]
+    valid_conversions: list[str]
 
-DT_F16  = UnquantizedDataType('F16')
-DT_F32  = UnquantizedDataType('F32')
-DT_I32  = UnquantizedDataType('I32')
-DT_BF16 = UnquantizedDataType('BF16')
+    def elements_to_bytes(self, n_elements: int) -> int:
+        return n_elements * self.dtype.itemsize
 
-DataType = Union[UnquantizedDataType]
+@dataclass(frozen=True)
+class UnquantizedDataType(DataType):
+    pass
 
-DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = {
-    DT_BF16: np.dtype(np.uint16),
-    DT_F16:  np.dtype(np.float16),
-    DT_F32:  np.dtype(np.float32),
-    DT_I32:  np.dtype(np.int32),
-}
+DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
 
-NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \
-    {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()}
+@dataclass(frozen=True)
+class QuantizedDataType(DataType):
+    block_size: int
+    quantized_dtype: np.dtype[Any]
+    ggml_type: gguf.GGMLQuantizationType
 
-SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
+    def quantize(self, arr: NDArray) -> NDArray:
+        raise NotImplementedError(f'Quantization for {self.name} not implemented')
+
+    def elements_to_bytes(self, n_elements: int) -> int:
+        assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
+        return self.quantized_dtype.itemsize * (n_elements // self.block_size)
+
+@dataclass(frozen=True)
+class Q8_0QuantizedDataType(QuantizedDataType):
+    # Mini Q8_0 quantization in Python!
+    def quantize(self, arr: NDArray) -> NDArray:
+        assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}'
+        assert arr.dtype == np.float32, f'Bad array type {arr.dtype}'
+        n_blocks = arr.size // self.block_size
+        blocks = arr.reshape((n_blocks, self.block_size))
+        # Much faster implementation of block quantization contributed by @Cebtenzzre
+        def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
+            d = abs(blocks).max(axis = 1) / np.float32(127)
+            with np.errstate(divide = 'ignore'):
+                qs = (blocks / d[:, None]).round()
+            qs[d == 0] = 0
+            yield from zip(d, qs)
+        return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
+
+DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
+    dtype = np.dtype(np.float32), valid_conversions = [],
+    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+
+# Quantized types skipped here because they may also map to np.float32
+NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
+for dt in (DT_BF16, DT_F16, DT_F32, DT_I32):
+    if dt.dtype in NUMPY_TYPE_TO_DATA_TYPE:
+        raise ValueError(f'Invalid duplicate data type {dt}')
+    NUMPY_TYPE_TO_DATA_TYPE[dt.dtype] = dt
+
+SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
     'BF16': DT_BF16,
     'F16': DT_F16,
     'F32': DT_F32,
@@ -73,20 +119,22 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
 # TODO: rename to LLAMAFileType
 # TODO: move to `gguf.py`
 class GGMLFileType(enum.IntEnum):
-    AllF32    = 0
-    MostlyF16 = 1  # except 1d tensors
+    AllF32     = 0
+    MostlyF16  = 1  # except 1d tensors
+    MostlyQ8_0 = 7  # except 1d tensors
 
-    def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType:
-        if len(tensor.shape) == 1:
-            # 1D tensors are always F32.
-            return DT_F32
-        elif self == GGMLFileType.AllF32:
-            return DT_F32
-        elif self == GGMLFileType.MostlyF16:
-            return DT_F16
-        else:
+    def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType:
+        dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
+        if dt is None:
             raise ValueError(self)
+        # 1D tensors are always F32.
+        return dt if len(tensor.shape) > 1 else DT_F32
 
+GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
+    GGMLFileType.AllF32    : DT_F32,
+    GGMLFileType.MostlyF16 : DT_F16,
+    GGMLFileType.MostlyQ8_0: DT_Q8_0,
+}
 
 #
 # hparams loading
@@ -96,7 +144,6 @@ class GGMLFileType(enum.IntEnum):
 class Params:
     n_vocab:    int
     n_embd:     int
-    n_mult:     int
     n_layer:    int
     n_ctx:      int
     n_ff:       int
@@ -104,19 +151,16 @@ class Params:
     n_head_kv:  int
     f_norm_eps: float
 
-    ftype: Optional[GGMLFileType] = None
+    f_rope_freq_base: float | None = None
+    f_rope_scale: float | None = None
+
+    ftype: GGMLFileType | None = None
+
+    # path to the directory containing the model files
+    path_model: Path | None = None
 
     @staticmethod
-    def find_n_mult(n_ff: int, n_embd: int) -> int:
-        # hardcoded magic range
-        for n_mult in range(8192, 1, -1):
-            calc_ff = (((8*n_embd) // 3 + n_mult - 1) // n_mult)*n_mult
-            if calc_ff == n_ff:
-                return n_mult
-        raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).")
-
-    @staticmethod
-    def guessed(model: 'LazyModel') -> 'Params':
+    def guessed(model: LazyModel) -> Params:
         # try transformer naming first
         n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
 
@@ -142,7 +186,6 @@ class Params:
         return Params(
             n_vocab    = n_vocab,
             n_embd     = n_embd,
-            n_mult     = n_mult,
             n_layer    = n_layer,
             n_ctx      = -1,
             n_ff       = n_ff,
@@ -152,18 +195,23 @@ class Params:
         )
 
     @staticmethod
-    def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab    = config["vocab_size"]
-        n_embd     = config["hidden_size"]
-        n_layer    = config["num_hidden_layers"]
-        n_ff       = config["intermediate_size"]
-        n_head     = config["num_attention_heads"]
-        n_head_kv  = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
-        f_norm_eps = config["rms_norm_eps"]
+        n_vocab          = config["vocab_size"]
+        n_embd           = config["hidden_size"]
+        n_layer          = config["num_hidden_layers"]
+        n_ff             = config["intermediate_size"]
+        n_head           = config["num_attention_heads"]
+        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
+        f_norm_eps       = config["rms_norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
 
-        n_mult = Params.find_n_mult(n_ff, n_embd)
+        rope_scaling = config.get("rope_scaling")
+        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
+            f_rope_scale = config["rope_scaling"].get("factor")
+        else:
+            f_rope_scale = None
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
@@ -174,32 +222,43 @@ class Params:
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
         return Params(
-            n_vocab    = n_vocab,
-            n_embd     = n_embd,
-            n_mult     = n_mult,
-            n_layer    = n_layer,
-            n_ctx      = n_ctx,
-            n_ff       = n_ff,
-            n_head     = n_head,
-            n_head_kv  = n_head_kv,
-            f_norm_eps = f_norm_eps,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
+            f_rope_scale     = f_rope_scale,
         )
 
     # LLaMA v2 70B params.json
-    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
     @staticmethod
-    def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+    def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_vocab    = config["vocab_size"]
-        n_embd     = config["dim"]
-        n_layer    = config["n_layers"]
-        n_mult     = config["multiple_of"]
-        n_ctx      = 2048 if config["norm_eps"] == 1e-06 else 4096 # hack to determine LLaMA v1 vs v2
-        n_ff       = -1
-        n_head     = config["n_heads"]
-        n_head_kv  = config["n_kv_heads"] if "n_kv_heads" in config else n_head
-        f_norm_eps = config["norm_eps"]
+        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
+        n_embd           = config["dim"]
+        n_layer          = config["n_layers"]
+        n_ff             = -1
+        n_head           = config["n_heads"]
+        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
+        f_norm_eps       = config["norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
+
+        # hack to determine LLaMA v1 vs v2 vs CodeLlama
+        if f_rope_freq_base == 1000000:
+            # CodeLlama
+            n_ctx = 16384
+        elif config["norm_eps"] == 1e-05:
+            # LLaMA v2
+            n_ctx = 4096
+        else:
+            # LLaMA v1
+            n_ctx = 2048
 
         if n_vocab == -1:
             n_vocab = model["tok_embeddings.weight"].shape[0]
@@ -208,19 +267,19 @@ class Params:
             n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
 
         return Params(
-            n_vocab    = n_vocab,
-            n_embd     = n_embd,
-            n_mult     = n_mult,
-            n_layer    = n_layer,
-            n_ctx      = n_ctx,
-            n_ff       = n_ff,
-            n_head     = n_head,
-            n_head_kv  = n_head_kv,
-            f_norm_eps = f_norm_eps,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
         )
 
     @staticmethod
-    def load(model_plus: 'ModelPlus') -> 'Params':
+    def load(model_plus: ModelPlus) -> Params:
         hf_config_path   = model_plus.paths[0].parent / "config.json"
         orig_config_path = model_plus.paths[0].parent / "params.json"
 
@@ -228,8 +287,12 @@ class Params:
             params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
         elif orig_config_path.exists():
             params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
-        else:
+        elif model_plus.format != 'none':
             params = Params.guessed(model_plus.model)
+        else:
+            raise ValueError('Cannot guess params when model format is none')
+
+        params.path_model = model_plus.paths[0].parent
 
         return params
 
@@ -239,19 +302,31 @@ class Params:
 #
 
 class BpeVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        added_tokens: Dict[str, int]
+        added_tokens: dict[str, int]
         if fname_added_tokens is not None:
+            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
             added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
-            added_tokens = {}
+            # Fall back to trying to find the added tokens in tokenizer.json
+            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
+            if not tokenizer_json_file.is_file():
+                added_tokens = {}
+            else:
+                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
+                added_tokens = dict(
+                    (item['content'], item['id'])
+                    for item in tokenizer_json.get('added_tokens', [])
+                    # Added tokens here can be duplicates of the main vocabulary.
+                    if item['content'] not in self.bpe_tokenizer )
 
         vocab_size: int = len(self.bpe_tokenizer)
         expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids      = sorted(added_tokens.values())
         if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
+            expected_end_id = vocab_size + len(actual_ids) - 1
+            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_list    = [text for (text, idx) in items]
@@ -260,33 +335,31 @@ class BpeVocab:
         self.fname_tokenizer      = fname_tokenizer
         self.fname_added_tokens   = fname_added_tokens
 
-    def bpe_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.bpe_tokenizer
-        from transformers.models.gpt2 import tokenization_gpt2
-        byte_encoder = tokenization_gpt2.bytes_to_unicode()
-        byte_decoder = {v: k for k, v in byte_encoder.items()}
-        for i, item in enumerate(tokenizer):
-            text: bytes = item.encode("utf-8")
-            score: float = -i
-            yield text, score, gguf.TokenType.USER_DEFINED
+        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
 
-    def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+        for i, _ in enumerate(tokenizer):
+            yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
+
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             score = -1000.0
-            yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
+            yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.bpe_tokens()
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+        return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 class SentencePieceVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: Dict[str, int]
+        added_tokens: dict[str, int]
         if fname_added_tokens is not None:
             added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
@@ -305,7 +378,7 @@ class SentencePieceVocab:
         self.fname_tokenizer = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
-    def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.sentencepiece_tokenizer
         for i in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(i)
@@ -329,20 +402,19 @@ class SentencePieceVocab:
 
             yield text, score, toktype
 
-    def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         for text in self.added_tokens_list:
             score = -1000.0
             yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED
 
-    def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         yield from self.sentencepiece_tokens()
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
-Vocab = Union[BpeVocab, SentencePieceVocab]
-
+Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
 #
 # data loading
@@ -352,7 +424,7 @@ Vocab = Union[BpeVocab, SentencePieceVocab]
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
     #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
-        n_head //= n_head_kv
+        n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
                 .swapaxes(1, 2)
                 .reshape(weights.shape))
@@ -362,18 +434,18 @@ class Tensor(metaclass=ABCMeta):
     data_type: DataType
 
     @abstractmethod
-    def astype(self, data_type: DataType) -> 'Tensor': ...
+    def astype(self, data_type: DataType) -> Tensor: ...
     @abstractmethod
-    def permute(self, n_head: int, n_head_kv: int) -> 'Tensor': ...
+    def permute(self, n_head: int, n_head_kv: int) -> Tensor: ...
     @abstractmethod
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ...
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def part(self, n_part: int) -> 'UnquantizedTensor': ...
+    def part(self, n_part: int) -> UnquantizedTensor: ...
     @abstractmethod
-    def to_ggml(self) -> 'GGMLCompatibleTensor': ...
+    def to_ggml(self) -> GGMLCompatibleTensor: ...
 
 
-def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray:
+def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
     assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}"
     fp32_arr = bf16_arr.astype(np.uint32) << 16
     return fp32_arr.view(np.float32)
@@ -386,27 +458,27 @@ class UnquantizedTensor(Tensor):
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
 
     def astype(self, data_type: DataType) -> Tensor:
-        dtype = DATA_TYPE_TO_NUMPY[data_type]
+        dtype = data_type.dtype
         if self.data_type == DT_BF16:
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
 
-    def to_ggml(self) -> 'UnquantizedTensor':
+    def to_ggml(self) -> UnquantizedTensor:
         return self
 
-    def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor':
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
-        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head))
+        return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
 
-    def part(self, n_part: int) -> 'UnquantizedTensor':
+    def part(self, n_part: int) -> UnquantizedTensor:
         r = self.ndarray.shape[0] // 3
         return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...])
 
-    def permute(self, n_head: int, n_head_kv: int) -> 'UnquantizedTensor':
+    def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor:
         return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv))
 
 
-def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray:
+def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray:
     tensor = lazy_tensor.load()
     assert isinstance(tensor, UnquantizedTensor)
 
@@ -422,38 +494,24 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv
     return tensor.ndarray
 
 
-GGMLCompatibleTensor = Union[UnquantizedTensor]
-
-
-class DeferredPermutedTensor(Tensor):
-    def __init__(self, base: Tensor, n_head: int, n_head_kv: int) -> None:
-        self.base = base
-        self.n_head = n_head
-        self.data_type = self.base.data_type
-
-    def astype(self, data_type: DataType) -> Tensor:
-        return self.base.astype(data_type).permute(self.n_head, self.n_head_kv)
-
-    def to_ggml(self) -> GGMLCompatibleTensor:
-        return self.base.to_ggml().permute(self.n_head, self.n_head_kv)
-
-    def permute(self, n_head: int, n_head_kv: int) -> Tensor:
-        raise Exception("shouldn't permute twice")
+GGMLCompatibleTensor = UnquantizedTensor
 
 
 @dataclass
 class LazyTensor:
     _load: Callable[[], Tensor]
-    shape: List[int]
+    shape: list[int]
     data_type: DataType
     description: str
 
     def load(self) -> Tensor:
         ret = self._load()
-        assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description)
+        # Should be okay if it maps to the same numpy type?
+        assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
+                (self.data_type, ret.data_type, self.description)
         return ret
 
-    def astype(self, data_type: DataType) -> 'LazyTensor':
+    def astype(self, data_type: DataType) -> LazyTensor:
         self.validate_conversion_to(data_type)
 
         def load() -> Tensor:
@@ -461,28 +519,28 @@ class LazyTensor:
         return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}')
 
     def validate_conversion_to(self, data_type: DataType) -> None:
-        if data_type == self.data_type:
-            return
+        if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions:
+            raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.')
 
 
-LazyModel = Dict[str, LazyTensor]
+LazyModel: TypeAlias = 'dict[str, LazyTensor]'
 
 
 @dataclass
 class ModelPlus:
     model: LazyModel
-    paths: List[Path]  # Where this was read from.
-    format: Literal['ggml', 'torch', 'safetensors']
-    vocab: Optional[Vocab]  # For GGML models (which have vocab built in), the vocab.
+    paths: list[Path]  # Where this was read from.
+    format: Literal['ggml', 'torch', 'safetensors', 'none']
+    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
-def merge_sharded(models: List[LazyModel]) -> LazyModel:
+def merge_sharded(models: list[LazyModel]) -> LazyModel:
     # Original LLaMA models have each file contain one part of each tensor.
     # Use a dict instead of a set to preserve order.
     names = {name: None for model in models for name in model}
 
     def convert(name: str) -> LazyTensor:
-        lazy_tensors: List[LazyTensor] = [model[name] for model in models]
+        lazy_tensors: list[LazyTensor] = [model[name] for model in models]
         if len(lazy_tensors) == 1:
             # only one file; don't go through this procedure since there might
             # be quantized tensors
@@ -510,7 +568,7 @@ def merge_sharded(models: List[LazyModel]) -> LazyModel:
     return {name: convert(name) for name in names}
 
 
-def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus:
+def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     formats = set(mp.format for mp in models_plus)
     assert len(formats) == 1, "different formats?"
     format = formats.pop()
@@ -538,12 +596,12 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe
         return lazy_tensor.load().permute(n_head, n_head_kv)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor:
+def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
-        return lazy_tensor.load().permute_part(n_part, n_head)
+        return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
     s = lazy_tensor.shape.copy()
     s[0] = s[0] // 3
-    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description)
+    return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
@@ -584,13 +642,11 @@ class LazyUnpickler(pickle.Unpickler):
         assert isinstance(pid[1], LazyStorageKind)
         data_type = pid[1].data_type
         filename_stem = pid[2]
-        filename = self.data_base_path + '/' + filename_stem
+        filename = f'{self.data_base_path}/{filename_stem}'
         info = self.zip_file.getinfo(filename)
 
         def load(offset: int, elm_count: int) -> NDArray:
-            dtype = DATA_TYPE_TO_NUMPY.get(data_type)
-            if dtype is None:
-                raise Exception("tensor stored in unsupported format")
+            dtype = data_type.dtype
             fp = self.zip_file.open(info)
             fp.seek(offset * dtype.itemsize)
             size = elm_count * dtype.itemsize
@@ -600,9 +656,8 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
         return LazyStorage(load=load, kind=pid[1], description=description)
 
-    # @staticmethod
+    @staticmethod
     def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
-                               # pyright: ignore[reportSelfClsParameterName]
                                requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
         assert isinstance(storage, LazyStorage)
 
@@ -612,13 +667,15 @@ class LazyUnpickler(pickle.Unpickler):
         description = f'pickled storage_offset={storage_offset} in {storage.description}'
         return LazyTensor(load, list(size), storage.kind.data_type, description)
 
-    # @staticmethod
+    @staticmethod
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: Dict[Any, Any] = {
-        ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2,
-        ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2,
+    CLASSES: dict[tuple[str, str], Any] = {
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
+        # the staticmethods have a __func__ attribute.
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
         ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
         ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
         ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
@@ -647,15 +704,15 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
 
 def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
     header_size, = struct.unpack('<Q', fp.read(8))
-    header: Dict[str, Dict[str, Any]] = json.loads(fp.read(header_size))
+    header: dict[str, dict[str, Any]] = json.loads(fp.read(header_size))
     # Use mmap for the actual data to avoid race conditions with the file offset.
     mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
     byte_buf = mapped[8 + header_size:]
 
-    def convert(info: Dict[str, Any]) -> LazyTensor:
+    def convert(info: dict[str, Any]) -> LazyTensor:
         data_type = SAFETENSORS_DATA_TYPES[info['dtype']]
-        numpy_dtype = DATA_TYPE_TO_NUMPY[data_type]
-        shape: List[int] = info['shape']
+        numpy_dtype = data_type.dtype
+        shape: list[int] = info['shape']
         begin, end = info['data_offsets']
         assert 0 <= begin <= end <= len(byte_buf)
         assert end - begin == math.prod(shape) * numpy_dtype.itemsize
@@ -694,23 +751,40 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int) -> Iterable[Out]:
+def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
     letting results pile up in memory.  Specifically, there is a max of one
     output value buffered per thread.'''
-    with concurrent.futures.ThreadPoolExecutor() as executor:
-        futures: List[concurrent.futures.Future[Out]] = []
-        items_rev = list(iterable)[::-1]
-        for i in range(min(concurrency, len(items_rev))):
-            futures.append(executor.submit(func, items_rev.pop()))
+    if concurrency < 2:
+        yield from map(func, iterable)
+        # Not reached.
+    iterable = iter(iterable)
+    executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor]
+    if use_processpool_executor:
+        executor_class = ProcessPoolExecutor
+    else:
+        executor_class = ThreadPoolExecutor
+    with executor_class(max_workers = max_workers) as executor:
+        futures: list[concurrent.futures.Future[Out]] = []
+        done = False
+        for _ in range(concurrency):
+            try:
+                futures.append(executor.submit(func, next(iterable)))
+            except StopIteration:
+                done = True
+                break
+
         while futures:
             result = futures.pop(0).result()
-            if items_rev:
-                futures.append(executor.submit(func, items_rev.pop()))
+            while not done and len(futures) < concurrency:
+                try:
+                    futures.append(executor.submit(func, next(iterable)))
+                except StopIteration:
+                    done = True
+                    break
             yield result
 
-
 def check_vocab_size(params: Params, vocab: Vocab) -> None:
     if params.n_vocab != vocab.vocab_size:
         assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
@@ -733,7 +807,15 @@ class OutputFile:
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
     def add_meta_arch(self, params: Params) -> None:
-        self.gguf.add_name                ("LLaMA")
+        name = "LLaMA"
+
+        # TODO: better logic to determine model name
+        if params.n_ctx == 4096:
+            name = "LLaMA v2"
+        elif params.path_model is not None:
+            name = str(params.path_model.parent).split('/')[-1]
+
+        self.gguf.add_name                (name)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -743,32 +825,44 @@ class OutputFile:
         self.gguf.add_head_count_kv       (params.n_head_kv)
         self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
 
-        if params.ftype:
+        if params.f_rope_freq_base is not None:
+            self.gguf.add_rope_freq_base(params.f_rope_freq_base)
+
+        if params.f_rope_scale is not None:
+            self.gguf.add_rope_scale_linear(params.f_rope_scale)
+
+        if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
         tokens = []
         scores = []
         toktypes = []
-        # NOTE: `all_tokens` returns the the base vocabulary and added tokens
-        # TODO: add special tokens?
+        # NOTE: `all_tokens` returns the base vocabulary and added tokens
         for text, score, toktype in vocab.all_tokens():
             tokens.append(text)
             scores.append(score)
             toktypes.append(toktype)
 
-        self.gguf.add_tokenizer_model("llama")
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
 
+    def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None:
+        svocab.add_to_gguf(self.gguf)
+
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
-        n_elements = 1
-        for dim in tensor.shape:
-            n_elements *= dim
-        data_type = DATA_TYPE_TO_NUMPY[tensor.data_type]
-        data_nbytes = n_elements * data_type.itemsize
-        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes)
+        n_elements = int(np.prod(tensor.shape))
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
+        data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype)
 
     def write_meta(self) -> None:
         self.gguf.write_header_to_file()
@@ -781,7 +875,7 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab) -> None:
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -789,12 +883,27 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
+
         of.write_meta()
 
         of.close()
 
     @staticmethod
-    def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None:
+    def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]:
+        name, lazy_tensor = item
+        tensor = lazy_tensor.load().to_ggml()
+        return (lazy_tensor.data_type, tensor.ndarray)
+
+    @staticmethod
+    def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray:
+        dt, arr = item
+        if not isinstance(dt, QuantizedDataType):
+            return arr
+        return dt.quantize(arr)
+
+    @staticmethod
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
         check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out)
@@ -802,6 +911,7 @@ class OutputFile:
         # meta data
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -810,27 +920,32 @@ class OutputFile:
         of.write_meta()
         of.write_tensor_info()
 
-        def do_item(item: Tuple[str, LazyTensor]) -> NDArray:
-            name, lazy_tensor = item
-            return lazy_tensor.load().to_ggml().ndarray
-
         # tensor data
-        ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8)
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True)
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
         for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}")
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
             of.gguf.write_tensor_data(ndarray)
 
         of.close()
 
-def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType:
-    wq_type = model[NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
+def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
     if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
         return GGMLFileType.MostlyF16
+    if output_type_str == "q8_0":
+        return GGMLFileType.MostlyQ8_0
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
@@ -841,7 +956,8 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
-    tmap = gguf.get_tensor_name_map(ARCH, params.n_layer)
+    tmap = gguf.TensorNameMap(ARCH, params.n_layer)
+    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
@@ -857,37 +973,31 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
             tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
             tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy        (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
         else:
             break
 
     out: LazyModel = {}
     for name, lazy_tensor in model.items():
-        name_new = name
-
-        if name in tmap:
-            name_new = tmap[name]
-        elif name.endswith(".weight") and name[:-7] in tmap:
-            name_new = tmap[name[:-7]] + ".weight"
-        elif name.endswith(".bias") and name[:-5] in tmap:
-            name_new = tmap[name[:-5]] + ".bias"
-        else:
+        tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
+        if name_new is None:
             raise Exception(f"Unexpected tensor name: {name}")
 
-        if gguf.should_skip_tensor_TMP(ARCH, params.n_layer, name_new):
+        if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
             continue
-        else:
-            print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type} | {lazy_tensor.shape}")
-            out[name_new] = lazy_tensor
+
+        print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
+        out[name_new] = lazy_tensor
 
     return out
 
-def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
+def nth_multifile_path(path: Path, n: int) -> Path | None:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns: List[Tuple[str, str]] = [
+    patterns: list[tuple[str, str]] = [
         # - x.00.pth, x.01.pth, etc.
         (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
         # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
@@ -903,11 +1013,11 @@ def nth_multifile_path(path: Path, n: int) -> Optional[Path]:
     return None
 
 
-def find_multifile_paths(path: Path) -> List[Path]:
+def find_multifile_paths(path: Path) -> list[Path]:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the whole list of paths in the model.
     '''
-    ret: List[Path] = []
+    ret: list[Path] = []
     for i in itertools.count():
         nth_path = nth_multifile_path(path, i)
         if nth_path is None:
@@ -938,7 +1048,7 @@ def load_some_model(path: Path) -> ModelPlus:
         path = files[0]
 
     paths = find_multifile_paths(path)
-    models_plus: List[ModelPlus] = []
+    models_plus: list[ModelPlus] = []
     for path in paths:
         print(f"Loading model file {path}")
         models_plus.append(lazy_load_file(path))
@@ -947,7 +1057,7 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, SentencePieceVocab]:
+def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
     # Be extra-friendly and accept either a file or a directory.  Also, if it's
     # a directory, it might be the model directory, and tokenizer.model might
     # be in the parent of that.
@@ -978,10 +1088,11 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence
         raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
 
-def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path:
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
     namestr = {
         GGMLFileType.AllF32:    "f32",
         GGMLFileType.MostlyF16: "f16",
+        GGMLFileType.MostlyQ8_0:"q8_0",
     }[file_type]
     ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
     if ret in model_paths:
@@ -1000,24 +1111,33 @@ def do_dump_model(model_plus: ModelPlus) -> None:
         print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}")
 
 
-def main(args_in: Optional[List[str]] = None) -> None:
+def main(args_in: list[str] | None = None) -> None:
     parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
     parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
-    parser.add_argument("--outtype",     choices=["f32", "f16"], help="output format (default: based on input)")
+    parser.add_argument("--outtype",     choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
     parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     parser.add_argument("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
     args = parser.parse_args(args_in)
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
+        return
 
-    model_plus = load_some_model(args.model)
+    if not args.vocab_only:
+        model_plus = load_some_model(args.model)
+    else:
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
+
+    if args.dump:
+        do_dump_model(model_plus)
+        return
 
     params = Params.load(model_plus)
     if params.n_ctx == -1:
@@ -1032,39 +1152,41 @@ def main(args_in: Optional[List[str]] = None) -> None:
         params.ftype = {
             "f32": GGMLFileType.AllF32,
             "f16": GGMLFileType.MostlyF16,
+            "q8_0": GGMLFileType.MostlyQ8_0,
         }[args.outtype]
 
     print(f"params = {params}")
 
     vocab: Vocab
     if args.vocab_only:
-        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
         assert args.outfile, "need --outfile if using --vocab-only"
+        # FIXME: Try to respect vocab_dir somehow?
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
+        return
+
+    if model_plus.vocab is not None and args.vocab_dir is None:
+        vocab = model_plus.vocab
     else:
-        if args.dump:
-            do_dump_model(model_plus)
-            return
+        vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
+        vocab = load_vocab(vocab_dir, args.vocabtype)
+    # FIXME: Try to respect vocab_dir somehow?
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
 
-        if model_plus.vocab is not None and args.vocab_dir is None:
-            vocab = model_plus.vocab
-        else:
-            vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
-            vocab = load_vocab(vocab_dir, args.vocabtype)
+    model   = model_plus.model
+    model   = convert_model_names(model, params)
+    ftype   = pick_output_type(model, args.outtype)
+    model   = convert_to_output_type(model, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
 
-        model   = model_plus.model
-        model   = convert_model_names(model, params)
-        ftype   = pick_output_type(model, args.outtype)
-        model   = convert_to_output_type(model, ftype)
-        outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    params.ftype = ftype
+    print(f"Writing {outfile}, format {ftype}")
 
-        params.ftype = ftype
-        print(f"Writing {outfile}, format {ftype}")
-
-        OutputFile.write_all(outfile, params, model, vocab)
-        print(f"Wrote {outfile}")
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
+    print(f"Wrote {outfile}")
 
 
 if __name__ == '__main__':
diff --git a/docs/BLIS.md b/docs/BLIS.md
index 9b3c30605..f3d2312b4 100644
--- a/docs/BLIS.md
+++ b/docs/BLIS.md
@@ -48,7 +48,7 @@ make -j
 According to the BLIS documentation, we could set the following
 environment variables to modify the behavior of openmp:
 
-```
+```bash
 export GOMP_GPU_AFFINITY="0-19"
 export BLIS_NUM_THREADS=14
 ```
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index d2176c910..de4cf7a69 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -21,14 +21,20 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(finetune)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(simple)
+    add_subdirectory(batched)
+    add_subdirectory(speculative)
+    add_subdirectory(parallel)
     add_subdirectory(embd-input)
     add_subdirectory(llama-bench)
+    add_subdirectory(beam-search)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()
+    add_subdirectory(export-lora)
 endif()
diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp
index 6fa55b319..8155101d0 100644
--- a/examples/baby-llama/baby-llama.cpp
+++ b/examples/baby-llama/baby-llama.cpp
@@ -1,43 +1,24 @@
 #include "ggml.h"
+#include "train.h"
+
 #include <vector>
 #include <cassert>
-#include <random>
+#include <cstdlib>
 #include <cstring>
+#include <random>
+#include <vector>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #ifdef LLAMA_DEFAULT_RMS_EPS
-static const float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
+constexpr float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS;
 #else
-static const float rms_norm_eps = 5e-6f;
+constexpr float rms_norm_eps = 5e-6f;
 #endif
 
-float frand() {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-struct random_normal_distribution {
-    std::mt19937 gen;
-    std::normal_distribution<float> nd;
-    float min;
-    float max;
-};
-
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
-    rnd->gen = std::mt19937(seed);
-    rnd->nd = std::normal_distribution<float>{mean, std};
-    rnd->min = min;
-    rnd->max = max;
-}
-
-float frand_normal(struct random_normal_distribution * rnd) {
-    const float r = rnd->nd(rnd->gen);
-    return ((r < rnd->min) ? (rnd->min) : (r > rnd->max) ? (rnd->max) : r);
-}
-
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -48,13 +29,9 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-struct ggml_tensor * randomize_tensor(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        float fmin,
-        float fmax) {
-
+static struct ggml_tensor * randomize_tensor(
+    struct ggml_tensor * tensor, int ndims, const int64_t ne[], float fmin, float fmax
+) {
     switch (ndims) {
         case 1:
             for (int i0 = 0; i0 < ne[0]; i0++) {
@@ -90,57 +67,7 @@ struct ggml_tensor * randomize_tensor(
             break;
         default:
             assert(false);
-    };
-
-    return tensor;
-}
-
-struct ggml_tensor * randomize_tensor_normal(
-        struct ggml_tensor * tensor,
-        int ndims,
-        const int64_t ne[],
-        struct random_normal_distribution * rnd) {
-    float scale = 1.0; // xavier
-    switch (ndims) {
-        case 1:
-            scale /= sqrtf(ne[0]);
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
-            }
-            break;
-        case 2:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
-                }
-            }
-            break;
-        case 3:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            scale /= sqrtf(ne[0]+ne[1]);
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
+    }
 
     return tensor;
 }
@@ -159,7 +86,7 @@ struct llama_hparams {
     }
 };
 
-uint32_t get_n_ff(const struct llama_hparams* hparams) {
+static uint32_t get_n_ff(const struct llama_hparams* hparams) {
     const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
     return n_ff;
 }
@@ -260,7 +187,7 @@ struct llama_model_lora {
     std::vector<llama_layer_lora> layers;
 };
 
-void init_model(struct llama_model * model) {
+static void init_model(struct llama_model * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -297,7 +224,7 @@ void init_model(struct llama_model * model) {
 }
 
 
-void init_model_lora(struct llama_model_lora * model) {
+static void init_model_lora(struct llama_model_lora * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
@@ -340,7 +267,7 @@ void init_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void set_param_model(struct llama_model * model) {
+static void set_param_model(struct llama_model * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -366,7 +293,7 @@ void set_param_model(struct llama_model * model) {
     }
 }
 
-void set_param_model_lora(struct llama_model_lora * model) {
+static void set_param_model_lora(struct llama_model_lora * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -397,69 +324,109 @@ void set_param_model_lora(struct llama_model_lora * model) {
     }
 }
 
-void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model(struct llama_model * model, int seed, float mean, float std, float min, float max) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->output,         model->output->n_dims,         model->output->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings , rnd);
+    randomize_tensor_normal(model->norm           , rnd);
+    randomize_tensor_normal(model->output         , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wq, layer.wq->n_dims, layer.wq->ne, &rnd);
-        randomize_tensor_normal(layer.wk, layer.wk->n_dims, layer.wk->ne, &rnd);
-        randomize_tensor_normal(layer.wv, layer.wv->n_dims, layer.wv->ne, &rnd);
-        randomize_tensor_normal(layer.wo, layer.wo->n_dims, layer.wo->ne, &rnd);
+        randomize_tensor_normal(layer.wq, rnd);
+        randomize_tensor_normal(layer.wk, rnd);
+        randomize_tensor_normal(layer.wv, rnd);
+        randomize_tensor_normal(layer.wo, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
 
-void randomize_model_lora(struct llama_model_lora * model, int seed, float mean, float std, float min, float max) {
+static void randomize_model_lora(
+    struct llama_model_lora * model, int seed, float mean, float std, float min, float max
+) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
-    randomize_tensor_normal(model->tok_embeddings, model->tok_embeddings->n_dims, model->tok_embeddings->ne, &rnd);
-    randomize_tensor_normal(model->norm,           model->norm->n_dims,           model->norm->ne,           &rnd);
-    randomize_tensor_normal(model->outputa,        model->outputa->n_dims,        model->outputa->ne,         &rnd);
-    randomize_tensor_normal(model->outputb,        model->outputb->n_dims,        model->outputb->ne,         &rnd);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(model->tok_embeddings, rnd);
+    randomize_tensor_normal(model->norm          , rnd);
+    randomize_tensor_normal(model->outputa       , rnd);
+    randomize_tensor_normal(model->outputb       , rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, layer.attention_norm->n_dims, layer.attention_norm->ne, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wqa, layer.wqa->n_dims, layer.wqa->ne, &rnd);
-        randomize_tensor_normal(layer.wqb, layer.wqb->n_dims, layer.wqb->ne, &rnd);
-        randomize_tensor_normal(layer.wka, layer.wka->n_dims, layer.wka->ne, &rnd);
-        randomize_tensor_normal(layer.wkb, layer.wkb->n_dims, layer.wkb->ne, &rnd);
-        randomize_tensor_normal(layer.wva, layer.wva->n_dims, layer.wva->ne, &rnd);
-        randomize_tensor_normal(layer.wvb, layer.wvb->n_dims, layer.wvb->ne, &rnd);
-        randomize_tensor_normal(layer.woa, layer.woa->n_dims, layer.woa->ne, &rnd);
-        randomize_tensor_normal(layer.wob, layer.wob->n_dims, layer.wob->ne, &rnd);
+        randomize_tensor_normal(layer.wqa, rnd);
+        randomize_tensor_normal(layer.wqb, rnd);
+        randomize_tensor_normal(layer.wka, rnd);
+        randomize_tensor_normal(layer.wkb, rnd);
+        randomize_tensor_normal(layer.wva, rnd);
+        randomize_tensor_normal(layer.wvb, rnd);
+        randomize_tensor_normal(layer.woa, rnd);
+        randomize_tensor_normal(layer.wob, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, layer.ffn_norm->n_dims, layer.ffn_norm->ne, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, layer.w1->n_dims, layer.w1->ne, &rnd);
-        randomize_tensor_normal(layer.w2, layer.w2->n_dims, layer.w2->ne, &rnd);
-        randomize_tensor_normal(layer.w3, layer.w3->n_dims, layer.w3->ne, &rnd);
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
+
+    free_random_normal_distribution(rnd);
 }
 
-bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+static void init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int n_batch) {
+    const auto & hparams = model->hparams;
+
+    const uint32_t n_ctx   = hparams.n_ctx;
+    const uint32_t n_embd  = hparams.n_embd;
+    const uint32_t n_layer = hparams.n_layer;
+
+    const int64_t n_mem      = n_layer*n_ctx*n_batch;
+    const int64_t n_elements = n_embd*n_mem;
+
+    // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
+
+    // struct ggml_init_params params;
+    // params.mem_size   = cache.buf.size;
+    // params.mem_buffer = cache.buf.addr;
+    // params.no_alloc   = false;
+    if (!cache->ctx) {
+        struct ggml_init_params params;
+        params.mem_size   = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        cache->ctx = ggml_init(params);
+
+        if (!cache->ctx) {
+            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
+            exit(1);
+        }
+    }
+
+    cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
+    cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
+}
+
+static bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_ctx   = hparams.n_ctx;
@@ -495,51 +462,15 @@ bool init_kv_cache(struct llama_kv_cache* cache, struct llama_model * model, int
     return true;
 }
 
-bool init_kv_cache_lora(struct llama_kv_cache* cache, struct llama_model_lora * model, int n_batch) {
-    const auto & hparams = model->hparams;
-
-    const uint32_t n_ctx   = hparams.n_ctx;
-    const uint32_t n_embd  = hparams.n_embd;
-    const uint32_t n_layer = hparams.n_layer;
-
-    const int64_t n_mem      = n_layer*n_ctx*n_batch;
-    const int64_t n_elements = n_embd*n_mem;
-
-    // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
-
-    // struct ggml_init_params params;
-    // params.mem_size   = cache.buf.size;
-    // params.mem_buffer = cache.buf.addr;
-    // params.no_alloc   = false;
-    if (!cache->ctx) {
-        struct ggml_init_params params;
-        params.mem_size   = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
-        params.mem_buffer = NULL;
-        params.no_alloc   = false;
-
-        cache->ctx = ggml_init(params);
-
-        if (!cache->ctx) {
-            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-            return false;
-        }
-    }
-
-    cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-    cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-
-    return true;
-}
-
-struct ggml_tensor * forward(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past) {
-
+static struct ggml_tensor * forward(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -556,6 +487,14 @@ struct ggml_tensor * forward(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -583,8 +522,8 @@ struct ggml_tensor * forward(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, 1]
             // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -756,42 +695,16 @@ struct ggml_tensor * forward(
     return inpL;
 }
 
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-}
-
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-}
-
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-}
-
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == ne3);
-}
-
-struct ggml_tensor * forward_batch(
-        struct llama_model    * model,
-        struct llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past,
-        const  int              n_batch) {
-
+static struct ggml_tensor * forward_batch(
+    struct llama_model    * model,
+    struct llama_kv_cache * cache,
+    struct ggml_context   * ctx0,
+    struct ggml_cgraph    * gf,
+    struct ggml_tensor    * tokens_input,
+    const  int              n_tokens,
+    const  int              n_past,
+    const  int              n_batch
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -810,9 +723,18 @@ struct ggml_tensor * forward_batch(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N*n_batch,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     assert_shape_2d(inpL, n_embd, N*n_batch);
+
     for (int il = 0; il < n_layer; ++il) {
         struct ggml_tensor * inpSA = inpL;
 
@@ -840,8 +762,8 @@ struct ggml_tensor * forward_batch(
             // wk   shape [n_embd, n_embd, 1, 1]
             // Qcur shape [n_embd/n_head, n_head, N, n_batch]
             // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), KQ_pos, n_rot, 0, 0);
             assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
             assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
 
@@ -1073,16 +995,15 @@ struct ggml_tensor * forward_batch(
     return inpL;
 }
 
-
-struct ggml_tensor * forward_lora(
-        struct llama_model_lora * model,
-        struct llama_kv_cache   * cache,
-        struct ggml_context     * ctx0,
-        struct ggml_cgraph      * gf,
-        struct ggml_tensor      * tokens_input,
-        const  int                n_tokens,
-        const  int                n_past) {
-
+static struct ggml_tensor * forward_lora(
+    struct llama_model_lora * model,
+    struct llama_kv_cache   * cache,
+    struct ggml_context     * ctx0,
+    struct ggml_cgraph      * gf,
+    struct ggml_tensor      * tokens_input,
+    const  int                n_tokens,
+    const  int                n_past
+) {
     const int N = n_tokens;
 
     struct llama_kv_cache& kv_self = *cache;
@@ -1100,6 +1021,14 @@ struct ggml_tensor * forward_lora(
     struct ggml_tensor * kc = kv_self.k;
     struct ggml_tensor * vc = kv_self.v;
 
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
     // inpL shape [n_embd,N,1,1]
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
     for (int il = 0; il < n_layer; ++il) {
@@ -1133,7 +1062,7 @@ struct ggml_tensor * forward_lora(
                                                         model->layers[il].wqb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
             struct ggml_tensor * Kcur = ggml_rope(ctx0,
                                             ggml_reshape_3d(ctx0,
                                                 ggml_mul_mat(ctx0,
@@ -1142,7 +1071,7 @@ struct ggml_tensor * forward_lora(
                                                         model->layers[il].wkb,
                                                         cur)),
                                                 n_embd/n_head, n_head, N),
-                                            n_past, n_rot, 0, 0);
+                                            KQ_pos, n_rot, 0, 0);
 
             // store key and value to memory
             {
@@ -1328,7 +1257,7 @@ struct ggml_tensor * forward_lora(
     return inpL;
 }
 
-void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
     assert(logits->n_dims == 2);
     assert(probs->n_dims == 2);
     assert(best_samples->n_dims == 1);
@@ -1359,7 +1288,10 @@ void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, str
     }
 }
 
-void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
+static void sample_softmax_batch(
+    struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
+    struct ggml_tensor * best_samples
+) {
     GGML_ASSERT(best_samples->n_dims == 2);
     GGML_ASSERT(logits->n_dims == 3);
     GGML_ASSERT(probs->n_dims == 3);
@@ -1393,7 +1325,7 @@ void sample_softmax_batch(struct ggml_context * ctx, struct ggml_tensor * logits
     }
 }
 
-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
         printf(" %.2f", p);
@@ -1401,7 +1333,7 @@ void print_row(struct ggml_tensor * probs, int i) {
     printf("\n");
 }
 
-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
     assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
@@ -1412,7 +1344,7 @@ void print_matrix(struct ggml_tensor * probs) {
     }
 }
 
-void print_token(int token, int n_vocab) {
+static void print_token(int token, int n_vocab) {
     for (int k = 0; k < token; ++k) {
         printf(" ");
     }
@@ -1423,14 +1355,14 @@ void print_token(int token, int n_vocab) {
     printf("\n");
 }
 
-void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
+static void print_tokens(struct ggml_tensor * tokens, int n_vocab) {
     for (int i=0; i<tokens->ne[0]; ++i) {
         int token = ggml_get_i32_1d(tokens, i);
         print_token(token, n_vocab);
     }
 }
 
-void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets(int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     float randomness = 0.0f;
@@ -1451,7 +1383,9 @@ void get_example_targets(int example_id, struct ggml_tensor * tokens_input, stru
     }
 }
 
-void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets) {
+static void get_example_targets_batch(
+    struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
+) {
     GGML_ASSERT(tokens_input->n_dims == 2);
     GGML_ASSERT(     targets->n_dims == 3);
     int n_tokens = tokens_input->ne[0];
@@ -1474,7 +1408,7 @@ void get_example_targets_batch(struct ggml_context * ctx, int example_id, struct
     }
 }
 
-void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
+static void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * targets, int n_shift) {
     int n_tokens = tokens_input->ne[0];
     int n_vocab = targets->ne[0];
     for (int i=0; i<n_tokens-n_shift; ++i) {
@@ -1485,12 +1419,16 @@ void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * tar
     }
 }
 
-struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * square_error_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     // todo: instead of a-b: a[1:]-b[:-1]
     return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, a, b)));
 }
 
-struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+static struct ggml_tensor * cross_entropy_loss(
+    struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b
+) {
     const float eps = 1e-3f;
     return
         ggml_sum(ctx,
@@ -1617,15 +1555,10 @@ int main(int argc, char ** argv) {
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
         struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
-        opt_params_adam.print_forward_graph = false;
-        opt_params_adam.print_backward_graph = false;
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
-        opt_params_adam.adam.n_iter = 16;
         opt_params_lbfgs.lbfgs.n_iter = 16;
-        // ggml_opt(ctx0, opt_params_adam, e);
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
         ggml_build_forward_expand(&gf, e);
diff --git a/examples/batched/CMakeLists.txt b/examples/batched/CMakeLists.txt
new file mode 100644
index 000000000..6aa178d4d
--- /dev/null
+++ b/examples/batched/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET batched)
+add_executable(${TARGET} batched.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/batched/README.md b/examples/batched/README.md
new file mode 100644
index 000000000..5d7303317
--- /dev/null
+++ b/examples/batched/README.md
@@ -0,0 +1,44 @@
+# llama.cpp/example/batched
+
+The example demonstrates batched generation from a given prompt
+
+```bash
+./batched ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is" 4
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 4, n_kv_req = 113
+
+ Hello my name is
+
+main: generating 4 sequences ...
+
+main: stream 0 finished
+main: stream 1 finished
+main: stream 2 finished
+main: stream 3 finished
+
+sequence 0:
+
+Hello my name is Shirley. I am a 25-year-old female who has been working for over 5 years as a b
+
+sequence 1:
+
+Hello my name is Renee and I'm a 32 year old female from the United States. I'm looking for a man between
+
+sequence 2:
+
+Hello my name is Diana. I am looking for a housekeeping job. I have experience with children and have my own transportation. I am
+
+sequence 3:
+
+Hello my name is Cody. I am a 3 year old neutered male. I am a very friendly cat. I am very playful and
+
+main: decoded 108 tokens in 3.57 s, speed: 30.26 t/s
+
+llama_print_timings:        load time =   587.00 ms
+llama_print_timings:      sample time =     2.56 ms /   112 runs   (    0.02 ms per token, 43664.72 tokens per second)
+llama_print_timings: prompt eval time =  4089.11 ms /   118 tokens (   34.65 ms per token,    28.86 tokens per second)
+llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+llama_print_timings:       total time =  4156.04 ms
+```
diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp
new file mode 100644
index 000000000..688ef2213
--- /dev/null
+++ b/examples/batched/batched.cpp
@@ -0,0 +1,255 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
+        return 1 ;
+    }
+
+    int n_parallel = 1;
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        params.prompt = argv[2];
+    }
+
+    if (argc >= 4) {
+        n_parallel = std::atoi(argv[3]);
+    }
+
+    if (params.prompt.empty()) {
+        params.prompt = "Hello my name is";
+    }
+
+    // total length of the sequences including the prompt
+    const int n_len = 32;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = n_kv_req;
+    ctx_params.n_batch = std::max(n_len, n_parallel);
+    ctx_params.n_threads = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    const int n_ctx    = llama_n_ctx(ctx);
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        return 1;
+    }
+
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
+
+    for (auto id : tokens_list) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+    }
+
+    fflush(stderr);
+
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
+    // assign the system KV cache to all parallel sequences
+    // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
+    for (int32_t i = 1; i < n_parallel; ++i) {
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
+    }
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+    }
+
+    // main loop
+
+    // we will store the parallel decoded sequences in this vector
+    std::vector<std::string> streams(n_parallel);
+
+    // remember the batch index of the last token for each parallel sequence
+    // we need this to determine which logits to sample from
+    std::vector<int32_t> i_batch(n_parallel, batch.n_tokens - 1);
+
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // prepare the next batch
+        batch.n_tokens = 0;
+
+        // sample the next token for each parallel sequence / stream
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            if (i_batch[i] < 0) {
+                // the stream has already finished
+                continue;
+            }
+
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            const int   top_k = 40;
+            const float top_p = 0.9f;
+            const float temp  = 0.4f;
+
+            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temp (ctx, &candidates_p, temp);
+
+            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
+
+            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream? -> mark the stream as finished
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                i_batch[i] = -1;
+                LOG_TEE("\n");
+                if (n_parallel > 1) {
+                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                }
+
+                continue;
+            }
+
+            // if there is only one stream, we print immediately to stdout
+            if (n_parallel == 1) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                fflush(stdout);
+            }
+
+            streams[i] += llama_token_to_piece(ctx, new_token_id);
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = i;
+            batch.logits[batch.n_tokens] = true;
+
+            i_batch[i] = batch.n_tokens;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        // all streams are finished
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    if (n_parallel > 1) {
+        LOG_TEE("\n");
+
+        for (int32_t i = 0; i < n_parallel; ++i) {
+            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}
diff --git a/examples/beam-search/CMakeLists.txt b/examples/beam-search/CMakeLists.txt
new file mode 100644
index 000000000..f0e37468b
--- /dev/null
+++ b/examples/beam-search/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET beam-search)
+add_executable(${TARGET} beam-search.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/beam-search/beam-search.cpp b/examples/beam-search/beam-search.cpp
new file mode 100644
index 000000000..f078ab8a8
--- /dev/null
+++ b/examples/beam-search/beam-search.cpp
@@ -0,0 +1,187 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+// Used for debugging to print out beam tokens.
+struct ostream_beam_view {
+    llama_context * ctx;
+    llama_beam_view beam_view;
+};
+
+static std::ostream & operator<<(std::ostream & os, const ostream_beam_view & obv) {
+    os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens(";
+    for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) {
+        os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]);
+    }
+    return os << ')';
+}
+
+// Put here anything you want back in beam_search_callback().
+struct beam_search_callback_data {
+    llama_context * ctx;
+    std::vector<llama_token> response;
+};
+
+// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
+// For example, eob can be flagged due to maximum token length, stop words, etc.
+static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+static void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) {
+    auto& callback_data = *static_cast<beam_search_callback_data*>(callback_data_ptr);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        callback_data.response.resize(callback_data.response.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        std::copy(tokens, tokens + n, callback_data.response.end() - n);
+        printf("%zu", n);
+    }
+    fflush(stdout);
+#if 1 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n";
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{callback_data.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+int main(int argc, char ** argv)
+{
+    gpt_params params;
+    //params.n_gpu_layers = 200;
+
+    //---------------------------------
+    // Print help :
+    //---------------------------------
+
+    if ( argc < 2 || argv[1][0] == '-' )
+    {
+        printf( "Usage: %s MODEL_PATH [BEAM_WIDTH=2] [PROMPT]\n" , argv[0] );
+        return 1 ;
+    }
+
+    //---------------------------------
+    // Load parameters :
+    //---------------------------------
+
+    params.model = argv[1];
+
+    params.n_beams = 2 < argc ? std::stoi(argv[2]) : 2;
+
+    if ( argc > 3 )
+    {
+        params.prompt = argv[3];
+    }
+
+    if ( params.prompt.empty() )
+    {
+        params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n";
+    }
+
+    //---------------------------------
+    // Init LLM :
+    //---------------------------------
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
+
+    if ( model == NULL )
+    {
+        fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
+        return 1;
+    }
+
+    //---------------------------------
+    // Tokenize the prompt :
+    //---------------------------------
+
+    std::vector<llama_token> tokens_list = llama_tokenize(ctx, params.prompt, true);
+
+    const size_t max_context_size     = llama_n_ctx( ctx );
+    const size_t max_tokens_list_size = max_context_size - 4 ;
+
+    if (tokens_list.size() > max_tokens_list_size)
+    {
+        fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" ,
+             __func__ , tokens_list.size() , max_tokens_list_size );
+        return 1;
+    }
+
+    fprintf( stderr, "\n\n" );
+
+    // Print the tokens from the prompt :
+
+    for( auto id : tokens_list )
+    {
+        std::cout << llama_token_to_piece(ctx, id);
+    }
+    std::cout << std::flush;
+
+    int n_past = 0;
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens_list.data(), tokens_list.size(), n_past, 0)))
+    {
+        fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
+        return 1;
+    }
+    n_past += tokens_list.size();
+
+    beam_search_callback_data callback_data{ctx, {}};
+    size_t const beam_width = static_cast<size_t>(params.n_beams);
+    int const n_predict = 256;
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict);
+
+    std::cout << "\n\n";
+    for (llama_token const token_id : callback_data.response) {
+        std::cout << llama_token_to_piece(ctx,token_id);
+    }
+    std::cout << std::endl;
+
+    llama_free( ctx );
+    llama_free_model( model );
+
+    llama_backend_free();
+
+    return 0;
+}
diff --git a/examples/benchmark/CMakeLists.txt b/examples/benchmark/CMakeLists.txt
index 3f3415350..14916d831 100644
--- a/examples/benchmark/CMakeLists.txt
+++ b/examples/benchmark/CMakeLists.txt
@@ -1,7 +1,8 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)
diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp
index f7215f43b..f1c382aa9 100644
--- a/examples/benchmark/benchmark-matmult.cpp
+++ b/examples/benchmark/benchmark-matmult.cpp
@@ -1,5 +1,6 @@
-#include "ggml.h"
 #include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 
 #include <locale.h>
 #include <assert.h>
@@ -20,7 +21,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
     struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
@@ -31,19 +32,19 @@ void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph,
     ggml_graph_compute(graph, &plan);
 }
 
-float tensor_sum_elements(const ggml_tensor * tensor) {
-    float sum = 0;
-    if (tensor->type==GGML_TYPE_F32) {
+static float tensor_sum_elements(const ggml_tensor * tensor) {
+    double sum = 0;
+    if (tensor->type == GGML_TYPE_F32) {
         for (int j = 0; j < tensor->ne[1]; j++) {
             for (int k = 0; k < tensor->ne[0]; k++) {
-                sum +=  ((float *) tensor->data)[j*tensor->ne[0]+k];
+                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
             }
         }
     }
     return sum;
 }
 
-void tensor_dump(const ggml_tensor * tensor, const char * name) {
+static void tensor_dump(const ggml_tensor * tensor, const char * name) {
     printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
         tensor->type, ggml_type_name(tensor->type),
         tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
@@ -58,7 +59,7 @@ struct benchmark_params_struct {
     int32_t n_iterations  = 10;
 };
 
-void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
+static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
@@ -99,7 +100,7 @@ int main(int argc, char ** argv)  {
         exit(1);
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
     printf("Starting Test\n");
 
     // create the ggml context
@@ -125,12 +126,15 @@ int main(int argc, char ** argv)  {
 
     //printf("Memsize required = %i\n", sizex*sizex);
 
+    // TODO: perform the bench for all types or for a user specified type
+    const ggml_type qtype = GGML_TYPE_Q4_1;
+
     size_t ctx_size = 0;
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
     ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
-    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
@@ -163,7 +167,7 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
     ggml_set_f32(m2, 2.0f);
 
-    printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
     // printf("Creating new tensor m11xm2\n");
     struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
 
@@ -181,17 +185,16 @@ int main(int argc, char ** argv)  {
 
     TENSOR_DUMP(gf.nodes[0]);
 
-    printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
+    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
     int32_t nelements = sizex*sizey;
-    int32_t ne[2] = { sizex, sizey };
 
     std::vector<int64_t> hist_cur(1 << 4, 0);
 
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -202,8 +205,8 @@ int main(int argc, char ** argv)  {
 
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
-    ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
+    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
@@ -220,7 +223,7 @@ int main(int argc, char ** argv)  {
     printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
 
 
-    // Let's use the F32 result from above as a reference for the q4_0 multiplication
+    // Let's use the F32 result from above as a reference for the quantized multiplication
     float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
@@ -250,7 +253,7 @@ int main(int argc, char ** argv)  {
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
         float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
-        float delta = abs(sum_of_Q4_result - sum_of_F32_reference);
+        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 
         if (delta > allowed_delta)  {
diff --git a/examples/chat-persistent.sh b/examples/chat-persistent.sh
index e0c251e5b..22f5b83d3 100755
--- a/examples/chat-persistent.sh
+++ b/examples/chat-persistent.sh
@@ -9,7 +9,7 @@ if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
     exit 1
 fi
 
-MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
+MODEL="${MODEL:-./models/llama-13b/ggml-model-q4_0.gguf}"
 PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
 USER_NAME="${USER_NAME:-User}"
 AI_NAME="${AI_NAME:-ChatLLaMa}"
@@ -61,9 +61,9 @@ fi
 
 if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
     echo 'Prompt cache does not exist, building...'
-    # Default batch_size to 8 here for better user feedback during initial prompt processing
+    # Default batch_size to 64 here for better user feedback during initial prompt processing
     ./main 2>>"$LOG" \
-        --batch_size 8 \
+        --batch_size 64 \
         "${OPTS[@]}" \
         --prompt-cache "$PROMPT_CACHE_FILE" \
         --file "$CUR_PROMPT_FILE" \
@@ -132,7 +132,7 @@ while read -e line; do
     # HACK get num tokens from debug message
     # TODO get both messages in one go
     if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
-        ! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+        ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
         echo >&2 "Couldn't get number of tokens from ./main output!"
         exit 1
     fi
diff --git a/examples/chat.sh b/examples/chat.sh
index 9a928ef05..d567acecd 100755
--- a/examples/chat.sh
+++ b/examples/chat.sh
@@ -11,6 +11,6 @@ cd ..
 #
 #   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
 #
-./main -m ./models/7B/ggml-model-q4_0.bin -c 512 -b 1024 -n 256 --keep 48 \
+./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
     --repeat_penalty 1.0 --color -i \
     -r "User:" -f prompts/chat-with-bob.txt
diff --git a/examples/convert-llama2c-to-ggml/README.md b/examples/convert-llama2c-to-ggml/README.md
index 868f57d6d..0f37d295b 100644
--- a/examples/convert-llama2c-to-ggml/README.md
+++ b/examples/convert-llama2c-to-ggml/README.md
@@ -12,15 +12,15 @@ usage: ./convert-llama2c-to-ggml [options]
 
 options:
   -h, --help                       show this help message and exit
-  --copy-vocab-from-model FNAME    model path from which to copy vocab (default 'models/ggml-vocab.bin')
+  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default 'models/7B/ggml-model-f16.gguf')
   --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model
   --llama2c-output-model FNAME     model path to save the converted llama2.c model (default ak_llama_model.bin')
 ```
 
-An example command is as follows:
+An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows:
 
-`$ ./convert-llama2c-to-ggml --copy-vocab-from-model <ggml-vocab.bin> --llama2c-model <llama2.c model path> --llama2c-output-model <ggml output model path>`
+`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
-Now you can use the model with command like:
+Now you can use the model with a command like:
 
-`$ ./main -m <ggml output model path> -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5`
+`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`
diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
index 469d6e3de..c291f0adf 100644
--- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "llama.h"
+#include "common.h"
 
 #include <unordered_map>
 #include <vector>
@@ -10,13 +11,60 @@
 #include <ctime>
 #include <random>
 #include <stdexcept>
+#include <sstream>
 #include <algorithm>
 #include <string>
 
+// GGUF keys & tensor names.
+
+#define KV_GENERAL_ARCHITECTURE          "general.architecture"
+#define KV_GENERAL_NAME                  "general.name"
+
+#define KV_TOKENIZER_MODEL               "tokenizer.ggml.model"
+#define KV_TOKENIZER_LIST                "tokenizer.ggml.tokens"
+#define KV_TOKENIZER_TOKEN_TYPE          "tokenizer.ggml.token_type"
+#define KV_TOKENIZER_SCORES              "tokenizer.ggml.scores"
+#define KV_TOKENIZER_BOS_ID              "tokenizer.ggml.bos_token_id"
+#define KV_TOKENIZER_EOS_ID              "tokenizer.ggml.eos_token_id"
+#define KV_TOKENIZER_UNK_ID              "tokenizer.ggml.unknown_token_id"
+#define KV_TOKENIZER_SEP_ID              "tokenizer.ggml.seperator_token_id"
+#define KV_TOKENIZER_PAD_ID              "tokenizer.ggml.padding_token_id"
+#define KV_TOKENIZER_HF_JSON             "tokenizer.huggingface.json"
+
+#define KV_CONTEXT_LENGTH                "llama.context_length"
+#define KV_EMBEDDING_LENGTH              "llama.embedding_length"
+#define KV_BLOCK_COUNT                   "llama.block_count"
+#define KV_FEED_FORWARD_LENGTH           "llama.feed_forward_length"
+#define KV_ATTENTION_HEAD_COUNT          "llama.attention.head_count"
+#define KV_ATTENTION_HEAD_COUNT_KV       "llama.attention.head_count_kv"
+#define KV_ATTENTION_LAYERNORM_RMS_EPS   "llama.attention.layer_norm_rms_epsilon"
+#define KV_ROPE_DIMENSION_COUNT          "llama.rope.dimension_count"
+
+#define TN_TOKEN_EMBD  "token_embd.weight"
+#define TN_OUTPUT_NORM "output_norm.weight"
+#define TN_OUTPUT      "output.weight"
+#define TN_ATTN_NORM   "blk.%d.attn_norm.weight"
+#define TN_ATTN_Q      "blk.%d.attn_q.weight"
+#define TN_ATTN_K      "blk.%d.attn_k.weight"
+#define TN_ATTN_V      "blk.%d.attn_v.weight"
+#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
+#define TN_FFN_NORM    "blk.%d.ffn_norm.weight"
+#define TN_FFN_GATE    "blk.%d.ffn_gate.weight"
+#define TN_FFN_DOWN    "blk.%d.ffn_down.weight"
+#define TN_FFN_UP      "blk.%d.ffn_up.weight"
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_VERSION_GGJT_V3   3
+
+#define TOKENIZER_NAME "llama"
+#define UNKNOWN_TOKEN_ID 0
+#define BOS_TOKEN_ID 1
+#define EOS_TOKEN_ID 2
+
 //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc.
 typedef struct {
     int dim; // transformer dimension
@@ -28,7 +76,7 @@ typedef struct {
     int seq_len; // max sequence length
 } Config;
 
-typedef struct {
+struct TransformerWeights {
     // token embedding table
     float* token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
@@ -49,10 +97,25 @@ typedef struct {
     // float* freq_cis_real; // (seq_len, dim/2)
     // float* freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    //float* wcls;
-} TransformerWeights;
+    float* wcls;
 
-void malloc_weights(TransformerWeights* w, Config* p) {
+    ~TransformerWeights() {
+        delete[] token_embedding_table;
+        delete[] rms_att_weight;
+        delete[] rms_ffn_weight;
+        delete[] wq;
+        delete[] wk;
+        delete[] wv;
+        delete[] wo;
+        delete[] w1;
+        delete[] w2;
+        delete[] w3;
+        delete[] rms_final_weight;
+        delete[] wcls;
+    }
+};
+
+static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
     // we calloc instead of malloc to keep valgrind happy
     w->token_embedding_table = new float[p->vocab_size * p->dim]();
     printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
@@ -86,9 +149,16 @@ void malloc_weights(TransformerWeights* w, Config* p) {
 
     w->rms_final_weight = new float[p->dim]();
     printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+
+    if (shared_weights) {
+        w->wcls = NULL;
+    } else {
+        w->wcls = new float[p->vocab_size * p->dim]();
+        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+    }
 }
 
-int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
+static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
     if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
     if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
     if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
@@ -100,24 +170,26 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) {
     if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
     if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
     if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+
+    // Skip freq_cis_real & freq_cis_imag
+    int head_size = p->dim / p->n_heads;
+    fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
+
+    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+
+    // Check we didn't forget to read anything
+    auto curr = ftell(f);
+    fseek(f, 0, SEEK_END);
+    auto end = ftell(f);
+    if (curr != end) {
+        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        return 1;
+    }
+
     return 0;
 }
 
-void free_weights(TransformerWeights* w) {
-    delete w->token_embedding_table;
-    delete w->rms_att_weight;
-    delete w->rms_ffn_weight;
-    delete w->wq;
-    delete w->wk;
-    delete w->wv;
-    delete w->wo;
-    delete w->w1;
-    delete w->w2;
-    delete w->w3;
-    delete w->rms_final_weight;
-}
-
-void print_sample_weights(TransformerWeights *w){
+static void print_sample_weights(TransformerWeights *w){
     printf("----- Quick print of first of the weight vales of all the variables\n");
     printf("%f\n", w->token_embedding_table[0]);
     printf("%f\n", w->rms_att_weight[0]);
@@ -131,6 +203,7 @@ void print_sample_weights(TransformerWeights *w){
     printf("%f\n", w->w2[0]);
     printf("%f\n", w->w3[0]);
     printf("%f\n", w->rms_att_weight[0]);
+    if (w->wcls) printf("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -155,6 +228,7 @@ struct my_llama_hparams {
     uint32_t n_vocab = 32000;
     uint32_t n_ctx   = 512;   // this is provided as user input?
     uint32_t n_embd  = 4096;
+    uint32_t n_ff    = 11008;
     uint32_t n_mult  = 4;
     uint32_t n_head  = 32;
     uint32_t n_layer = 32;
@@ -186,6 +260,8 @@ struct my_llama_layer {
 struct my_llama_model {
     struct ggml_context * ctx = NULL;
 
+    std::string name;
+
     my_llama_hparams hparams;
 
     struct ggml_tensor * tok_embeddings;
@@ -248,30 +324,25 @@ struct train_params {
     int mem_compute1_gb;
 };
 
-uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
-    const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
-    return n_ff;
-}
-
-void print_params(struct my_llama_hparams * params) {
+static void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
     printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
     printf("%s: n_embd:  %d\n", __func__, params->n_embd);
     printf("%s: n_mult:  %d\n", __func__, params->n_mult);
     printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, get_n_ff(params));
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
     printf("%s: n_layer: %d\n", __func__, params->n_layer);
     printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
 
-void init_model(struct my_llama_model * model) {
+static void init_model(struct my_llama_model * model) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_embd  = hparams.n_embd;
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
-    const uint32_t n_ff = get_n_ff(&hparams);
+    const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
     model->train_its = 0;
@@ -337,17 +408,17 @@ void init_model(struct my_llama_model * model) {
     }
 }
 
-float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
     float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
     return *ptr;
 }
 
-int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
+static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
     int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
     return *ptr;
 }
 
-void print_row(struct ggml_tensor * probs, int i) {
+static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
         printf(" %f", p);
@@ -355,7 +426,7 @@ void print_row(struct ggml_tensor * probs, int i) {
     printf("\n");
 }
 
-void print_matrix(struct ggml_tensor * probs) {
+static void print_matrix(struct ggml_tensor * probs) {
     assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
@@ -429,10 +500,10 @@ struct llama_file {
         errno = 0;
         std::size_t ret = std::fread(ptr, size, 1, fp);
         if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
+            die_fmt("fread failed: %s", strerror(errno));
         }
         if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+            die("unexpectedly reached end of file");
         }
     }
 
@@ -453,21 +524,6 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
-        if (ret != 1) {
-            throw std::runtime_error(format("write error: %s", strerror(errno)));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
     ~llama_file() {
         if (fp) {
             std::fclose(fp);
@@ -475,31 +531,7 @@ struct llama_file {
     }
 };
 
-void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
-    if (tensor == NULL) {
-        file->write_u32(0);
-        file->write_u32(0);
-        file->write_u32(GGML_TYPE_F32);
-        file->seek((0-file->tell()) & 31, SEEK_CUR);
-        return;
-    }
-    const char * name = ggml_get_name(tensor);
-    uint32_t name_len = strlen(name);
-    uint32_t nd = tensor->n_dims;
-    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
-                       (uint32_t)tensor->ne[1],
-                       (uint32_t)tensor->ne[2],
-                       (uint32_t)tensor->ne[3] };
-    file->write_u32(nd);
-    file->write_u32(name_len);
-    file->write_u32(tensor->type);
-    file->write_raw(ne, sizeof(ne[0]) * nd);
-    file->write_raw(name, name_len);
-    file->seek((0-file->tell()) & 31, SEEK_CUR);
-    file->write_raw(tensor->data, ggml_nbytes(tensor));
-}
-
-bool is_ggml_file(const char *filename) {
+static bool is_ggml_file(const char * filename) {
     llama_file file(filename, "rb");
     if (file.size < 4) {
         return false;
@@ -508,45 +540,104 @@ bool is_ggml_file(const char *filename) {
     return magic == GGUF_MAGIC;
 }
 
-void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
-    // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary
+static std::string llama_escape_whitespaces(const std::string & text) {
+    std::ostringstream out;
+    for (char c : text) {
+        if (c == ' ') out << "\xe2\x96\x81";
+        else out << c;
+    }
+    return out.str();
+}
+
+static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
     if (is_ggml_file(filename)) {
+        struct ggml_context * ctx_data = NULL;
 
-        struct llama_context_params llama_params = llama_context_default_params();
-        llama_params.vocab_only = true;
+        struct gguf_init_params params = {
+            /*.no_alloc = */ false,
+            /*.ctx      = */ &ctx_data,
+        };
 
-        struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params);
-        struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+        struct gguf_context * ctx = gguf_init_from_file(filename, params);
+        GGML_ASSERT(ctx != NULL);
+
+        const int model_idx = gguf_find_key(ctx, KV_TOKENIZER_MODEL);
+        GGML_ASSERT(model_idx >= 0);
+        std::string tokenizer_name = gguf_get_val_str(ctx, model_idx);
+        GGML_ASSERT(tokenizer_name == TOKENIZER_NAME);
+
+        const int token_idx = gguf_find_key(ctx, KV_TOKENIZER_LIST);
+        GGML_ASSERT(token_idx >= 0);
+
+        const int score_idx = gguf_find_key(ctx, KV_TOKENIZER_SCORES);
+        GGML_ASSERT(score_idx >= 0);
+        const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+
+        const int toktype_idx = gguf_find_key(ctx, KV_TOKENIZER_TOKEN_TYPE);
+        GGML_ASSERT(toktype_idx >= 0);
+        const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
+
+        const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
 
-        const int n_vocab = llama_n_vocab(lctx);
         vocab->id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
-            vocab->id_to_token[i].text  = llama_token_get_text(lctx, i);
-            vocab->id_to_token[i].score = llama_token_get_score(lctx, i);
-            vocab->id_to_token[i].type  = llama_token_get_type(lctx, i);
-            vocab->token_to_id.emplace(vocab->id_to_token[i].text, i);
+
+        for (uint32_t i = 0; i < n_vocab; i++) {
+            std::string word = gguf_get_arr_str(ctx, token_idx, i);
+
+            vocab->token_to_id[word] = i;
+
+            auto & token_data = vocab->id_to_token[i];
+            token_data.text  = std::move(word);
+            token_data.score = scores[i];
+            token_data.type  = (llama_token_type) toktypes[i];
         }
-        llama_free(lctx);
-        llama_free_model(lmodel);
-    } else { // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename);
+        ggml_free(ctx_data);
+        gguf_free(ctx);
+    } else {
+        // assume llama2.c vocabulary
+        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
         llama_file file(filename, "rb");
+        if (!file.fp) {
+            die_fmt("%s: %s", strerror(errno), filename);
+        }
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
+        for (llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
-            vocab->id_to_token[i].text = text;
-            vocab->id_to_token[i].score = score;
-            vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED;
-            vocab->token_to_id.emplace(text, i);
+
+            unsigned char byte_val;
+            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            if (id == UNKNOWN_TOKEN_ID) {
+                text = "<unk>";
+                type = LLAMA_TOKEN_TYPE_UNKNOWN;
+            } else if (id == BOS_TOKEN_ID) {
+                text = "<s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (id == EOS_TOKEN_ID) {
+                text = "</s>";
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (text.empty()) {
+                type = LLAMA_TOKEN_TYPE_CONTROL;
+            } else if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) {
+                // Text of byte tokens is already in the expected format.
+                type = LLAMA_TOKEN_TYPE_BYTE;
+            } else {
+                type = LLAMA_TOKEN_TYPE_NORMAL;
+            }
+            text = llama_escape_whitespaces(text);
+
+            vocab->id_to_token[id].text = text;
+            vocab->id_to_token[id].score = score;
+            vocab->id_to_token[id].type = type;
+            vocab->token_to_id.emplace(text, id);
         }
     }
 }
 
-void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * karpathy_weights){
+static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
     switch (gg_weights->n_dims){
         case 1:
@@ -582,90 +673,123 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar
     }
 }
 
-void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
-    struct llama_file file(filename, "wb");
-    if (file.fp == NULL) {
-        return;
+static void save_as_llama_model(
+    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+) {
+    // convert AK weights into GG weights one by one.
+    // w->token_embedding_table -> model->tok_embeddings
+    // float*                   -> struct ggml_tensor
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
+    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    //print_row(model->norm, 0);
+
+    // for rms-att-weight
+    int row_length = model->hparams.n_embd;
+    int n_ff = model->hparams.n_ff;
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
+        auto & layer = model->layers[i];
+        // 1d
+        convert_weights_ak_to_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
+        convert_weights_ak_to_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
+
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
+        convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
+        convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+
+        convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
+        convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
+        convert_weights_ak_to_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
     }
 
-#pragma message("TODO: implement file saving using gguf")
-    (void) vocab;
-    (void) model;
-    (void) w;
-//    // write_magic
-//    file.write_u32(LLAMA_FILE_MAGIC);   // magic
-//    file.write_u32(LLAMA_FILE_VERSION); // version
-//    // write_hparams
-//    file.write_u32(model->hparams.n_vocab);
-//    file.write_u32(model->hparams.n_embd);
-//    file.write_u32(model->hparams.n_mult);
-//    file.write_u32(model->hparams.n_head);
-//    file.write_u32(model->hparams.n_layer);
-//    file.write_u32(model->hparams.n_rot);
-//    file.write_u32(LLAMA_FTYPE_ALL_F32);
-//
-//    // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk.
-//    uint32_t n_vocab = model->hparams.n_vocab;
-//    for (uint32_t i = 0; i < n_vocab; i++) {
-//        const auto & token_data = vocab->id_to_token.at(i);
-//        file.write_u32((uint32_t) token_data.tok.size());
-//        file.write_raw(token_data.tok.data(), token_data.tok.size());
-//        file.write_raw(&token_data.score, sizeof(token_data.score));
-//    }
-//
-//    // stuff AK weights into GG weights one by one.
-//    // w->token_embedding_table -> model->tok_embeddings
-//    // float*                   -> struct ggml_tensor
-//    stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
-//    stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table);
-//
-//    stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
-//    //print_row(model->norm, 0);
-//
-//    // for rms-att-weight
-//    int row_length = model->hparams.n_embd;
-//    const auto & hparams = model->hparams;
-//    //int n_ff = model->hparams.n_embd;
-//    int n_ff = get_n_ff(&hparams);
-//
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
-//        auto & layer = model->layers[i];
-//        // 1d
-//        stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.ffn_norm      , &w->rms_ffn_weight[i*row_length]);
-//
-//        // from 3d matrix layer x dim x dim to 2d matrix dim x dim
-//        stuff_karpathy_weights_into_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wv            , &w->wv[i*row_length*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.wo            , &w->wo[i*row_length*row_length]);
-//
-//        stuff_karpathy_weights_into_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
-//        stuff_karpathy_weights_into_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
-//        stuff_karpathy_weights_into_gg(layer.w3            , &w->w3[i*row_length*n_ff]);
-//    }
-//    // write tensors
-//    write_tensor(&file, model->tok_embeddings);
-//    write_tensor(&file, model->norm);
-//    write_tensor(&file, model->output); // ?
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
-//        auto & layer = model->layers[i];
-//
-//        write_tensor(&file, layer.attention_norm);
-//        write_tensor(&file, layer.wq);
-//        write_tensor(&file, layer.wk);
-//        write_tensor(&file, layer.wv);
-//        write_tensor(&file, layer.wo);
-//        write_tensor(&file, layer.ffn_norm);
-//        write_tensor(&file, layer.w1);
-//        write_tensor(&file, layer.w2);
-//        write_tensor(&file, layer.w3);
-//    }
+    struct gguf_context * ctx = gguf_init_empty();
+
+    std::vector<const char*> tokens;
+    std::vector<float> scores;
+    std::vector<llama_token_type> token_types;
+    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+        tokens.push_back(token_data.text.c_str());
+        scores.push_back(token_data.score);
+        token_types.push_back(token_data.type);
+    }
+    gguf_set_arr_str(ctx, KV_TOKENIZER_LIST, tokens.data(), tokens.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_SCORES, GGUF_TYPE_FLOAT32, scores.data(), scores.size());
+    gguf_set_arr_data(ctx, KV_TOKENIZER_TOKEN_TYPE, GGUF_TYPE_INT32, token_types.data(), token_types.size());
+
+    gguf_set_val_str(ctx, KV_TOKENIZER_MODEL, TOKENIZER_NAME);
+
+    gguf_set_val_str(ctx, KV_GENERAL_ARCHITECTURE, "llama");
+    gguf_set_val_str(ctx, KV_GENERAL_NAME, "llama");
+
+    // special tokens
+    gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
+    gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
+
+    gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
+    gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
+    gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    // n_head_kv is optional, default to n_head
+    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
+    gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
+    gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
+
+    // write tensors
+    ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD);
+    gguf_add_tensor(ctx, model->tok_embeddings);
+
+    ggml_set_name(model->norm, TN_OUTPUT_NORM);
+    gguf_add_tensor(ctx, model->norm);
+
+    ggml_set_name(model->output, TN_OUTPUT);
+    gguf_add_tensor(ctx, model->output);
+
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        ggml_format_name(layer.wq, TN_ATTN_Q, i);
+        gguf_add_tensor(ctx, layer.wq);
+
+        ggml_format_name(layer.wk, TN_ATTN_K, i);
+        gguf_add_tensor(ctx, layer.wk);
+
+        ggml_format_name(layer.wv, TN_ATTN_V, i);
+        gguf_add_tensor(ctx, layer.wv);
+
+        ggml_format_name(layer.wo, TN_ATTN_OUTPUT, i);
+        gguf_add_tensor(ctx, layer.wo);
+
+        ggml_format_name(layer.attention_norm, TN_ATTN_NORM, i);
+        gguf_add_tensor(ctx, layer.attention_norm);
+
+        ggml_format_name(layer.w1, TN_FFN_GATE, i);
+        gguf_add_tensor(ctx, layer.w1);
+
+        ggml_format_name(layer.w2, TN_FFN_DOWN, i);
+        gguf_add_tensor(ctx, layer.w2);
+
+        ggml_format_name(layer.w3, TN_FFN_UP, i);
+        gguf_add_tensor(ctx, layer.w3);
+
+        ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i);
+        gguf_add_tensor(ctx, layer.ffn_norm);
+    }
+
+    gguf_write_to_file(ctx, filename, false);
+    gguf_free(ctx);
 }
 
-struct train_params get_default_train_params() {
+static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/ggml-vocab.bin";
+    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
     params.fn_train_data     = "shakespeare.txt";
     params.fn_checkpoint_in  = "checkpoint.bin";
@@ -713,18 +837,18 @@ struct train_params get_default_train_params() {
     return params;
 }
 
-void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+static void print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                       show this help message and exit\n");
-    fprintf(stderr, "  --copy-vocab-from-model FNAME    llama2.c vocabulary or ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model);
+    fprintf(stderr, "  --copy-vocab-from-model FNAME    path of gguf llama model or llama2.c vocabulary from which to copy vocab (default '%s')\n", params->fn_vocab_model);
     fprintf(stderr, "  --llama2c-model FNAME            [REQUIRED] model path from which to load Karpathy's llama2.c model\n");
     fprintf(stderr, "  --llama2c-output-model FNAME     model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model);
     fprintf(stderr, "\n");
 }
 
-bool params_parse(int argc, char ** argv, struct train_params * params) {
+static bool params_parse(int argc, char ** argv, struct train_params * params) {
     bool invalid_param = false;
     bool reqd_param_found = false;
     std::string arg;
@@ -779,21 +903,32 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
     return true;
 }
 
+static std::string basename(const std::string &path) {
+    size_t pos = path.find_last_of("/\\");
+    if (pos == std::string::npos) {
+        return path;
+    }
+    return path.substr(pos + 1);
+}
+
 int main(int argc, char ** argv) {
     struct train_params params = get_default_train_params();
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
     Config config;
-    TransformerWeights weights;
+    TransformerWeights weights = {};
     {
         FILE *file = fopen(params.fn_llama2c_model, "rb");
         if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
         // read in the config header
         if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        auto shared_weights = config.vocab_size > 0;
+        config.vocab_size = abs(config.vocab_size);
+
         // read in the Transformer weights
-        malloc_weights(&weights, &config);
-        if(checkpoint_init_weights(&weights, &config, file)) { return 1; }
+        malloc_weights(&weights, &config, shared_weights);
+        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
         fclose(file);
     }
 
@@ -804,6 +939,7 @@ int main(int argc, char ** argv) {
     model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
     model.hparams.n_ctx   = params.n_ctx;
     model.hparams.n_embd  = config.dim; //params.n_embd;
+    model.hparams.n_ff    = config.hidden_dim;
     model.hparams.n_mult  = 32;//params.n_mult;
     model.hparams.n_head  = config.n_heads; //params.n_head;
     model.hparams.n_layer = config.n_layers; //params.n_layer;
@@ -817,11 +953,11 @@ int main(int argc, char ** argv) {
     model.ctx = ggml_init(lcparams);
 
     init_model(&model);
+    model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
     printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
-    free_weights(&weights);
     return 0;
 }
diff --git a/examples/embd-input/embd-input-lib.cpp b/examples/embd-input/embd-input-lib.cpp
index 8a6ad882e..99e6bdad5 100644
--- a/examples/embd-input/embd-input-lib.cpp
+++ b/examples/embd-input/embd-input-lib.cpp
@@ -1,8 +1,5 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
+#include "build-info.h"
+#include "common.h"
 #include "embd-input.h"
 
 #include <cassert>
@@ -23,11 +20,11 @@ extern "C" {
 struct MyModel* create_mymodel(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return nullptr;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = uint32_t(time(NULL));
@@ -51,8 +48,7 @@ struct MyModel* create_mymodel(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
     struct MyModel * ret = new MyModel();
     ret->ctx = ctx;
@@ -74,7 +70,7 @@ bool eval_float(void * model, float * input, int N){
     MyModel * mymodel = (MyModel*)model;
     llama_context * ctx = mymodel->ctx;
     gpt_params params = mymodel->params;
-    int n_emb = llama_n_embd(ctx);
+    int n_emb = llama_n_embd(llama_get_model(ctx));
     int n_past = mymodel->n_past;
     int n_batch = N; // params.n_batch;
 
@@ -83,7 +79,8 @@ bool eval_float(void * model, float * input, int N){
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past, params.n_threads)) {
+        llama_batch batch = {  int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
+        if (llama_decode(ctx, batch)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -104,7 +101,7 @@ bool eval_tokens(void * model, std::vector<llama_token> tokens) {
         if (n_eval > params.n_batch) {
             n_eval = params.n_batch;
         }
-        if (llama_eval(ctx, &tokens[i], n_eval, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -135,7 +132,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
 
     // out of user input, sample next token
     const float   temp            = params.temp;
-    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
+    const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : params.top_k;
     const float   top_p           = params.top_p;
     const float   tfs_z           = params.tfs_z;
     const float   typical_p       = params.typical_p;
@@ -151,7 +148,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
     llama_token id = 0;
     {
         auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
+        auto n_vocab = llama_n_vocab(llama_get_model(ctx));
 
         // Apply params.logit_bias map
         for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
@@ -186,11 +183,11 @@ llama_token sampling_id(struct MyModel* mymodel) {
             if (mirostat == 1) {
                 static float mirostat_mu = 2.0f * mirostat_tau;
                 const int mirostat_m = 100;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
             } else if (mirostat == 2) {
                 static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
             } else {
                 // Temperature sampling
@@ -198,7 +195,7 @@ llama_token sampling_id(struct MyModel* mymodel) {
                 llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
                 llama_sample_typical(ctx, &candidates_p, typical_p, 1);
                 llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                llama_sample_temperature(ctx, &candidates_p, temp);
+                llama_sample_temp(ctx, &candidates_p, temp);
                 id = llama_sample_token(ctx, &candidates_p);
             }
         }
@@ -214,7 +211,7 @@ const char * sampling(struct MyModel * mymodel) {
     if (id == llama_token_eos(ctx)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_str(ctx, id);
+        ret = llama_token_to_piece(ctx, id);
     }
     eval_id(mymodel, id);
     return ret.c_str();
diff --git a/examples/embd-input/embd-input-test.cpp b/examples/embd-input/embd-input-test.cpp
index e5e040f62..dc4a0e488 100644
--- a/examples/embd-input/embd-input-test.cpp
+++ b/examples/embd-input/embd-input-test.cpp
@@ -8,7 +8,7 @@ int main(int argc, char** argv) {
     auto mymodel = create_mymodel(argc, argv);
     int N = 10;
     int max_tgt_len = 500;
-    int n_embd = llama_n_embd(mymodel->ctx);
+    int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
 
     // add random float embd to test evaluation
     float * data = new float[N*n_embd];
diff --git a/examples/embd-input/embd-input.h b/examples/embd-input/embd-input.h
index efb5ba5e2..eff5e3b84 100644
--- a/examples/embd-input/embd-input.h
+++ b/examples/embd-input/embd-input.h
@@ -3,7 +3,6 @@
 
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 extern "C" {
 
diff --git a/examples/embedding/README.md b/examples/embedding/README.md
index fe8f5dcc6..6929454c5 100644
--- a/examples/embedding/README.md
+++ b/examples/embedding/README.md
@@ -1,3 +1,21 @@
-# embedding
+# llama.cpp/example/embedding
 
-TODO
+This example demonstrates generate high-dimensional embedding vector of a given text with llama.cpp.
+
+## Quick Start
+
+To get started right away, run the following command, making sure to use the correct path for the model you have:
+
+### Unix-based systems (Linux, macOS, etc.):
+
+```bash
+./embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
+```
+
+### Windows:
+
+```powershell
+embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
+```
+
+The above command will output space-separated float values.
diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp
index 38395c75b..14075609e 100644
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <ctime>
 
@@ -11,18 +11,13 @@
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
     params.embedding = true;
 
-    if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
-    }
-
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -47,18 +42,22 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
     int n_past = 0;
 
-    // Add a space in front of the first character to match OG llama tokenizer behavior
-    params.prompt.insert(0, 1, ' ');
-
     // tokenize the prompt
     auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
 
@@ -67,20 +66,20 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
         fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
+            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
         fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)params.n_ctx) {
+    if (embd_inp.size() > (size_t)n_ctx) {
         fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), params.n_ctx);
+                __func__, embd_inp.size(), n_ctx);
         return 1;
     }
 
     while (!embd_inp.empty()) {
         int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }
@@ -88,8 +87,8 @@ int main(int argc, char ** argv) {
         embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
     }
 
-    const int n_embd = llama_n_embd(ctx);
-    const auto embeddings = llama_get_embeddings(ctx);
+    const int n_embd = llama_n_embd(model);
+    const auto * embeddings = llama_get_embeddings(ctx);
 
     for (int i = 0; i < n_embd; i++) {
         printf("%f ", embeddings[i]);
diff --git a/examples/export-lora/CMakeLists.txt b/examples/export-lora/CMakeLists.txt
new file mode 100644
index 000000000..cbbdaec67
--- /dev/null
+++ b/examples/export-lora/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET export-lora)
+add_executable(${TARGET} export-lora.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/export-lora/README.md b/examples/export-lora/README.md
new file mode 100644
index 000000000..0cf3e8e45
--- /dev/null
+++ b/examples/export-lora/README.md
@@ -0,0 +1,26 @@
+# export-lora
+
+Apply LORA adapters to base model and export the resulting model.
+
+```
+usage: export-lora [options]
+
+options:
+  -h, --help                         show this help message and exit
+  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
+  -o FNAME, --model-out FNAME        path to save exported model (default '')
+  -l FNAME, --lora FNAME             apply LoRA adapter
+  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
+  -t N, --threads N                  number of threads to use during computation (default: 4)
+```
+
+For example:
+
+```bash
+./bin/export-lora \
+    -m open-llama-3b-v2-q8_0.gguf \
+    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
+    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
+```
+
+Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.
diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp
new file mode 100644
index 000000000..d803cfd5c
--- /dev/null
+++ b/examples/export-lora/export-lora.cpp
@@ -0,0 +1,474 @@
+
+#include "common.h"
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#include <vector>
+#include <string>
+#include <thread>
+
+static const size_t tensor_alignment = 32;
+
+struct lora_info {
+    std::string filename;
+    float scale;
+};
+
+struct export_lora_params {
+    std::string fn_model_base;
+    std::string fn_model_out;
+    std::vector<struct lora_info> lora;
+    int n_threads;
+};
+
+struct lora_data {
+    struct lora_info     info;
+    std::vector<uint8_t> data;
+    struct ggml_context * ctx;
+
+    uint32_t lora_r;
+    uint32_t lora_alpha;
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    bool eof() {
+        return tell() >= size;
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static struct export_lora_params get_default_export_lora_params() {
+    struct export_lora_params result;
+    result.fn_model_base = "";
+    result.fn_model_out  = "";
+    result.n_threads = GGML_DEFAULT_N_THREADS;
+    return result;
+}
+
+static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
+    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
+    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
+    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
+    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
+}
+
+static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct export_lora_params default_params = get_default_export_lora_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-m" || arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "-o" || arg == "--model-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_out = argv[i];
+        } else if (arg == "-l" || arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            lora.scale = 1.0f;
+            params->lora.push_back(lora);
+        } else if (arg == "-s" || arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            lora.scale = std::stof(argv[i]);
+            params->lora.push_back(lora);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_threads = std::stoi(argv[i]);
+            if (params->n_threads <= 0) {
+                params->n_threads = std::thread::hardware_concurrency();
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
+            export_lora_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+
+    if (params->fn_model_base == default_params.fn_model_base) {
+        fprintf(stderr, "error: please specify a filename for model-base.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (params->fn_model_out == default_params.fn_model_out) {
+        fprintf(stderr, "error: please specify a filename for model-out.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    return true;
+}
+
+static void free_lora(struct lora_data * lora) {
+    if (lora->ctx != NULL) {
+        ggml_free(lora->ctx);
+    }
+    delete lora;
+}
+
+static struct lora_data * load_lora(struct lora_info * info) {
+    struct lora_data * result = new struct lora_data;
+    result->info = *info;
+    result->ctx = NULL;
+    result->lora_r     = 1;
+    result->lora_alpha = 1;
+
+    struct llama_file file(info->filename.c_str(), "rb");
+    if (file.fp == NULL) {
+        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
+            info->filename.c_str());
+        free_lora(result);
+        return NULL;
+    }
+
+    struct ggml_init_params params_ggml;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_buffer = NULL;
+    params_ggml.no_alloc   = true;
+    result->ctx = ggml_init(params_ggml);
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    uint32_t magic   = file.read_u32();
+    if (magic != LLAMA_FILE_MAGIC_LORA) {
+        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
+    }
+    uint32_t version = file.read_u32();
+    if (version != 1) {
+        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
+    }
+    result->lora_r     = file.read_u32();
+    result->lora_alpha = file.read_u32();
+    // read tensor infos from file
+    std::vector<char> name_buf;
+    std::vector<struct ggml_tensor *> tensors;
+    std::vector<size_t> tensors_offset;
+    size_t total_nbytes_pad = 0;
+    while(!file.eof()) {
+        int64_t ne[4]   = {1,1,1,1};
+        uint32_t n_dims  = file.read_u32();
+        uint32_t namelen = file.read_u32();
+        uint32_t type    = file.read_u32();
+        for (uint32_t k = 0; k < n_dims; ++k) {
+            ne[k] = (int64_t)file.read_u32();
+        }
+        name_buf.clear();
+        name_buf.resize(namelen + 1, '\0');
+        file.read_raw(name_buf.data(), namelen);
+        file.seek((0-file.tell()) & 31, SEEK_CUR);
+        size_t offset = file.tell();
+        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
+        ggml_set_name(tensor, name_buf.data());
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        total_nbytes_pad += nbytes_pad;
+        tensors.push_back(tensor);
+        tensors_offset.push_back(offset);
+        file.seek(nbytes, SEEK_CUR);
+    }
+    // read tensor data
+    result->data.resize(total_nbytes_pad);
+    size_t data_offset = 0;
+    for (size_t i = 0; i < tensors.size(); ++i) {
+        struct ggml_tensor * tensor = tensors[i];
+        size_t offset     = tensors_offset[i];
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        file.seek(offset, SEEK_SET);
+        tensor->data = result->data.data() + data_offset;
+        file.read_raw(tensor->data, nbytes);
+        data_offset += nbytes_pad;
+    }
+    return result;
+}
+
+
+static struct ggml_cgraph * build_graph_lora(
+    struct ggml_context * ctx,
+    struct ggml_tensor * tensor,
+    struct ggml_tensor * lora_a,
+    struct ggml_tensor * lora_b,
+    float scaling
+) {
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
+    if (scaling != 1.0f) {
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+    }
+    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
+
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand (gf, res);
+    return gf;
+}
+
+static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
+    if (lora->ctx == NULL) {
+        return false;
+    }
+    std::string name = ggml_get_name(tensor);
+    std::string name_a = name + std::string(".loraA");
+    std::string name_b = name + std::string(".loraB");
+    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
+    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
+    if (lora_a == NULL || lora_b == NULL) {
+        return false;
+    }
+
+    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
+
+    struct ggml_init_params params;
+    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
+    struct ggml_context * ctx = NULL;
+    struct ggml_allocr * alloc = NULL;
+    struct ggml_cgraph * gf = NULL;
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+    ggml_free(ctx);
+
+    static std::vector<uint8_t> data_compute;
+    data_compute.resize(alloc_size + tensor_alignment);
+
+    ctx   = ggml_init(params);
+    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+    ggml_allocr_alloc_graph(alloc, gf);
+    ggml_allocr_free(alloc);
+
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> data_work;
+    data_work.resize(cplan.work_size);
+    cplan.work_data = data_work.data();
+
+    ggml_graph_compute(gf, &cplan);
+
+    ggml_free(ctx);
+    return true;
+}
+
+static void export_lora(struct export_lora_params * params) {
+    // load all loras
+    std::vector<struct lora_data *> loras;
+    for (size_t i = 0; i < params->lora.size(); ++i) {
+        struct lora_data * lora = load_lora(&params->lora[i]);
+        if (lora != NULL) {
+            loras.push_back(lora);
+        }
+    }
+    if (loras.size() == 0) {
+        fprintf(stderr, "warning: no lora adapters will be applied.\n");
+    }
+
+    // open input file
+    struct llama_file fin(params->fn_model_base.c_str(), "rb");
+    if (!fin.fp) {
+        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
+    }
+
+    // open base model gguf, read tensors without their data
+    struct ggml_context * ctx_in;
+    struct gguf_init_params params_gguf;
+    params_gguf.no_alloc = true;
+    params_gguf.ctx      = &ctx_in;
+    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
+
+    // create new gguf
+    struct gguf_context * gguf_out = gguf_init_empty();
+
+    // copy meta data from base model: kv and tensors
+    gguf_set_kv(gguf_out, gguf_in);
+    int n_tensors = gguf_get_n_tensors(gguf_in);
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+        gguf_add_tensor(gguf_out, tensor);
+    }
+
+    // create output file
+    struct llama_file fout(params->fn_model_out.c_str(), "wb");
+    if (!fout.fp) {
+        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    }
+
+    // write gguf meta data
+    std::vector<uint8_t> meta;
+    meta.resize(gguf_get_meta_size(gguf_out));
+    gguf_get_meta_data(gguf_out, meta.data());
+    fout.write_raw(meta.data(), meta.size());
+
+    std::vector<uint8_t> data;
+    std::vector<uint8_t> padding;
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+
+        // read tensor data
+        data.resize(ggml_nbytes(tensor));
+        tensor->data = data.data();
+        size_t offset = gguf_get_tensor_offset(gguf_in, i);
+        fin.seek(offset + meta.size(), SEEK_SET);
+        fin.read_raw(data.data(), data.size());
+
+        // apply all loras
+        for (size_t k = 0; k < loras.size(); ++k) {
+            apply_lora(tensor, loras[k], params->n_threads);
+        }
+
+        // write tensor data + padding
+        padding.clear();
+        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
+
+        GGML_ASSERT(fout.tell() == offset + meta.size());
+        // fout.seek(offset + meta.size(), SEEK_SET);
+        fout.write_raw(data.data(), data.size());
+        fout.write_raw(padding.data(), padding.size());
+
+        if (i % 2 == 0) {
+            printf(".");
+        }
+    }
+    printf("\n");
+
+    // close gguf
+    gguf_free(gguf_out);
+    gguf_free(gguf_in);
+
+    // free loras
+    for (size_t i = 0; i < loras.size(); ++i) {
+        free_lora(loras[i]);
+    }
+}
+
+int main(int argc, char ** argv) {
+    struct export_lora_params params = get_default_export_lora_params();
+
+    if (!export_lora_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    export_lora(&params);
+
+    return 0;
+}
diff --git a/examples/finetune/CMakeLists.txt b/examples/finetune/CMakeLists.txt
new file mode 100644
index 000000000..2b52d21cf
--- /dev/null
+++ b/examples/finetune/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/finetune/README.md b/examples/finetune/README.md
new file mode 100644
index 000000000..36e62578c
--- /dev/null
+++ b/examples/finetune/README.md
@@ -0,0 +1,90 @@
+# finetune
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# finetune LORA adapter
+./bin/finetune \
+        --model-base open-llama-3b-v2-q8_0.gguf \
+        --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
+        --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \
+        --train-data "shakespeare.txt" \
+        --save-every 10 \
+        --threads 6 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing
+
+# predict
+./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+```
+
+Finetune output files will be saved every N iterations (config with `--save-every N`).
+The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
+So in above example after 10 iterations these files will be written:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+After 10 more iterations:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
+
+llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
+These LORA adapters can then be used by `main` together with the base model, like in the 'predict' example command above.
+
+In `main` you can also load multiple LORA adapters, which will then be mixed together.
+
+For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
+  --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
+```
+
+You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`.
+
+For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
+
+```bash
+./bin/main -m open-llama-3b-v2-q8_0.gguf \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
+  --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
+```
+
+The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
+
+Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
+If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.
+
+The default LORA rank can be specified with `--lora-r N`.
+The LORA rank can be configured for each model tensor type separately with these command line options:
+
+```bash
+  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4)
+  --rank-att-norm N          LORA rank for attention norm tensor (default 1)
+  --rank-ffn-norm N          LORA rank for feed-forward norm tensor (default 1)
+  --rank-out-norm N          LORA rank for output norm tensor (default 1)
+  --rank-tok-embd N          LORA rank for token embeddings tensor (default 4)
+  --rank-out N               LORA rank for output tensor (default 4)
+  --rank-wq N                LORA rank for wq tensor (default 4)
+  --rank-wk N                LORA rank for wk tensor (default 4)
+  --rank-wv N                LORA rank for wv tensor (default 4)
+  --rank-wo N                LORA rank for wo tensor (default 4)
+  --rank-w1 N                LORA rank for w1 tensor (default 4)
+  --rank-w2 N                LORA rank for w2 tensor (default 4)
+  --rank-w3 N                LORA rank for w3 tensor (default 4)
+```
+
+The LORA rank of 'norm' tensors should always be 1.
+
+To see all available options use `finetune --help`.
diff --git a/examples/finetune/convert-finetune-checkpoint-to-gguf.py b/examples/finetune/convert-finetune-checkpoint-to-gguf.py
new file mode 100644
index 000000000..c8e14da87
--- /dev/null
+++ b/examples/finetune/convert-finetune-checkpoint-to-gguf.py
@@ -0,0 +1,489 @@
+#!/usr/bin/env python3
+# finetune checkpoint --> gguf conversion
+
+import argparse
+import gguf
+import os
+import struct
+import sys
+import numpy as np
+from pathlib import Path
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output"
+LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm"
+LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q"
+LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k"
+LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v"
+LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output"
+LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm"
+LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate"
+LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down"
+LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        if tuple(ne) != tuple(self.ne):
+            raise ValueError(f"Tensor.load: Expected number of elements {str(self.ne)} does not match what is read from file {str(ne)}")
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version != 1:
+            raise ValueError('Invalid version of optimization context in checkpoint file')
+
+        self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+        self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+        self.adam_m  = Tensor('f', [self.nx])
+        self.adam_v  = Tensor('f', [self.nx])
+        self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+        self.lbfgs_x    = Tensor('f', [self.nx])
+        self.lbfgs_xp   = Tensor('f', [self.nx])
+        self.lbfgs_g    = Tensor('f', [self.nx])
+        self.lbfgs_gp   = Tensor('f', [self.nx])
+        self.lbfgs_d    = Tensor('f', [self.nx])
+        self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+        self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+        self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+        # forgot to save type in version 1:
+        # guess self.type from number of remaining bytes
+        size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                [self.adam_m, self.adam_v]
+                                +([self.adam_pf] if (self.past > 0) else [])])
+        size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                 self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                 self.lbfgs_lmal, self.lbfgs_lmys,
+                                 self.lbfgs_lms, self.lbfgs_lmy]
+                                 +([self.lbfgs_pf] if (self.past > 0) else [])])
+        # due to alignment padding the size might not by exact
+        # but the difference in size for both types is significant,
+        # so we can just use whichever is closest
+        remaining = len(data) - offset
+        if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+            self.type = 0
+        else:
+            self.type = 1
+
+        if self.type == 0:
+            offset = self.adam_m.load(data, offset)
+            offset = self.adam_v.load(data, offset)
+            offset = self.adam_pf.load(data,offset)
+
+            self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        elif self.type == 1:
+            offset = self.lbfgs_x.load(data, offset)
+            offset = self.lbfgs_xp.load(data, offset)
+            offset = self.lbfgs_g.load(data, offset)
+            offset = self.lbfgs_gp.load(data, offset)
+            offset = self.lbfgs_d.load(data, offset)
+            offset = self.lbfgs_pf.load(data, offset)
+            offset = self.lbfgs_lmal.load(data, offset)
+            offset = self.lbfgs_lmys.load(data, offset)
+            offset = self.lbfgs_lms.load(data, offset)
+            offset = self.lbfgs_lmy.load(data, offset)
+
+            self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError(f"Invalid optimizer type '{self.type}'")
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class LoraParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_rank_attention_norm  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wq              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wk              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wv              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wo              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_ffn_norm        = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w1              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w2              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w3              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_tok_embeddings  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_norm            = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_output          = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,  self.n_rank_tok_embeddings)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, self.n_rank_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT,      self.n_rank_output)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,   self.n_rank_attention_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_Q,      self.n_rank_wq)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_K,      self.n_rank_wk)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_V,      self.n_rank_wv)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,    self.n_rank_wo)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_NORM,    self.n_rank_ffn_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_GATE,    self.n_rank_w1)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,    self.n_rank_w2)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_UP,      self.n_rank_w3)
+
+class ModelParams:
+    def __init__(self, n_ff = None):
+        self.n_ff = n_ff
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        if self.n_ff is None:
+            # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+            return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+        else:
+            return self.n_ff
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None, suffix=".weight"):
+    return gguf.TENSOR_NAMES[key].format(bid=bid) + suffix
+
+class Layer:
+    def __init__(self, params, lora_params, bid):
+        self.bid = bid
+        self.att_norm_a = Tensor('f', [lora_params.n_rank_attention_norm, params.n_embd])
+        self.att_norm_b = Tensor('f', [lora_params.n_rank_attention_norm, 1])
+        self.wq_a       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wq_b       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wk_a       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wk_b       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wv_a       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wv_b       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wo_a       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.wo_b       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.ffn_norm_a = Tensor('f', [lora_params.n_rank_ffn_norm, params.n_embd])
+        self.ffn_norm_b = Tensor('f', [lora_params.n_rank_ffn_norm, 1])
+        self.w1_a       = Tensor('f', [lora_params.n_rank_w1, params.n_embd])
+        self.w1_b       = Tensor('f', [lora_params.n_rank_w1, params.get_n_ff()])
+        self.w2_a       = Tensor('f', [lora_params.n_rank_w2, params.get_n_ff()])
+        self.w2_b       = Tensor('f', [lora_params.n_rank_w2, params.n_embd])
+        self.w3_a       = Tensor('f', [lora_params.n_rank_w3, params.n_embd])
+        self.w3_b       = Tensor('f', [lora_params.n_rank_w3, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm_a.load(data, offset)
+        offset = self.att_norm_b.load(data, offset)
+        offset = self.wq_a.load(data, offset)
+        offset = self.wq_b.load(data, offset)
+        offset = self.wk_a.load(data, offset)
+        offset = self.wk_b.load(data, offset)
+        offset = self.wv_a.load(data, offset)
+        offset = self.wv_b.load(data, offset)
+        offset = self.wo_a.load(data, offset)
+        offset = self.wo_b.load(data, offset)
+        offset = self.ffn_norm_a.load(data, offset)
+        offset = self.ffn_norm_b.load(data, offset)
+        offset = self.w1_a.load(data, offset)
+        offset = self.w1_b.load(data, offset)
+        offset = self.w2_a.load(data, offset)
+        offset = self.w2_b.load(data, offset)
+        offset = self.w3_a.load(data, offset)
+        offset = self.w3_b.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_a"))
+        self.att_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_b"))
+        self.wq_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_a"))
+        self.wq_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_b"))
+        self.wk_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_a"))
+        self.wk_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_b"))
+        self.wv_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_a"))
+        self.wv_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_b"))
+        self.wo_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_a"))
+        self.wo_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_b"))
+        self.ffn_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_a"))
+        self.ffn_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_b"))
+        self.w1_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_a"))
+        self.w1_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_b"))
+        self.w2_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_a"))
+        self.w2_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_b"))
+        self.w3_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_a"))
+        self.w3_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_b"))
+
+class LoraModel:
+    def __init__(self, n_ff = None):
+        self.params = ModelParams(n_ff = n_ff)
+        self.lora_params = LoraParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+        offset = self.lora_params.load(data, offset)
+
+        self.tok_embd_a = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_embd])
+        self.tok_embd_b = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_vocab])
+        self.norm_a     = Tensor('f', [self.lora_params.n_rank_norm, self.params.n_embd])
+        self.norm_b     = Tensor('f', [self.lora_params.n_rank_norm, 1])
+        self.output_a   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_embd])
+        self.output_b   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_vocab])
+
+        offset = self.tok_embd_a.load(data, offset)
+        offset = self.tok_embd_b.load(data, offset)
+        offset = self.norm_a.load(data, offset)
+        offset = self.norm_b.load(data, offset)
+        offset = self.output_a.load(data, offset)
+        offset = self.output_b.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, self.lora_params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+        self.lora_params.save_gguf(gguf_writer)
+
+        self.tok_embd_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_a"))
+        self.tok_embd_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_b"))
+        self.norm_a.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_a"))
+        self.norm_b.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_b"))
+        self.output_a.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_a"))
+        self.output_b.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_b"))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class LoraCheckpoint:
+    def __init__(self, n_ff = None):
+        self.model = LoraModel(n_ff = n_ff)
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcl':
+            raise ValueError(f"File header magic indicates, that this is no finetune-lora checkpoint file. Expected 'ggcl', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_FINETUNE_LORA)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert finetune checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input finetune checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help = 'Output GGUF filename', required=True)
+    parser.add_argument('--ff', type = int, help = "Feedforward size, if not provided compute from n_mult. Provide this if you get 'ValueError: Tensor.load: Expected number of elements does not match what is read from file'", required=False)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(cfg)
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = LoraCheckpoint(n_ff = cfg.ff)
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    print("    gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("    gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("    gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+
+if __name__ == '__main__':
+    main()
diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp
new file mode 100644
index 000000000..9ae4bc198
--- /dev/null
+++ b/examples/finetune/finetune.cpp
@@ -0,0 +1,1940 @@
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "llama.h"
+#include "common.h"
+#include "train.h"
+#include <unordered_map>
+#include <vector>
+#include <cassert>
+#include <climits>
+#include <cstring>
+#include <cstdarg>
+#include <ctime>
+#include <random>
+#include <stdexcept>
+#include <algorithm>
+#include <string>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static const size_t tensor_alignment = 32;
+
+struct my_llama_hparams {
+    uint32_t n_vocab    = 32000;
+    uint32_t n_ctx      = 512;
+    uint32_t n_embd     = 4096;
+    uint32_t n_ff       = 11008;
+    uint32_t n_head     = 32;
+    uint32_t n_head_kv  = 32;
+    uint32_t n_layer    = 32;
+
+    // float f_norm_eps     = 1e-5f; // falcon
+    float f_norm_rms_eps = 1e-5f; // llama
+
+    float rope_freq_base  = 10000.0f;
+    float rope_freq_scale = 1.0f;
+
+    uint32_t n_gqa() const {
+        return n_head/n_head_kv;
+    }
+
+    uint32_t n_embd_head() const {
+        return n_embd/n_head;
+    }
+
+    uint32_t n_embd_gqa() const {
+        return n_embd/n_gqa();
+    }
+
+    bool operator!=(const my_llama_hparams& other) const {
+        return memcmp(this, &other, sizeof(other));
+    }
+};
+
+struct my_llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct my_llama_model {
+    struct my_llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<my_llama_layer> layers;
+};
+
+struct my_llama_lora_hparams {
+    uint32_t lora_r = 1;
+    uint32_t lora_alpha = 1;
+    uint32_t n_rank_attention_norm = 1;
+    uint32_t n_rank_wq = 4;
+    uint32_t n_rank_wk = 4;
+    uint32_t n_rank_wv = 4;
+    uint32_t n_rank_wo = 4;
+    uint32_t n_rank_ffn_norm = 1;
+    uint32_t n_rank_w1 = 4;
+    uint32_t n_rank_w2 = 4;
+    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_tok_embeddings = 4;
+    uint32_t n_rank_norm = 1;
+    uint32_t n_rank_output = 4;
+
+    bool operator!=(const my_llama_lora_hparams& other) const {
+        return memcmp(this, &other, sizeof(other));
+    }
+};
+
+struct my_llama_lora_layer {
+    // normalization
+    struct ggml_tensor * attention_norm_a;
+    struct ggml_tensor * attention_norm_b;
+
+    // attention
+    struct ggml_tensor * wq_a;
+    struct ggml_tensor * wq_b;
+    struct ggml_tensor * wk_a;
+    struct ggml_tensor * wk_b;
+    struct ggml_tensor * wv_a;
+    struct ggml_tensor * wv_b;
+    struct ggml_tensor * wo_a;
+    struct ggml_tensor * wo_b;
+
+    // normalization
+    struct ggml_tensor * ffn_norm_a;
+    struct ggml_tensor * ffn_norm_b;
+
+    // ff
+    struct ggml_tensor * w1_a;
+    struct ggml_tensor * w1_b;
+    struct ggml_tensor * w2_a;
+    struct ggml_tensor * w2_b;
+    struct ggml_tensor * w3_a;
+    struct ggml_tensor * w3_b;
+};
+
+struct my_llama_lora {
+    struct ggml_context * ctx = NULL;
+    std::vector<uint8_t> data;
+
+    my_llama_lora_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings_a;
+    struct ggml_tensor * tok_embeddings_b;
+
+    struct ggml_tensor * norm_a;
+    struct ggml_tensor * norm_b;
+    struct ggml_tensor * output_a;
+    struct ggml_tensor * output_b;
+
+    std::vector<my_llama_lora_layer> layers;
+};
+
+// gguf constants
+static const char * LLM_KV_TRAINING_TYPE_FINETUNE_LORA   = "finetune_lora";
+static const char * LLM_KV_TRAINING_TYPE                 = "training.type";
+
+static const char * LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd";
+static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v";
+static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down";
+static const char * LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up";
+
+// gguf constants (sync with gguf.py)
+
+static const char * LLM_KV_GENERAL_ARCHITECTURE        = "general.architecture";
+static const char * LLM_KV_GENERAL_FILE_TYPE           = "general.file_type";
+
+static const char * LLM_KV_CONTEXT_LENGTH              = "%s.context_length";
+static const char * LLM_KV_EMBEDDING_LENGTH            = "%s.embedding_length";
+static const char * LLM_KV_BLOCK_COUNT                 = "%s.block_count";
+static const char * LLM_KV_FEED_FORWARD_LENGTH         = "%s.feed_forward_length";
+static const char * LLM_KV_ATTENTION_HEAD_COUNT        = "%s.attention.head_count";
+static const char * LLM_KV_ATTENTION_HEAD_COUNT_KV     = "%s.attention.head_count_kv";
+static const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon";
+static const char * LLM_KV_ROPE_DIMENSION_COUNT        = "%s.rope.dimension_count";
+static const char * LLM_KV_ROPE_FREQ_BASE              = "%s.rope.freq_base"; // TODO load in llama.cpp
+static const char * LLM_KV_ROPE_SCALE_LINEAR           = "%s.rope.scale_linear";
+
+static const char * LLM_TENSOR_TOKEN_EMBD    = "token_embd";
+static const char * LLM_TENSOR_OUTPUT_NORM   = "output_norm";
+static const char * LLM_TENSOR_OUTPUT        = "output";
+static const char * LLM_TENSOR_ATTN_NORM     = "blk.%d.attn_norm";
+static const char * LLM_TENSOR_ATTN_Q        = "blk.%d.attn_q";
+static const char * LLM_TENSOR_ATTN_K        = "blk.%d.attn_k";
+static const char * LLM_TENSOR_ATTN_V        = "blk.%d.attn_v";
+static const char * LLM_TENSOR_ATTN_OUT      = "blk.%d.attn_output";
+static const char * LLM_TENSOR_FFN_NORM      = "blk.%d.ffn_norm";
+static const char * LLM_TENSOR_FFN_GATE      = "blk.%d.ffn_gate";
+static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
+static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
+
+static void print_params(struct my_llama_hparams * params) {
+    printf("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:    %u\n", __func__, params->n_embd);
+    printf("%s: n_ff:      %u\n", __func__, params->n_ff);
+    printf("%s: n_head:    %u\n", __func__, params->n_head);
+    printf("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    printf("%s: n_layer:   %u\n", __func__, params->n_layer);
+    printf("%s: norm_rms_eps          : %f\n", __func__, params->f_norm_rms_eps);
+    printf("%s: rope_freq_base        : %f\n", __func__, params->rope_freq_base);
+    printf("%s: rope_freq_scale       : %f\n", __func__, params->rope_freq_scale);
+}
+
+static void print_lora_params(struct my_llama_lora_hparams * params) {
+    printf("%s: n_rank_attention_norm : %u\n", __func__, params->n_rank_attention_norm);
+    printf("%s: n_rank_wq             : %u\n", __func__, params->n_rank_wq);
+    printf("%s: n_rank_wk             : %u\n", __func__, params->n_rank_wk);
+    printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
+    printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
+    printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
+    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
+    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
+    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
+    printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
+    printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
+}
+
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+{ \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        die_fmt("key not found in model: %s", skey.c_str()); \
+    } \
+}
+
+static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_hparams * hparams, const char * expected_arch) {
+    std::string arch;
+
+    GGUF_GET_KEY(ctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
+    if (expected_arch != NULL) {
+        if (arch != expected_arch) {
+            printf("%s: arch=%s expected_arch=%s\n", __func__, arch.c_str(), expected_arch);
+        }
+        GGML_ASSERT(arch == expected_arch);
+    }
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [&arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch.c_str());
+        return keybuf.data();
+    };
+
+    GGUF_GET_KEY(ctx, hparams->n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_EMBEDDING_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_ctx,          gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_FEED_FORWARD_LENGTH));
+    GGUF_GET_KEY(ctx, hparams->n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
+    GGUF_GET_KEY(ctx, hparams->n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32,  true, kv(LLM_KV_BLOCK_COUNT));
+
+    // n_head_kv is optional, default to n_head
+    hparams->n_head_kv = hparams->n_head;
+    GGUF_GET_KEY(ctx, hparams->n_head_kv,      gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV));
+
+    float rope_freq_scale = 1.0f;
+    GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+    GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
+    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    if (rope_freq_scale != 1.0f) {
+        hparams->rope_freq_scale = 1.0f / rope_freq_scale;
+    }
+}
+
+static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t n_ctx) {
+    auto & hparams = model->hparams;
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
+        return tn_buf.data();
+    };
+
+
+    // get parameters directly from gguf file
+    {
+        struct gguf_init_params params = {
+            /*.no_alloc = */ false,
+            /*.ctx      = */ NULL,
+        };
+        struct gguf_context * mctx = gguf_init_from_file(fn_model, params);
+
+        load_model_hparams_gguf(mctx, &hparams, "llama");
+
+        gguf_free(mctx);
+    }
+    hparams.n_vocab = llama_n_vocab(input);
+    hparams.n_ctx = n_ctx;
+
+    // get tensors from llama_model (possibly mmapped)
+    model->tok_embeddings = llama_get_model_tensor(input, tn(LLM_TENSOR_TOKEN_EMBD));
+    model->norm           = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT_NORM));
+    model->output         = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT));
+
+    assert_shape_2d(model->tok_embeddings, hparams.n_embd, hparams.n_vocab);
+    assert_shape_1d(model->norm,           hparams.n_embd);
+    assert_shape_2d(model->output,         hparams.n_embd, hparams.n_vocab);
+
+    model->layers.resize(hparams.n_layer);
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        layer.attention_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_NORM, i));
+        layer.wq             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_Q, i));
+        layer.wk             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_K, i));
+        layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
+        layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
+        layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
+        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+
+        assert_shape_1d(layer.attention_norm, hparams.n_embd);
+        assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
+        assert_shape_2d(layer.wk,             hparams.n_embd, hparams.n_embd_gqa());
+        assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
+        assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
+        assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
+        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+    }
+}
+
+static void set_param_lora(struct my_llama_lora * lora) {
+    const uint32_t n_layer = lora->layers.size();
+
+    struct ggml_context* ctx = lora->ctx;
+
+    ggml_set_param(ctx, lora->tok_embeddings_a);
+    ggml_set_param(ctx, lora->tok_embeddings_b);
+    ggml_set_param(ctx, lora->norm_a);
+    ggml_set_param(ctx, lora->norm_b);
+    ggml_set_param(ctx, lora->output_a);
+    ggml_set_param(ctx, lora->output_b);
+
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+
+        ggml_set_param(ctx, layer.attention_norm_a);
+        ggml_set_param(ctx, layer.attention_norm_b);
+        ggml_set_param(ctx, layer.wq_a);
+        ggml_set_param(ctx, layer.wq_b);
+        ggml_set_param(ctx, layer.wk_a);
+        ggml_set_param(ctx, layer.wk_b);
+        ggml_set_param(ctx, layer.wv_a);
+        ggml_set_param(ctx, layer.wv_b);
+        ggml_set_param(ctx, layer.wo_a);
+        ggml_set_param(ctx, layer.wo_b);
+        ggml_set_param(ctx, layer.ffn_norm_a);
+        ggml_set_param(ctx, layer.ffn_norm_b);
+        ggml_set_param(ctx, layer.w1_a);
+        ggml_set_param(ctx, layer.w1_b);
+        ggml_set_param(ctx, layer.w2_a);
+        ggml_set_param(ctx, layer.w2_b);
+        ggml_set_param(ctx, layer.w3_a);
+        ggml_set_param(ctx, layer.w3_b);
+    }
+}
+
+static void alloc_lora(struct ggml_allocr * alloc, struct my_llama_lora * lora) {
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_a);
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_b);
+    ggml_allocr_alloc(alloc, lora->norm_a);
+    ggml_allocr_alloc(alloc, lora->norm_b);
+    ggml_allocr_alloc(alloc, lora->output_a);
+    ggml_allocr_alloc(alloc, lora->output_b);
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm_a);
+        ggml_allocr_alloc(alloc, layer.attention_norm_b);
+        ggml_allocr_alloc(alloc, layer.wq_a);
+        ggml_allocr_alloc(alloc, layer.wq_b);
+        ggml_allocr_alloc(alloc, layer.wk_a);
+        ggml_allocr_alloc(alloc, layer.wk_b);
+        ggml_allocr_alloc(alloc, layer.wv_a);
+        ggml_allocr_alloc(alloc, layer.wv_b);
+        ggml_allocr_alloc(alloc, layer.wo_a);
+        ggml_allocr_alloc(alloc, layer.wo_b);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_a);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_b);
+        ggml_allocr_alloc(alloc, layer.w1_a);
+        ggml_allocr_alloc(alloc, layer.w1_b);
+        ggml_allocr_alloc(alloc, layer.w2_a);
+        ggml_allocr_alloc(alloc, layer.w2_b);
+        ggml_allocr_alloc(alloc, layer.w3_a);
+        ggml_allocr_alloc(alloc, layer.w3_b);
+    }
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_a->grad);
+    ggml_allocr_alloc(alloc, lora->tok_embeddings_b->grad);
+    ggml_allocr_alloc(alloc, lora->norm_a->grad);
+    ggml_allocr_alloc(alloc, lora->norm_b->grad);
+    ggml_allocr_alloc(alloc, lora->output_a->grad);
+    ggml_allocr_alloc(alloc, lora->output_b->grad);
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm_a->grad);
+        ggml_allocr_alloc(alloc, layer.attention_norm_b->grad);
+        ggml_allocr_alloc(alloc, layer.wq_a->grad);
+        ggml_allocr_alloc(alloc, layer.wq_b->grad);
+        ggml_allocr_alloc(alloc, layer.wk_a->grad);
+        ggml_allocr_alloc(alloc, layer.wk_b->grad);
+        ggml_allocr_alloc(alloc, layer.wv_a->grad);
+        ggml_allocr_alloc(alloc, layer.wv_b->grad);
+        ggml_allocr_alloc(alloc, layer.wo_a->grad);
+        ggml_allocr_alloc(alloc, layer.wo_b->grad);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_a->grad);
+        ggml_allocr_alloc(alloc, layer.ffn_norm_b->grad);
+        ggml_allocr_alloc(alloc, layer.w1_a->grad);
+        ggml_allocr_alloc(alloc, layer.w1_b->grad);
+        ggml_allocr_alloc(alloc, layer.w2_a->grad);
+        ggml_allocr_alloc(alloc, layer.w2_b->grad);
+        ggml_allocr_alloc(alloc, layer.w3_a->grad);
+        ggml_allocr_alloc(alloc, layer.w3_b->grad);
+    }
+}
+
+static void init_lora(const struct my_llama_model * model, struct my_llama_lora * lora) {
+    const auto & lparams = lora->hparams;
+
+    const uint32_t n_embd     = model->hparams.n_embd;
+    const uint32_t n_embd_gqa = model->hparams.n_embd_gqa();
+    const uint32_t n_layer    = model->hparams.n_layer;
+    const uint32_t n_vocab    = model->hparams.n_vocab;
+    const uint32_t n_ff       = model->hparams.n_ff;
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, const char * suffix, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix);
+        return tn_buf.data();
+    };
+
+    // context for lora tensors without their data
+    struct ggml_init_params ctx_lora_params;
+    ctx_lora_params.mem_size   = ggml_tensor_overhead()*2*(6 + n_layer*18);
+    ctx_lora_params.mem_buffer = NULL;
+    ctx_lora_params.no_alloc   = true;
+
+    struct ggml_context * ctx = ggml_init(ctx_lora_params);
+    lora->ctx = ctx;
+
+    lora->tok_embeddings_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_embd);
+    lora->tok_embeddings_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_vocab);
+    lora->norm_a           = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, n_embd);
+    lora->norm_b           = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, 1);
+    lora->output_a         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_embd);
+    lora->output_b         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_vocab);
+
+    ggml_set_name(lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.lora_a"));
+    ggml_set_name(lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.lora_b"));
+    ggml_set_name(lora->norm_a,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_a"));
+    ggml_set_name(lora->norm_b,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_b"));
+    ggml_set_name(lora->output_a,         tn(LLM_TENSOR_OUTPUT,      ".weight.lora_a"));
+    ggml_set_name(lora->output_b,         tn(LLM_TENSOR_OUTPUT,      ".weight.lora_b"));
+
+    lora->layers.resize(n_layer);
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+
+        layer.attention_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, n_embd);
+        layer.attention_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, 1);
+
+        layer.wq_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd);
+        layer.wq_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd);
+        layer.wk_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd);
+        layer.wk_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd_gqa);
+        layer.wv_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd);
+        layer.wv_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd_gqa);
+        layer.wo_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd);
+        layer.wo_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd);
+
+        layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
+        layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
+
+        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
+        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
+        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
+        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
+        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
+        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+
+        ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
+        ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
+        ggml_set_name(layer.wq_a,             tni(LLM_TENSOR_ATTN_Q,    ".weight.lora_a", i));
+        ggml_set_name(layer.wq_b,             tni(LLM_TENSOR_ATTN_Q,    ".weight.lora_b", i));
+        ggml_set_name(layer.wk_a,             tni(LLM_TENSOR_ATTN_K,    ".weight.lora_a", i));
+        ggml_set_name(layer.wk_b,             tni(LLM_TENSOR_ATTN_K,    ".weight.lora_b", i));
+        ggml_set_name(layer.wv_a,             tni(LLM_TENSOR_ATTN_V,    ".weight.lora_a", i));
+        ggml_set_name(layer.wv_b,             tni(LLM_TENSOR_ATTN_V,    ".weight.lora_b", i));
+        ggml_set_name(layer.wo_a,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_a", i));
+        ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
+        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+    }
+
+    set_param_lora(lora);
+
+    // measure data size
+    struct ggml_allocr * alloc = NULL;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc_lora(alloc, lora);
+
+    // allocate data
+    lora->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    ggml_allocr_free(alloc);
+    alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
+    alloc_lora(alloc, lora);
+    ggml_allocr_free(alloc);
+}
+
+static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) {
+    const uint32_t n_layer = lora->layers.size();
+
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
+
+    randomize_tensor_normal(lora->tok_embeddings_a, rnd);
+    randomize_tensor_normal(lora->tok_embeddings_b, rnd);
+    randomize_tensor_normal(lora->norm_a,           rnd);
+    randomize_tensor_normal(lora->norm_b,           rnd);
+    randomize_tensor_normal(lora->output_a,         rnd);
+    randomize_tensor_normal(lora->output_b,         rnd);
+
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = lora->layers[i];
+        randomize_tensor_normal(layer.attention_norm_a, rnd);
+        randomize_tensor_normal(layer.attention_norm_b, rnd);
+
+        randomize_tensor_normal(layer.wq_a, rnd);
+        randomize_tensor_normal(layer.wq_b, rnd);
+        randomize_tensor_normal(layer.wk_a, rnd);
+        randomize_tensor_normal(layer.wk_b, rnd);
+        randomize_tensor_normal(layer.wv_a, rnd);
+        randomize_tensor_normal(layer.wv_b, rnd);
+        randomize_tensor_normal(layer.wo_a, rnd);
+        randomize_tensor_normal(layer.wo_b, rnd);
+
+        randomize_tensor_normal(layer.ffn_norm_a, rnd);
+        randomize_tensor_normal(layer.ffn_norm_b, rnd);
+
+        randomize_tensor_normal(layer.w1_a, rnd);
+        randomize_tensor_normal(layer.w1_b, rnd);
+        randomize_tensor_normal(layer.w2_a, rnd);
+        randomize_tensor_normal(layer.w2_b, rnd);
+        randomize_tensor_normal(layer.w3_a, rnd);
+        randomize_tensor_normal(layer.w3_b, rnd);
+    }
+
+    free_random_normal_distribution(rnd);
+}
+
+static struct ggml_tensor * llama_build_lora_finetune_graphs(
+        struct my_llama_model * model,
+        struct my_llama_lora  * lora,
+        struct ggml_allocr    * alloc,
+        struct ggml_context   * ctx,
+        struct ggml_cgraph    * gf,
+        struct ggml_cgraph    * gb,
+        struct ggml_cgraph    * gb_tmp,
+        struct ggml_tensor  * * logits,
+        struct ggml_tensor    * tokens_input,
+        struct ggml_tensor    * targets,
+        const  int              n_tokens,
+        const  int              n_batch,
+        const  bool             enable_flash_attn,
+        const  bool             enable_checkpointing) {
+
+    ggml_set_scratch(ctx, { 0, 0, nullptr, });
+    const int n_past = 0;
+    const int N = n_tokens;
+    const auto & hparams  = model->hparams;
+    const int n_ctx       = hparams.n_ctx;
+    const int n_vocab     = hparams.n_vocab;
+    const int n_embd      = hparams.n_embd;
+    const int n_layer     = hparams.n_layer;
+    const int n_head      = hparams.n_head;
+    const int n_head_kv   = hparams.n_head_kv;
+    const int n_ff        = hparams.n_ff;
+    const int n_rot       = hparams.n_embd_head();
+    const int n_embd_head = hparams.n_embd_head();
+    const int n_embd_gqa  = hparams.n_embd_gqa();
+    const float rms_norm_eps    = hparams.f_norm_rms_eps;
+    const float rope_freq_base  = hparams.rope_freq_base;
+    const float rope_freq_scale = hparams.rope_freq_scale;
+
+    GGML_ASSERT((size_t) n_layer == lora->layers.size());
+
+    auto set_name = [](struct ggml_tensor * t, const char * n) {
+        ggml_set_name(t, n);
+        if (t->grad) {
+            ggml_format_name(t->grad, "%s->grad", n);
+        }
+    };
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
+        }
+    }
+
+    // rope has so much parameters that we make a custom function for it
+    auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
+                (struct ggml_tensor * t) -> struct ggml_tensor * {
+        // not capturing these, to silcence warnings
+        const int rope_mode = 0;
+
+        return ggml_rope_custom(ctx,
+            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            rope_freq_base, rope_freq_scale);
+    };
+
+    set_name(tokens_input, "tokens_input");
+    set_name(targets,      "targets");
+
+    GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
+
+    auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
+        if (ggml_is_quantized(a->type)) {
+            return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
+        } else if (a->type == GGML_TYPE_F32) {
+            return ggml_add(ctx, a, b);
+        } else {
+            die_fmt("%s: Finetuning on tensors with type '%s' is not yet supported.\n",
+                __func__, ggml_type_name(a->type));
+        }
+    };
+
+    struct ggml_tensor * tok_embeddings = add_to_f32(ctx, model->tok_embeddings, ggml_mul_mat(ctx, lora->tok_embeddings_a, lora->tok_embeddings_b));
+    struct ggml_tensor * norm           = add_to_f32(ctx, model->norm, ggml_mul_mat(ctx, lora->norm_a, lora->norm_b));
+    struct ggml_tensor * output         = add_to_f32(ctx, model->output, ggml_mul_mat(ctx, lora->output_a, lora->output_b));
+
+    struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch);  set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch);
+    struct ggml_tensor * t01 = ggml_get_rows(ctx, tok_embeddings, t00);        set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch);
+
+    struct ggml_tensor * cur = t01;
+
+    std::vector<struct ggml_tensor *> checkpoints;
+    if (enable_checkpointing) {
+        checkpoints.push_back(tokens_input);
+        checkpoints.push_back(targets);
+        checkpoints.push_back(t00);
+        checkpoints.push_back(t01);
+    }
+
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct my_llama_layer & layer = model->layers[il];
+        struct my_llama_lora_layer & llayer = lora->layers[il];
+
+        struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
+        struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
+        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
+        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
+        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+
+        struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
+        struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
+        struct ggml_tensor * t04 = ggml_mul          (ctx, t03, t02);                                set_name(t04, "t04");     assert_shape_2d(t04, n_embd, N*n_batch);
+        struct ggml_tensor * t05 = ggml_mul_mat      (ctx, wq, t04);                                 set_name(t05, "t05");     assert_shape_2d(t05, n_embd, N*n_batch);
+        struct ggml_tensor * t06 = ggml_reshape_4d   (ctx, t05, n_embd_head, n_head, N, n_batch);    set_name(t06, "t06");     assert_shape_4d(t06, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t07 = rope              (t06);                                          set_name(t07, "t07");     assert_shape_4d(t07, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t08 = ggml_mul_mat      (ctx, wk, t04);                                 set_name(t08, "t08");     assert_shape_2d(t08, n_embd_gqa, N*n_batch);
+        struct ggml_tensor * t09 = ggml_reshape_4d   (ctx, t08, n_embd_head, n_head_kv, N, n_batch); set_name(t09, "t09");     assert_shape_4d(t09, n_embd_head, n_head_kv, N, n_batch);
+        struct ggml_tensor * t10 = rope              (t09);                                          set_name(t10, "t10");     assert_shape_4d(t10, n_embd_head, n_head_kv, N, n_batch);
+
+        struct ggml_tensor * t11;
+        if (ggml_is_quantized(wv->type)) {
+            struct ggml_tensor * t11_1 = ggml_mul_mat  (ctx, wv, t04);                               set_name(t11_1, "t11_1"); assert_shape_2d(t11_1, n_embd_gqa, N*n_batch);
+            struct ggml_tensor * t11_2 = ggml_transpose(ctx, t11_1);                                 set_name(t11_2, "t11_2"); assert_shape_2d(t11_2, N*n_batch, n_embd_gqa);
+                                 t11   = ggml_cont     (ctx, t11_2);                                 set_name(t11, "t11");     assert_shape_2d(t11, N*n_batch, n_embd_gqa);
+        } else {
+                                 t11   = ggml_mul_mat  (ctx, t04, wv);                               set_name(t11, "t11");     assert_shape_2d(t11, N*n_batch, n_embd_gqa);
+        }
+
+        struct ggml_tensor * t12 = ggml_reshape_4d   (ctx, t11, N, n_batch, n_embd_head, n_head_kv); set_name(t12, "t12");     assert_shape_4d(t12, N, n_batch, n_embd_head, n_head_kv);
+        struct ggml_tensor * t13 = ggml_permute      (ctx, t07, 0, 2, 1, 3);                         set_name(t13, "t13");     assert_shape_4d(t13, n_embd_head, N, n_head, n_batch);
+        struct ggml_tensor * t14 = ggml_permute      (ctx, t10, 0, 2, 1, 3);                         set_name(t14, "t14");     assert_shape_4d(t14, n_embd_head, N, n_head_kv, n_batch);
+        struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
+        struct ggml_tensor * t16;
+        if (enable_flash_attn) {
+            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+        } else {
+            struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past);             set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_3 = ggml_soft_max_inplace     (ctx, t16_2);                     set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch);
+            t16 = ggml_mul_mat(ctx, t15, t16_3);                                                     set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+        }
+        struct ggml_tensor * t17 = ggml_permute      (ctx, t16, 0, 2, 1, 3);                         set_name(t17, "t17");     assert_shape_4d(t17, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t18 = ggml_cont         (ctx, t17);                                     set_name(t18, "t18");     assert_shape_4d(t18, n_embd_head, n_head, N, n_batch);
+        struct ggml_tensor * t19 = ggml_reshape_2d   (ctx, t18, n_embd, N*n_batch);                  set_name(t19, "t19");     assert_shape_2d(t19, n_embd, N*n_batch);
+        struct ggml_tensor * t20 = ggml_mul_mat      (ctx, wo, t19);                                 set_name(t20, "t20");     assert_shape_2d(t20, n_embd, N*n_batch);
+        struct ggml_tensor * t21 = ggml_add          (ctx, t20, cur);                                set_name(t21, "t21");     assert_shape_2d(t21, n_embd, N*n_batch);
+        struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
+        struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
+        struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
+        struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
+        cur = t30;
+        if (enable_checkpointing) {
+            checkpoints.push_back(cur);
+        }
+    }
+    struct ggml_tensor * t31   = ggml_rms_norm          (ctx, cur, rms_norm_eps);                    set_name(t31, "t31");     assert_shape_2d(t31, n_embd, N*n_batch);
+    struct ggml_tensor * t32   = ggml_repeat            (ctx, norm, t31);                            set_name(t32, "t32");     assert_shape_2d(t32, n_embd, N*n_batch);
+    struct ggml_tensor * t33   = ggml_mul               (ctx, t32, t31);                             set_name(t33, "t33");     assert_shape_2d(t33, n_embd, N*n_batch);
+    struct ggml_tensor * t34   = ggml_mul_mat           (ctx, output, t33);                          set_name(t34, "t34");     assert_shape_2d(t34, n_vocab, N*n_batch);
+    struct ggml_tensor * t35   = ggml_reshape_3d        (ctx, t34, n_vocab, N, n_batch);             set_name(t35, "t35");     assert_shape_3d(t35, n_vocab, N, n_batch);
+    struct ggml_tensor * t36   = ggml_cross_entropy_loss(ctx, t35, targets);                         set_name(t36, "t36");     assert_shape_1d(t36, 1);
+
+    if (enable_checkpointing) {
+        checkpoints.push_back(t31);
+        checkpoints.push_back(t32);
+        checkpoints.push_back(t33);
+        checkpoints.push_back(t34);
+        checkpoints.push_back(t35);
+        checkpoints.push_back(t36);
+    }
+
+    ggml_build_forward_expand(gf, t36);
+
+    if (enable_checkpointing) {
+        ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
+    } else {
+        *gb = *gf;
+        ggml_build_backward_expand(ctx, gf, gb, true);
+    }
+
+    GGML_ASSERT(alloc != NULL);
+
+    // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
+    int n_leafs_before = gb->n_leafs;
+    int n_nodes_before = gb->n_nodes;
+    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
+    // output tensors
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+    // input gradient
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+    GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
+    ggml_allocr_alloc(alloc, t36->grad);
+    // KQ_pos
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+
+    // make sure base model tensors data cannot be used in viewable operations
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
+    for (int il = 0; il < n_layer; ++il) {
+        struct my_llama_layer & layer = model->layers[il];
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
+    }
+
+    // allocating checkpoints in one block to reduce memory fragmentation
+    // note: they will be freed in reverse order
+    for (unsigned int i = 0; i < checkpoints.size(); ++i) {
+        if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
+            ggml_allocr_alloc(alloc, checkpoints[i]);
+        }
+    }
+
+    ggml_allocr_alloc_graph(alloc, gb);
+
+    // remove the additional nodes and leafs
+    for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
+        gb->leafs[i] = NULL;
+    }
+    for (int i = n_nodes_before; i < gb->n_nodes; ++i) {
+        gb->nodes[i] = NULL;
+    }
+    gb->n_leafs = n_leafs_before;
+    gb->n_nodes = n_nodes_before;
+
+    *logits = t35;
+    return t36;
+}
+
+static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora) {
+    // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
+
+    std::string arch;
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+
+    GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
+    GGML_ASSERT(arch == "llama");
+
+    uint32_t ftype_u;
+    GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE);
+    GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32);
+
+    struct my_llama_hparams hparams;
+    load_model_hparams_gguf(fctx, &hparams, arch.c_str());
+
+    // parameters that define tensor shapes must match
+    GGML_ASSERT(hparams.n_embd    == model->hparams.n_embd);
+    GGML_ASSERT(hparams.n_ff      == model->hparams.n_ff);
+    GGML_ASSERT(hparams.n_head    == model->hparams.n_head);
+    GGML_ASSERT(hparams.n_head_kv == model->hparams.n_head_kv);
+    GGML_ASSERT(hparams.n_layer   == model->hparams.n_layer);
+
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_tok_embeddings, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_norm,           gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_output,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_attention_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wq,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_Q);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wk,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_K);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+
+    init_lora(model, lora);
+
+    copy_tensor_by_name(lora->tok_embeddings_a, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_a));
+    copy_tensor_by_name(lora->tok_embeddings_b, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_b));
+    copy_tensor_by_name(lora->norm_a,           f_ggml_ctx, ggml_get_name(lora->norm_a));
+    copy_tensor_by_name(lora->norm_b,           f_ggml_ctx, ggml_get_name(lora->norm_b));
+    copy_tensor_by_name(lora->output_a,         f_ggml_ctx, ggml_get_name(lora->output_a));
+    copy_tensor_by_name(lora->output_b,         f_ggml_ctx, ggml_get_name(lora->output_b));
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        copy_tensor_by_name(layer.attention_norm_a, f_ggml_ctx, ggml_get_name(layer.attention_norm_a));
+        copy_tensor_by_name(layer.attention_norm_b, f_ggml_ctx, ggml_get_name(layer.attention_norm_b));
+        copy_tensor_by_name(layer.wq_a,             f_ggml_ctx, ggml_get_name(layer.wq_a));
+        copy_tensor_by_name(layer.wq_b,             f_ggml_ctx, ggml_get_name(layer.wq_b));
+        copy_tensor_by_name(layer.wk_a,             f_ggml_ctx, ggml_get_name(layer.wk_a));
+        copy_tensor_by_name(layer.wk_b,             f_ggml_ctx, ggml_get_name(layer.wk_b));
+        copy_tensor_by_name(layer.wv_a,             f_ggml_ctx, ggml_get_name(layer.wv_a));
+        copy_tensor_by_name(layer.wv_b,             f_ggml_ctx, ggml_get_name(layer.wv_b));
+        copy_tensor_by_name(layer.wo_a,             f_ggml_ctx, ggml_get_name(layer.wo_a));
+        copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
+        copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
+        copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
+        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
+        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
+        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
+        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
+        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
+        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+    }
+}
+
+static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora) {
+    const char * arch = "llama";
+    enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
+
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch);
+        return keybuf.data();
+    };
+
+    gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
+    gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
+
+    gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH),              model->hparams.n_ctx);
+    gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH),            model->hparams.n_embd);
+    gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH),         model->hparams.n_ff);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT),        model->hparams.n_head);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV),     model->hparams.n_head_kv);
+    gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT),                 model->hparams.n_layer);
+    gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT),        model->hparams.n_embd_head());
+    gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps);
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE),              model->hparams.rope_freq_base);
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR),           model->hparams.rope_freq_scale);
+
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,   lora->hparams.n_rank_tok_embeddings);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM,  lora->hparams.n_rank_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT,       lora->hparams.n_rank_output);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,    lora->hparams.n_rank_attention_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_Q,       lora->hparams.n_rank_wq);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_K,       lora->hparams.n_rank_wk);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+
+    gguf_add_tensor(fctx, lora->tok_embeddings_a);
+    gguf_add_tensor(fctx, lora->tok_embeddings_b);
+    gguf_add_tensor(fctx, lora->norm_a);
+    gguf_add_tensor(fctx, lora->norm_b);
+    gguf_add_tensor(fctx, lora->output_a);
+    gguf_add_tensor(fctx, lora->output_b);
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+
+        gguf_add_tensor(fctx, layer.attention_norm_a);
+        gguf_add_tensor(fctx, layer.attention_norm_b);
+        gguf_add_tensor(fctx, layer.wq_a);
+        gguf_add_tensor(fctx, layer.wq_b);
+        gguf_add_tensor(fctx, layer.wk_a);
+        gguf_add_tensor(fctx, layer.wk_b);
+        gguf_add_tensor(fctx, layer.wv_a);
+        gguf_add_tensor(fctx, layer.wv_b);
+        gguf_add_tensor(fctx, layer.wo_a);
+        gguf_add_tensor(fctx, layer.wo_b);
+        gguf_add_tensor(fctx, layer.ffn_norm_a);
+        gguf_add_tensor(fctx, layer.ffn_norm_b);
+        gguf_add_tensor(fctx, layer.w1_a);
+        gguf_add_tensor(fctx, layer.w1_b);
+        gguf_add_tensor(fctx, layer.w2_a);
+        gguf_add_tensor(fctx, layer.w2_b);
+        gguf_add_tensor(fctx, layer.w3_a);
+        gguf_add_tensor(fctx, layer.w3_b);
+    }
+}
+
+static void load_checkpoint_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    std::string train_type = LLM_KV_TRAINING_TYPE_FINETUNE_LORA;
+    GGUF_GET_KEY(fctx, train_type, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE);
+    GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_FINETUNE_LORA);
+
+    load_train_state_gguf(fctx, f_ggml_ctx, train);
+    load_llama_lora_gguf(fctx, f_ggml_ctx, model, lora);
+}
+
+static void save_checkpoint_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    gguf_set_val_str(fctx, LLM_KV_TRAINING_TYPE, LLM_KV_TRAINING_TYPE_FINETUNE_LORA);
+    save_llama_lora_gguf(fctx, model, lora);
+    save_train_state_gguf(fctx, train);
+}
+
+static bool load_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    struct ggml_context * f_ggml_ctx;
+    struct gguf_init_params params;
+    params.no_alloc = false;
+    params.ctx = &f_ggml_ctx;
+    struct gguf_context * fctx = gguf_init_from_file(filename, params);
+    if (fctx == NULL) {
+        return false;
+    }
+
+    load_checkpoint_lora_gguf(fctx, f_ggml_ctx, model, lora, train);
+
+    gguf_free(fctx);
+    return true;
+}
+
+static void save_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct gguf_context * fctx = gguf_init_empty();
+
+    save_checkpoint_lora_gguf(fctx, model, lora, train);
+
+    // write file
+    const bool only_meta = false;
+    gguf_write_to_file(fctx, filename, only_meta);
+    gguf_free(fctx);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
+        } else {
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
+        }
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor, const char * name) {
+    if (tensor == NULL) {
+        file->write_u32(0);
+        file->write_u32(0);
+        file->write_u32(GGML_TYPE_F32);
+        file->seek((0-file->tell()) & 31, SEEK_CUR);
+        return;
+    }
+    if (name == NULL) {
+        name = ggml_get_name(tensor);
+    }
+    uint32_t name_len = strlen(name);
+    uint32_t nd = tensor->n_dims;
+    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
+                       (uint32_t)tensor->ne[1],
+                       (uint32_t)tensor->ne[2],
+                       (uint32_t)tensor->ne[3] };
+    file->write_u32(nd);
+    file->write_u32(name_len);
+    file->write_u32(tensor->type);
+    file->write_raw(ne, sizeof(ne[0]) * nd);
+    file->write_raw(name, name_len);
+    file->seek((0-file->tell()) & 31, SEEK_CUR);
+    file->write_raw(tensor->data, ggml_nbytes(tensor));
+}
+
+static void save_as_llama_lora(const char * filename, struct my_llama_lora * lora) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct llama_file file(filename, "wb");
+    if (file.fp == NULL) {
+        return;
+    }
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+
+    auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix);
+        return tn_buf.data();
+    };
+
+    auto tni = [&tn_buf](const char * key, int bid, const char * suffix) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix);
+        return tn_buf.data();
+    };
+
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
+    // write_magic
+    file.write_u32(LLAMA_FILE_MAGIC_LORA);   // magic
+    file.write_u32(1); // version
+    // write_hparams
+    file.write_u32(lora->hparams.lora_r);
+    file.write_u32(lora->hparams.lora_alpha);
+    // write tensors
+    write_tensor(&file, lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.loraA"));
+    write_tensor(&file, lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD,  ".weight.loraB"));
+    write_tensor(&file, lora->norm_a,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraA"));
+    write_tensor(&file, lora->norm_b,           tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraB"));
+    write_tensor(&file, lora->output_a,         tn(LLM_TENSOR_OUTPUT,      ".weight.loraA"));
+    write_tensor(&file, lora->output_b,         tn(LLM_TENSOR_OUTPUT,      ".weight.loraB"));
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        write_tensor(&file, layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraA"));
+        write_tensor(&file, layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraB"));
+        write_tensor(&file, layer.wq_a,             tni(LLM_TENSOR_ATTN_Q,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wq_b,             tni(LLM_TENSOR_ATTN_Q,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wk_a,             tni(LLM_TENSOR_ATTN_K,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wk_b,             tni(LLM_TENSOR_ATTN_K,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wv_a,             tni(LLM_TENSOR_ATTN_V,    i, ".weight.loraA"));
+        write_tensor(&file, layer.wv_b,             tni(LLM_TENSOR_ATTN_V,    i, ".weight.loraB"));
+        write_tensor(&file, layer.wo_a,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraA"));
+        write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+    }
+}
+
+struct train_params {
+    struct train_params_common common;
+
+    const char * fn_model_base;
+    const char * fn_lora_out;
+
+    bool only_write_lora;
+
+    float f_norm_rms_eps;
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    bool custom_f_norm_rms_eps;
+    bool custom_rope_freq_base;
+    bool custom_rope_freq_scale;
+
+    int32_t lora_r;
+    int32_t lora_alpha;
+    bool custom_lora_alpha;
+
+    uint32_t n_rank_attention_norm;
+    uint32_t n_rank_wq;
+    uint32_t n_rank_wk;
+    uint32_t n_rank_wv;
+    uint32_t n_rank_wo;
+    uint32_t n_rank_ffn_norm;
+    uint32_t n_rank_w1;
+    uint32_t n_rank_w2;
+    uint32_t n_rank_w3;
+    uint32_t n_rank_tok_embeddings;
+    uint32_t n_rank_norm;
+    uint32_t n_rank_output;
+
+    bool custom_n_rank_attention_norm;
+    bool custom_n_rank_wq;
+    bool custom_n_rank_wk;
+    bool custom_n_rank_wv;
+    bool custom_n_rank_wo;
+    bool custom_n_rank_ffn_norm;
+    bool custom_n_rank_w1;
+    bool custom_n_rank_w2;
+    bool custom_n_rank_w3;
+    bool custom_n_rank_tok_embeddings;
+    bool custom_n_rank_norm;
+    bool custom_n_rank_output;
+};
+
+static struct train_params get_default_train_params() {
+    struct train_params params;
+    params.common = get_default_train_params_common();
+    params.fn_model_base     = "";
+    params.fn_lora_out       = "ggml-lora-ITERATION-f32.gguf";
+
+    params.only_write_lora = false;
+
+    params.f_norm_rms_eps  = 1e-5f;
+    params.rope_freq_base  = 10000.0f;
+    params.rope_freq_scale = 1.0f;
+
+    params.custom_f_norm_rms_eps  = false;
+    params.custom_rope_freq_base  = false;
+    params.custom_rope_freq_scale = false;
+
+    params.lora_r      = 4;
+    params.lora_alpha  = 4;
+    params.custom_lora_alpha = false;
+
+    params.n_rank_attention_norm = 1;
+    params.n_rank_wq             = 4;
+    params.n_rank_wk             = 4;
+    params.n_rank_wv             = 4;
+    params.n_rank_wo             = 4;
+    params.n_rank_ffn_norm       = 1;
+    params.n_rank_w1             = 4;
+    params.n_rank_w2             = 4;
+    params.n_rank_w3             = 4;
+    params.n_rank_tok_embeddings = 4;
+    params.n_rank_norm           = 1;
+    params.n_rank_output         = 4;
+
+    params.custom_n_rank_attention_norm = false;
+    params.custom_n_rank_wq             = false;
+    params.custom_n_rank_wk             = false;
+    params.custom_n_rank_wv             = false;
+    params.custom_n_rank_wo             = false;
+    params.custom_n_rank_ffn_norm       = false;
+    params.custom_n_rank_w1             = false;
+    params.custom_n_rank_w2             = false;
+    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_tok_embeddings = false;
+    params.custom_n_rank_norm           = false;
+    params.custom_n_rank_output         = false;
+
+    return params;
+}
+
+static void train_print_usage(int argc, char ** argv, const struct train_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                 show this help message and exit\n");
+
+    fprintf(stderr, "  --model-base FNAME         model path from which to load base model (default '%s')\n", params->fn_model_base);
+    fprintf(stderr, "  --lora-out FNAME           path to save llama lora (default '%s')\n", params->fn_lora_out);
+    fprintf(stderr, "  --only-write-lora          only save llama lora, don't do any training.  use this if you only want to convert a checkpoint to a lora adapter.\n");
+    fprintf(stderr, "  --norm-rms-eps F           RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps);
+    fprintf(stderr, "  --rope-freq-base F         Frequency base for ROPE (default %f)\n", params->rope_freq_base);
+    fprintf(stderr, "  --rope-freq-scale F        Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
+    fprintf(stderr, "  --lora-alpha N             LORA alpha : resulting LORA scaling is alpha/r. (default %d)\n", params->lora_alpha);
+    fprintf(stderr, "  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default %d)\n", params->lora_r);
+    fprintf(stderr, "  --rank-att-norm N          LORA rank for attention norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-ffn-norm N          LORA rank for feed-forward norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-out-norm N          LORA rank for output norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n");
+    fprintf(stderr, "  --rank-tok-embd N          LORA rank for token embeddings tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-out N               LORA rank for output tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wq N                LORA rank for wq tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+
+    print_common_train_usage(argc, argv, &params->common);
+}
+
+static bool train_params_parse(int argc, char ** argv, struct train_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct train_params default_params = get_default_train_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (consume_common_train_arg(argc, argv, &i, &params->common, &invalid_param)) {
+            if (invalid_param) {
+                break;
+            } else if (params->common.print_usage) {
+                train_print_usage(argc, argv, &default_params);
+                exit(0);
+            }
+        } else if (arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "--lora-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_lora_out = argv[i];
+        } else if (arg == "--only-write-lora") {
+            params->only_write_lora = true;
+        } else if (arg == "--norm-rms-eps") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->f_norm_rms_eps = std::stof(argv[i]);
+            params->custom_f_norm_rms_eps = true;
+        } else if (arg == "--rope-freq-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->rope_freq_base = std::stof(argv[i]);
+            params->custom_rope_freq_base = true;
+        } else if (arg == "--rope-freq-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->rope_freq_scale = std::stof(argv[i]);
+            params->custom_rope_freq_scale = true;
+        } else if (arg == "--lora-alpha") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->lora_alpha = std::stoi(argv[i]);
+            params->custom_lora_alpha = true;
+        } else if (arg == "--lora-r") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->lora_r = std::stoi(argv[i]);
+        } else if (arg == "--rank-att-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_attention_norm = std::stoi(argv[i]);
+            params->custom_n_rank_attention_norm = true;
+        } else if (arg == "--rank-ffn-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_ffn_norm = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_norm = true;
+        } else if (arg == "--rank-out-norm") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_norm = std::stoi(argv[i]);
+            params->custom_n_rank_norm = true;
+        } else if (arg == "--rank-tok-embd") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_tok_embeddings = std::stoi(argv[i]);
+            params->custom_n_rank_tok_embeddings = true;
+        } else if (arg == "--rank-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_output = std::stoi(argv[i]);
+            params->custom_n_rank_output = true;
+        } else if (arg == "--rank-wq") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wq = std::stoi(argv[i]);
+            params->custom_n_rank_wq = true;
+        } else if (arg == "--rank-wk") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wk = std::stoi(argv[i]);
+            params->custom_n_rank_wk = true;
+        } else if (arg == "--rank-wv") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wv = std::stoi(argv[i]);
+            params->custom_n_rank_wv = true;
+        } else if (arg == "--rank-wo") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_wo = std::stoi(argv[i]);
+            params->custom_n_rank_wo = true;
+        } else if (arg == "--rank-w1") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w1 = std::stoi(argv[i]);
+            params->custom_n_rank_w1 = true;
+        } else if (arg == "--rank-w2") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w2 = std::stoi(argv[i]);
+            params->custom_n_rank_w2 = true;
+        } else if (arg == "--rank-w3") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_rank_w3 = std::stoi(argv[i]);
+            params->custom_n_rank_w3 = true;
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            train_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        train_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    finish_processing_train_args(&params->common);
+    return true;
+}
+
+struct save_train_files_data {
+    const char            * fn_checkpoint_out;
+    const char            * fn_lora_out;
+    const char            * pattern_fn_it;
+    const char            * fn_latest;
+    struct my_llama_model * model;
+    struct my_llama_lora  * lora;
+};
+
+static void save_train_files(void * vdata, struct train_state * train) {
+    struct save_train_files_data * data   = (struct save_train_files_data *) vdata;
+
+    int64_t iter = train->opt->iter;
+
+    if (strlen(data->fn_checkpoint_out) > 0) {
+        save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->model, data->lora, train);
+        save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->model, data->lora, train);
+    }
+    if (strlen(data->fn_lora_out) > 0) {
+        save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->lora);
+        save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->lora);
+    }
+}
+
+static int64_t get_parameter_count(struct my_llama_lora* lora) {
+    int64_t nx = 0;
+    nx += ggml_nelements(lora->tok_embeddings_a);
+    nx += ggml_nelements(lora->tok_embeddings_b);
+    nx += ggml_nelements(lora->norm_a);
+    nx += ggml_nelements(lora->norm_b);
+    nx += ggml_nelements(lora->output_a);
+    nx += ggml_nelements(lora->output_b);
+
+    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
+        auto & layer = lora->layers[i];
+        nx += ggml_nelements(layer.attention_norm_a);
+        nx += ggml_nelements(layer.attention_norm_b);
+        nx += ggml_nelements(layer.wq_a);
+        nx += ggml_nelements(layer.wq_b);
+        nx += ggml_nelements(layer.wk_a);
+        nx += ggml_nelements(layer.wk_b);
+        nx += ggml_nelements(layer.wv_a);
+        nx += ggml_nelements(layer.wv_b);
+        nx += ggml_nelements(layer.wo_a);
+        nx += ggml_nelements(layer.wo_b);
+        nx += ggml_nelements(layer.ffn_norm_a);
+        nx += ggml_nelements(layer.ffn_norm_b);
+        nx += ggml_nelements(layer.w1_a);
+        nx += ggml_nelements(layer.w1_b);
+        nx += ggml_nelements(layer.w2_a);
+        nx += ggml_nelements(layer.w2_b);
+        nx += ggml_nelements(layer.w3_a);
+        nx += ggml_nelements(layer.w3_b);
+    }
+    return nx;
+}
+
+int main(int argc, char ** argv) {
+    struct train_params params = get_default_train_params();
+
+    if (!train_params_parse(argc, argv, &params)) {
+        return 1;
+    }
+
+    if (params.common.seed == LLAMA_DEFAULT_SEED) {
+        params.common.seed = time(NULL);
+    }
+    printf("%s: seed: %u\n", __func__, params.common.seed);
+    srand(params.common.seed);
+
+    struct llama_model_params llama_mparams = llama_model_default_params();
+    llama_mparams.vocab_only = false;
+
+    printf("%s: model base = '%s'\n", __func__, params.fn_model_base);
+    struct llama_model * lmodel = llama_load_model_from_file(params.fn_model_base, llama_mparams);
+
+    struct llama_context_params llama_cparams = llama_context_default_params();
+    struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_cparams);
+
+    struct my_llama_model model;
+    init_model(lmodel, &model, params.fn_model_base, params.common.n_ctx);
+
+    struct my_llama_lora lora;
+
+    struct train_state      * train = init_train_state();
+    struct ggml_opt_context * opt   = train->opt;
+
+    // set params from command line
+    if (params.custom_f_norm_rms_eps) {
+        model.hparams.f_norm_rms_eps  = params.f_norm_rms_eps;
+    }
+    if (params.custom_rope_freq_base) {
+        model.hparams.rope_freq_base  = params.rope_freq_base;
+    }
+    if (params.custom_rope_freq_scale) {
+        model.hparams.rope_freq_scale = params.rope_freq_scale;
+    }
+    lora.hparams.lora_r                = params.lora_r;
+    lora.hparams.lora_alpha            = params.custom_lora_alpha            ? params.lora_alpha            : params.lora_r;
+    uint32_t n_rank_attention_norm     = params.custom_n_rank_attention_norm ? params.n_rank_attention_norm : 1;
+    uint32_t n_rank_wq                 = params.custom_n_rank_wq             ? params.n_rank_wq             : params.lora_r;
+    uint32_t n_rank_wk                 = params.custom_n_rank_wk             ? params.n_rank_wk             : params.lora_r;
+    uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
+    uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
+    uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
+    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
+    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
+    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
+    uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
+    uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
+    lora.hparams.n_rank_attention_norm = n_rank_attention_norm;
+    lora.hparams.n_rank_wq             = n_rank_wq;
+    lora.hparams.n_rank_wk             = n_rank_wk;
+    lora.hparams.n_rank_wv             = n_rank_wv;
+    lora.hparams.n_rank_wo             = n_rank_wo;
+    lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
+    lora.hparams.n_rank_w1             = n_rank_w1;
+    lora.hparams.n_rank_w2             = n_rank_w2;
+    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
+    lora.hparams.n_rank_norm           = n_rank_norm;
+    lora.hparams.n_rank_output         = n_rank_output;
+
+    // set opt params from command line
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    opt->params.print_forward_graph     = false;
+    opt->params.print_backward_graph    = false;
+    opt->params.n_threads               = params.common.n_threads;
+    opt->params.past                    = params.common.opt_past;
+    opt->params.delta                   = params.common.opt_delta;
+    opt->params.max_no_improvement      = params.common.opt_max_no_improvement;
+    opt->params.n_gradient_accumulation = params.common.n_gradient_accumulation;
+    opt->params.adam.n_iter             = params.common.adam_n_iter;
+    opt->params.adam.sched              = 1.0f;
+    opt->params.adam.alpha              = params.common.adam_alpha;
+    opt->params.adam.decay              = params.common.adam_decay;
+    opt->params.adam.decay_min_ndim     = params.common.adam_decay_min_ndim;
+    opt->params.adam.beta1              = params.common.adam_beta1;
+    opt->params.adam.beta2              = params.common.adam_beta2;
+    opt->params.adam.gclip              = params.common.adam_gclip;
+    opt->params.adam.eps_f              = params.common.adam_eps_f;
+
+    ggml_allocr * alloc = NULL;
+
+    printf("%s: init model\n", __func__);
+    bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
+
+    if (existed) {
+        // overwrite last n_ctx with user provided n_ctx
+        if (params.common.custom_n_ctx) {
+            model.hparams.n_ctx = params.common.n_ctx;
+        }
+
+        const bool opt_param_count_changed = (
+           (lora.hparams.n_rank_attention_norm != n_rank_attention_norm)
+        || (lora.hparams.n_rank_wq             != n_rank_wq)
+        || (lora.hparams.n_rank_wk             != n_rank_wk)
+        || (lora.hparams.n_rank_wv             != n_rank_wv)
+        || (lora.hparams.n_rank_wo             != n_rank_wo)
+        || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
+        || (lora.hparams.n_rank_w1             != n_rank_w1)
+        || (lora.hparams.n_rank_w2             != n_rank_w2)
+        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
+        || (lora.hparams.n_rank_norm           != n_rank_norm)
+        || (lora.hparams.n_rank_output         != n_rank_output)
+        );
+
+        const bool opt_past_changed = opt->params.past != params.common.opt_past;
+
+        if (opt_param_count_changed) {
+            print_lora_params(&lora.hparams);
+            die("Provided rank differs from checkpoint file. To use different rank start finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting.");
+            // need to discard previous optimizer gradient statistics and opt_init with new shapes
+            // TODO
+        }
+        if (opt_past_changed) {
+            die("Optimizer parameter '--opt-past N' differs from checkpoint file. To use different value finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting");
+            // need to discard previous optimizer past function value statistics and opt_init with new shapes
+            // TODO
+        }
+    } else { // existed == false
+        init_lora(&model, &lora);
+        randomize_lora(&lora, params.common.seed, 0.0f, 1.0f, -1.0f, +1.0f);
+        if (!params.only_write_lora) {
+            ggml_opt_init(opt->ctx, opt, opt->params, get_parameter_count(&lora));
+        }
+    }
+    opt->iter = train->train_its;
+
+    print_params(&model.hparams);
+    print_lora_params(&lora.hparams);
+    printf("%s: total train_iterations %llu\n", __func__, (long long unsigned) train->train_its);
+    printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
+    printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
+    printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
+    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + lora.data.size()), (float) (ggml_used_mem(lora.ctx) + lora.data.size()) / (1024.0f*1024.0f));
+
+    if (params.only_write_lora) {
+        save_train_files_data save_data;
+        save_data.fn_checkpoint_out = "";
+        save_data.fn_lora_out       = params.fn_lora_out;
+        save_data.pattern_fn_it     = params.common.pattern_fn_it;
+        save_data.fn_latest         = params.common.fn_latest;
+        save_data.model             = &model;
+        save_data.lora              = &lora;
+
+        save_train_files(&save_data, train);
+
+        free_train_state(train);
+        ggml_free(lora.ctx);
+        llama_free(lctx);
+        llama_free_model(lmodel);
+        return 0;
+    }
+
+    printf("%s: opt_size  = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f));
+    printf("%s: opt iter %d\n", __func__, opt->iter);
+
+    int n_tokens = model.hparams.n_ctx;
+    int n_vocab  = model.hparams.n_vocab;
+    int n_batch  = params.common.n_batch;
+
+
+    std::vector<uint8_t> mem_input_data;
+    std::vector<uint8_t> mem_compute_data;
+
+    // context for input tensors without their data
+    struct ggml_init_params ctx_input_params = {
+        ggml_tensor_overhead() * 2, // mem_size
+        NULL,                       // mem_buffer
+        true,                       // no_alloc
+    };
+    struct ggml_context * ctx_input = ggml_init(ctx_input_params);
+
+    // the input tensors
+    struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
+    struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
+
+    // measure required memory for input tensors
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+    ggml_allocr_free(alloc);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
+
+    // allocate input tensors
+    mem_input_data.resize(max_input_size);
+    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    ggml_allocr_free(alloc);
+
+    // context for compute tensors without their data
+    size_t estimated_compute_size_wo_data = (
+        ggml_tensor_overhead()*GGML_MAX_NODES*2
+      + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*(
+            params.common.use_checkpointing ? 3 : 2
+        )
+    );
+    struct ggml_init_params ctx_compute_params = {
+        estimated_compute_size_wo_data, // mem_size
+        NULL,                           // mem_buffer
+        true,                           // no_alloc
+    };
+    struct ggml_context * ctx_compute = NULL;
+
+    struct ggml_tensor * loss   = NULL;
+    struct ggml_tensor * logits = NULL;
+
+    struct ggml_cgraph * gf     = NULL;
+    struct ggml_cgraph * gb     = NULL;
+    struct ggml_cgraph * gb_tmp = NULL;
+
+    // measure required memory for compute tensors
+    size_t best_compute_size = SIZE_MAX;
+    enum ggml_cgraph_eval_order best_order = GGML_CGRAPH_EVAL_ORDER_COUNT;
+    // find best evaluation order
+    for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
+        ctx_compute = ggml_init(ctx_compute_params);
+        alloc = ggml_allocr_new_measure(tensor_alignment);
+        gf = ggml_new_graph(ctx_compute);
+        gf->order = (enum ggml_cgraph_eval_order) order;
+        gb = ggml_new_graph(ctx_compute);
+        gb_tmp = params.common.use_checkpointing
+            ? ggml_new_graph(ctx_compute)
+            : NULL;
+        loss = llama_build_lora_finetune_graphs(
+            &model, &lora, alloc, ctx_compute,
+            gf, gb, gb_tmp,
+            &logits, tokens_input, target_probs,
+            n_tokens, n_batch,
+            params.common.use_flash,
+            params.common.use_checkpointing
+        );
+        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        if (max_compute_size < best_compute_size) {
+            best_compute_size = max_compute_size;
+            best_order = gf->order;
+        }
+        ggml_allocr_free(alloc);
+        ggml_free(ctx_compute);
+    }
+    size_t max_compute_size = best_compute_size;
+    printf("%s: compute_size = %zu bytes (%.1f MB)\n", __func__, max_compute_size, (float) max_compute_size / (1024.0f*1024.0f));
+    printf("%s: evaluation order = %s\n", __func__,
+        (best_order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? "LEFT_TO_RIGHT" :
+        (best_order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? "RIGHT_TO_LEFT" :
+        "invalid");
+
+    // allocate compute tensors
+    mem_compute_data.resize(max_compute_size);
+    ctx_compute = ggml_init(ctx_compute_params);
+    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    gf = ggml_new_graph(ctx_compute);
+    gf->order = best_order;
+    gb = ggml_new_graph(ctx_compute);
+    gb_tmp = params.common.use_checkpointing
+        ? ggml_new_graph(ctx_compute)
+        : NULL;
+    loss = llama_build_lora_finetune_graphs(
+        &model, &lora, alloc, ctx_compute,
+        gf, gb, gb_tmp,
+        &logits, tokens_input, target_probs,
+        n_tokens, n_batch,
+        params.common.use_flash,
+        params.common.use_checkpointing
+    );
+    ggml_allocr_free(alloc);
+
+    // tokenize data
+    std::vector<llama_token> train_tokens;
+    std::vector<size_t> train_samples_begin;
+    std::vector<size_t> train_samples_size;
+    printf("%s: tokenize training data\n", __func__);
+    tokenize_file(lctx,
+            params.common.fn_train_data,
+            params.common.sample_start,
+            params.common.include_sample_start,
+            params.common.overlapping_samples,
+            n_tokens,
+            train_tokens,
+            train_samples_begin,
+            train_samples_size);
+    GGML_ASSERT(train_samples_begin.size() == train_samples_size.size());
+
+    printf("%s: number of training tokens: %zu\n", __func__, train_tokens.size());
+
+    std::vector<size_t> token_noccurs;
+    token_noccurs.resize(model.hparams.n_vocab, 0);
+    for (unsigned int i = 0; i < train_tokens.size(); ++i) {
+        ++token_noccurs[train_tokens[i]];
+    }
+    int n_unique_tokens = 0;
+    for (unsigned int i = 0; i < token_noccurs.size(); ++i) {
+        if (token_noccurs[i] == 0) continue;
+        ++n_unique_tokens;
+    }
+    printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens);
+
+    size_t shuffle_samples_hash = compute_samples_hash(params.common.fn_train_data, train_samples_begin.data(), train_samples_size.data(), train_samples_size.size());
+    const bool changed_train_data = (shuffle_samples_hash != train->shuffle_samples_hash) || (train->shuffle_sample_count != train_samples_size.size());
+    if (changed_train_data) {
+        printf("%s: train data seems to have changed. restarting shuffled epoch.\n", __func__);
+    }
+    if (params.common.force_reshuffle) {
+        printf("%s: forced reshuffling of data. restarting with newly shuffled epoch.\n", __func__);
+    }
+    if ((train->shuffle_rng_state_current == "") || changed_train_data || params.common.force_reshuffle) {
+        train->shuffle_rng_state_current = mt19937_seed_to_state(params.common.seed);
+        train->shuffle_sample_count = train_samples_size.size();
+        train->shuffle_next_sample = 0;
+        train->shuffle_samples_hash = shuffle_samples_hash;
+    }
+    std::vector<size_t> train_shuffled_samples_offs;
+    std::vector<size_t> train_shuffled_samples_begin;
+    std::vector<size_t> train_shuffled_samples_size;
+    train_shuffled_samples_offs.resize(train_samples_begin.size());
+    train_shuffled_samples_begin.resize(train_samples_begin.size());
+    train_shuffled_samples_size.resize(train_samples_size.size());
+    train->shuffle_rng_state_next = shuffle_samples(
+        train->shuffle_rng_state_current,
+        train_shuffled_samples_offs.data(),
+        train_shuffled_samples_begin.data(),
+        train_shuffled_samples_size.data(),
+        train_samples_begin.data(),
+        train_samples_size.data(),
+        train_samples_size.size());
+
+    printf("%s: begin training\n", __func__);
+
+    save_train_files_data save_data;
+    save_data.fn_checkpoint_out = params.common.fn_checkpoint_out;
+    save_data.fn_lora_out       = params.fn_lora_out;
+    save_data.pattern_fn_it     = params.common.pattern_fn_it;
+    save_data.fn_latest         = params.common.fn_latest;
+    save_data.model             = &model;
+    save_data.lora              = &lora;
+
+    struct train_opt_callback_data opt_cb_data;
+    opt_cb_data.params                 = &params.common;
+    opt_cb_data.train                  = train;
+    opt_cb_data.save_cb                = &save_train_files;
+    opt_cb_data.save_data              = &save_data;
+    opt_cb_data.lctx                   = lctx;
+    opt_cb_data.last_save_iter         = opt->iter;
+    opt_cb_data.tokens_data            = train_tokens.data();
+    opt_cb_data.tokens_size            = train_tokens.size();
+    opt_cb_data.samples_begin          = train_samples_begin.data();
+    opt_cb_data.samples_size           = train_samples_size.data();
+    opt_cb_data.shuffled_samples_offs  = train_shuffled_samples_offs.data();
+    opt_cb_data.shuffled_samples_begin = train_shuffled_samples_begin.data();
+    opt_cb_data.shuffled_samples_size  = train_shuffled_samples_size.data();
+    opt_cb_data.samples_count          = train_samples_size.size();
+    opt_cb_data.tokens_input           = tokens_input;
+    opt_cb_data.target_probs           = target_probs;
+    opt_cb_data.first_iter             = opt->iter;
+    opt_cb_data.first_epoch            = train->train_epochs;
+    opt_cb_data.iter_at_last_epoch     = -1;
+    opt_cb_data.last_time              = ggml_time_ms();
+    opt_cb_data.millis_per_iter        = 0.0;
+
+    // measure required memory for work buffer
+    size_t max_work_size = ggml_graph_plan(gb, params.common.n_threads).work_size + GGML_OBJECT_SIZE;
+    printf("%s: work_size = %zu bytes (%.1f MB)\n", __func__, max_work_size, (float) max_work_size / (1024.0f*1024.0f));
+
+    // context for work buffer
+    struct ggml_init_params ctx_work_params = {
+        max_work_size, // mem_size
+        NULL,          // mem_buffer
+        false,         // no_alloc
+    };
+    struct ggml_context * ctx_work = ggml_init(ctx_work_params);
+
+    int64_t t0 = ggml_time_ms();
+
+    ggml_opt_resume_g(ctx_work, opt, loss, gf, gb, &train_opt_callback, (void *) &opt_cb_data);
+
+    ggml_free(ctx_work);
+    ggml_free(ctx_compute);
+    ggml_free(ctx_input);
+
+    int64_t t1 = ggml_time_ms();
+    printf("%s: total training time: ", __func__);
+    print_duration((double) (t1 - t0));
+    printf("\n");
+
+    int new_iters = opt->iter - opt_cb_data.last_save_iter;
+    if (new_iters > 0) {
+        train->train_its     += new_iters;
+        train->train_tokens  += new_iters * opt->params.n_gradient_accumulation * n_batch * n_tokens;
+
+        save_train_files(&save_data, train);
+        opt_cb_data.last_save_iter = opt->iter;
+    }
+
+    ggml_free(opt->ctx);
+    free_train_state(train);
+    ggml_free(lora.ctx);
+    llama_free(lctx);
+    llama_free_model(lmodel);
+    return 0;
+}
diff --git a/examples/gguf/CMakeLists.txt b/examples/gguf/CMakeLists.txt
new file mode 100644
index 000000000..7d1806af3
--- /dev/null
+++ b/examples/gguf/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET gguf)
+add_executable(${TARGET} gguf.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp
index dee00df87..9ab63a293 100644
--- a/examples/gguf/gguf.cpp
+++ b/examples/gguf/gguf.cpp
@@ -13,14 +13,14 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-template<typename T>
+template <typename T>
 static std::string to_string(const T & val) {
     std::stringstream ss;
     ss << val;
     return ss.str();
 }
 
-bool gguf_ex_write(const std::string & fname) {
+static bool gguf_ex_write(const std::string & fname) {
     struct gguf_context * ctx = gguf_init_empty();
 
     gguf_set_val_u8  (ctx, "some.parameter.uint8",    0x12);
@@ -30,6 +30,9 @@ bool gguf_ex_write(const std::string & fname) {
     gguf_set_val_u32 (ctx, "some.parameter.uint32",   0x12345678);
     gguf_set_val_i32 (ctx, "some.parameter.int32",   -0x12345679);
     gguf_set_val_f32 (ctx, "some.parameter.float32",  0.123456789f);
+    gguf_set_val_u64 (ctx, "some.parameter.uint64",   0x123456789abcdef0ull);
+    gguf_set_val_i64 (ctx, "some.parameter.int64",   -0x123456789abcdef1ll);
+    gguf_set_val_f64 (ctx, "some.parameter.float64",  0.1234567890123456789);
     gguf_set_val_bool(ctx, "some.parameter.bool",     true);
     gguf_set_val_str (ctx, "some.parameter.string",   "hello world");
 
@@ -73,7 +76,7 @@ bool gguf_ex_write(const std::string & fname) {
 
     gguf_write_to_file(ctx, fname.c_str(), false);
 
-    fprintf(stdout, "%s: wrote file '%s;\n", __func__, fname.c_str());
+    printf("%s: wrote file '%s;\n", __func__, fname.c_str());
 
     ggml_free(ctx_data);
     gguf_free(ctx);
@@ -82,7 +85,7 @@ bool gguf_ex_write(const std::string & fname) {
 }
 
 // just read tensor info
-bool gguf_ex_read_0(const std::string & fname) {
+static bool gguf_ex_read_0(const std::string & fname) {
     struct gguf_init_params params = {
         /*.no_alloc = */ false,
         /*.ctx      = */ NULL,
@@ -90,20 +93,20 @@ bool gguf_ex_read_0(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
 
     // kv
     {
         const int n_kv = gguf_get_n_kv(ctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
 
@@ -113,10 +116,10 @@ bool gguf_ex_read_0(const std::string & fname) {
 
         const int keyidx = gguf_find_key(ctx, findkey);
         if (keyidx == -1) {
-            fprintf(stdout, "%s: find key: %s not found.\n", __func__, findkey);
+            printf("%s: find key: %s not found.\n", __func__, findkey);
         } else {
             const char * key_value = gguf_get_val_str(ctx, keyidx);
-            fprintf(stdout, "%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
+            printf("%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);
         }
     }
 
@@ -124,13 +127,13 @@ bool gguf_ex_read_0(const std::string & fname) {
     {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -140,7 +143,7 @@ bool gguf_ex_read_0(const std::string & fname) {
 }
 
 // read and create ggml_context containing the tensors and their data
-bool gguf_ex_read_1(const std::string & fname) {
+static bool gguf_ex_read_1(const std::string & fname) {
     struct ggml_context * ctx_data = NULL;
 
     struct gguf_init_params params = {
@@ -150,20 +153,20 @@ bool gguf_ex_read_1(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
-    fprintf(stdout, "%s: version:      %d\n", __func__, gguf_get_version(ctx));
-    fprintf(stdout, "%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
-    fprintf(stdout, "%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
+    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
+    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
+    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));
 
     // kv
     {
         const int n_kv = gguf_get_n_kv(ctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
 
@@ -171,13 +174,13 @@ bool gguf_ex_read_1(const std::string & fname) {
     {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
 
@@ -186,13 +189,13 @@ bool gguf_ex_read_1(const std::string & fname) {
         const int n_tensors = gguf_get_n_tensors(ctx);
 
         for (int i = 0; i < n_tensors; ++i) {
-            fprintf(stdout, "%s: reading tensor %d data\n", __func__, i);
+            printf("%s: reading tensor %d data\n", __func__, i);
 
             const char * name = gguf_get_tensor_name(ctx, i);
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            fprintf(stdout, "%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;
@@ -216,7 +219,7 @@ bool gguf_ex_read_1(const std::string & fname) {
         }
     }
 
-    fprintf(stdout, "%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));
+    printf("%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));
 
     ggml_free(ctx_data);
     gguf_free(ctx);
@@ -226,7 +229,7 @@ bool gguf_ex_read_1(const std::string & fname) {
 
 int main(int argc, char ** argv) {
     if (argc < 3) {
-        fprintf(stdout, "usage: %s data.gguf r|w\n", argv[0]);
+        printf("usage: %s data.gguf r|w\n", argv[0]);
         return -1;
     }
 
diff --git a/examples/gptneox-wip/falcon-main.cpp b/examples/gptneox-wip/falcon-main.cpp
index 43b6a29f3..e9197f6b5 100644
--- a/examples/gptneox-wip/falcon-main.cpp
+++ b/examples/gptneox-wip/falcon-main.cpp
@@ -305,9 +305,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)
 
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
     } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
     }
 
     return cur;
@@ -333,21 +333,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         return false;
     }
 
-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
 
     // print all kv
     #if 0
     {
         const int n_kv = gguf_get_n_kv(ggufctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ggufctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
     #endif
@@ -357,21 +357,21 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         int keyidx;
 
         keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 
     // check required metadata
@@ -382,11 +382,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "falcon") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
             return false;
         }
 
@@ -394,11 +394,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "falcon.tensor_data_layout");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "jploski") != 0) {
-                fprintf(stdout, "%s: model tensor data layout not supported!\n", __func__);
+                printf("%s: model tensor data layout not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model tensor data layout not found!\n", __func__);
+            printf("%s: gguf model tensor data layout not found!\n", __func__);
             return false;
         }
 
@@ -455,11 +455,11 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
 
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
             return false;
         }
 
@@ -467,22 +467,22 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
 
         if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
             return false;
         }
 
         int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");
 
         if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
             return false;
         }
 
         hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
         hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);
 
-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
 
         for (size_t i = 0; i < hparams.n_vocab; i++) {
             std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -523,12 +523,12 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
 
-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
 
     }
 
@@ -543,13 +543,13 @@ bool falcon_model_load(const std::string & fname, falcon_model & model, gpt2bpe_
     {
         const int n_tensors = gguf_get_n_tensors(ggufctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ggufctx, i);
             const size_t offset = gguf_get_tensor_offset(ggufctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
     #endif
@@ -953,7 +953,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
diff --git a/examples/gptneox-wip/gptneox-main.cpp b/examples/gptneox-wip/gptneox-main.cpp
index 04af50245..b76bafaa8 100644
--- a/examples/gptneox-wip/gptneox-main.cpp
+++ b/examples/gptneox-wip/gptneox-main.cpp
@@ -318,9 +318,9 @@ struct ggml_tensor * get_tensor_ex( struct ggml_context * ctx, std::string name)
 
     struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
     if( cur == NULL ) {
-        fprintf(stdout, "%s: tensor '%s' not found!\n", __func__, name.c_str());
+        printf("%s: tensor '%s' not found!\n", __func__, name.c_str());
     } else {
-//        fprintf(stdout, "%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
+//        printf("%s: n_dims = %d, name = '%s'\n", __func__, cur->n_dims, cur->name);
     }
 
     return cur;
@@ -346,21 +346,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         return false;
     }
 
-    fprintf(stdout, "%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
-    fprintf(stdout, "%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
-    fprintf(stdout, "%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
+    printf("%s: gguf version     = %d\n", __func__, gguf_get_version(ggufctx));
+    printf("%s: gguf alignment   = %zu\n", __func__, gguf_get_alignment(ggufctx));
+    printf("%s: gguf data offset = %zu\n", __func__, gguf_get_data_offset(ggufctx));
 
     // print all kv
     #if 0
     {
         const int n_kv = gguf_get_n_kv(ggufctx);
 
-        fprintf(stdout, "%s: n_kv: %d\n", __func__, n_kv);
+        printf("%s: n_kv: %d\n", __func__, n_kv);
 
         for (int i = 0; i < n_kv; ++i) {
             const char * key = gguf_get_key(ggufctx, i);
 
-            fprintf(stdout, "%s: kv[%d]: key = %s\n", __func__, i, key);
+            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
         }
     }
     #endif
@@ -370,21 +370,21 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         int keyidx;
 
         keyidx = gguf_find_key(ggufctx, "general.name");
-        if (keyidx != -1) { fprintf(stdout, "%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model name           = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.description");
-        if (keyidx != -1) { fprintf(stdout, "%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model description    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.author");
-        if (keyidx != -1) { fprintf(stdout, "%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model author         = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.license");
-        if (keyidx != -1) { fprintf(stdout, "%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model license        = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.architecture");
-        if (keyidx != -1) { fprintf(stdout, "%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model architecture   = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "general.file_type");
-        if (keyidx != -1) { fprintf(stdout, "%s: model file type      = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model file type      = %" PRIu32 "\n", __func__, gguf_get_val_u32(ggufctx, keyidx)); }
         keyidx = gguf_find_key(ggufctx, "gptneox.tensor_data_layout");
-        if (keyidx != -1) { fprintf(stdout, "%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
-        keyidx = gguf_find_key(ggufctx, "general.source.hugginface.repository");
-        if (keyidx != -1) { fprintf(stdout, "%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        if (keyidx != -1) { printf("%s: model data layout    = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
+        keyidx = gguf_find_key(ggufctx, "general.source.huggingface.repository");
+        if (keyidx != -1) { printf("%s: model source HF repo = %s\n", __func__, gguf_get_val_str(ggufctx, keyidx)); }
     }
 
     // check required metadata
@@ -395,11 +395,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         keyidx = gguf_find_key(ggufctx, "general.architecture");
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gptneox") != 0) {
-                fprintf(stdout, "%s: model architecture not supported!\n", __func__);
+                printf("%s: model architecture not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: gguf model architecture not found!\n", __func__);
+            printf("%s: gguf model architecture not found!\n", __func__);
             return false;
         }
 
@@ -456,11 +456,11 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
 
         if (keyidx != -1) {
             if ( strcmp(gguf_get_val_str(ggufctx, keyidx), "gpt2") != 0) {
-                fprintf(stdout, "%s: tokenizer model not supported!\n", __func__);
+                printf("%s: tokenizer model not supported!\n", __func__);
                 return false;
             }
         } else {
-            fprintf(stdout, "%s: tokenizer model not found!\n", __func__);
+            printf("%s: tokenizer model not found!\n", __func__);
             return false;
         }
 
@@ -468,22 +468,22 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         int tokens_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.tokens");
 
         if (tokens_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer vocab not found!\n", __func__);
+            printf("%s: gpt2 tokenizer vocab not found!\n", __func__);
             return false;
         }
 
         int merges_keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.merges");
 
         if (merges_keyidx == -1) {
-            fprintf(stdout, "%s: gpt2 tokenizer merges not found!\n", __func__);
+            printf("%s: gpt2 tokenizer merges not found!\n", __func__);
             return false;
         }
 
         hparams.n_vocab = gguf_get_arr_n(ggufctx,tokens_keyidx);
         hparams.n_merges = gguf_get_arr_n(ggufctx,merges_keyidx);
 
-        fprintf(stdout, "%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
-        fprintf(stdout, "%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
+        printf("%s: gpt2 tokenizer vocab  = %zu\n", __func__, hparams.n_vocab);
+        printf("%s: gpt2 tokenizer merges = %zu\n", __func__, hparams.n_merges);
 
         for (size_t i = 0; i < hparams.n_vocab; i++) {
             std::string word = gguf_get_arr_str(ggufctx, tokens_keyidx, i);
@@ -524,12 +524,12 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.separator_token_id"); if( keyidx != -1 ) { vocab.special_sep_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
         keyidx = gguf_find_key(ggufctx, "tokenizer.ggml.padding_token_id"); if( keyidx != -1 ) {   vocab.special_pad_id = (int32_t)gguf_get_val_u32(ggufctx, keyidx); }
 
-        if( vocab.special_bos_id != -1 ) { fprintf(stdout, "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
-        if( vocab.special_eos_id != -1 ) { fprintf(stdout, "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
-        if( vocab.special_unk_id != -1 ) { fprintf(stdout, "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
-        if( vocab.special_sep_id != -1 ) { fprintf(stdout, "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
-        if( vocab.special_pad_id != -1 ) { fprintf(stdout, "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
-        if( vocab.linefeed_id    != -1 ) { fprintf(stdout, "%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
+        if( vocab.special_bos_id != -1 ) { printf("%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].c_str() ); }
+        if( vocab.special_eos_id != -1 ) { printf("%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].c_str() ); }
+        if( vocab.special_unk_id != -1 ) { printf("%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].c_str() ); }
+        if( vocab.special_sep_id != -1 ) { printf("%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].c_str() ); }
+        if( vocab.special_pad_id != -1 ) { printf("%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].c_str() ); }
+        if( vocab.linefeed_id    != -1 ) { printf("%s: LF token  = %d\n",      __func__, vocab.linefeed_id ); }
     }
 
 
@@ -543,13 +543,13 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
     {
         const int n_tensors = gguf_get_n_tensors(ggufctx);
 
-        fprintf(stdout, "%s: n_tensors: %d\n", __func__, n_tensors);
+        printf("%s: n_tensors: %d\n", __func__, n_tensors);
 
         for (int i = 0; i < n_tensors; ++i) {
             const char * name   = gguf_get_tensor_name  (ggufctx, i);
             const size_t offset = gguf_get_tensor_offset(ggufctx, i);
 
-            fprintf(stdout, "%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
+            printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
         }
     }
     #endif
@@ -660,9 +660,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2
 ggml_tensor * gpt_neox_ff(
         const gpt_neox_block &block,
         ggml_context * ctx0,
-        ggml_tensor * inp) {
+        ggml_tensor * inp,
+        const gpt_neox_hparams &hparams) {
 
-    ggml_tensor * cur = ggml_norm(ctx0, inp);
+    ggml_tensor * cur = ggml_norm(ctx0, inp, hparams.norm_eps);
 
     cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, block.ln_2_g, cur), cur), ggml_repeat(ctx0, block.ln_2_b, cur));
     cur = ggml_mul_mat(ctx0, block.c_mlp_fc_w, cur);
@@ -753,7 +754,7 @@ bool gpt_neox_eval(
         // self-attention
         {
             {
-                cur = ggml_norm(ctx0, inpL);
+                cur = ggml_norm(ctx0, inpL, hparams.norm_eps);
 
                 cur = ggml_add(ctx0,
                         ggml_mul(ctx0, ggml_repeat(ctx0, model.blocks[il].ln_1_g, cur), cur),
@@ -844,7 +845,7 @@ bool gpt_neox_eval(
         if (hparams.par_res == 0) {
             struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
 
-            cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF);
+            cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF, hparams);
 
             // input for next layer
             inpL = ggml_add(ctx0, cur, inpFF);
@@ -853,7 +854,7 @@ bool gpt_neox_eval(
 
             // this is independent of the self-attention result, so it could be done in parallel to the self-attention
             // note here we pass inpL instead of cur
-            cur = gpt_neox_ff(model.blocks[il], ctx0, inpL);
+            cur = gpt_neox_ff(model.blocks[il], ctx0, inpL, hparams);
 
             // layer input + FF
             cur  = ggml_add(ctx0, cur, inpFF);
@@ -867,7 +868,7 @@ bool gpt_neox_eval(
 
     // norm
     {
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, hparams.norm_eps);
 
         // inpL = ln_f_g*inpL + ln_f_b
         inpL = ggml_add(ctx0,
@@ -924,7 +925,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
diff --git a/examples/infill/CMakeLists.txt b/examples/infill/CMakeLists.txt
new file mode 100644
index 000000000..046f9b1e7
--- /dev/null
+++ b/examples/infill/CMakeLists.txt
@@ -0,0 +1,8 @@
+set(TARGET infill)
+add_executable(${TARGET} infill.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()
diff --git a/examples/infill/README.md b/examples/infill/README.md
new file mode 100644
index 000000000..8c97f719b
--- /dev/null
+++ b/examples/infill/README.md
@@ -0,0 +1,41 @@
+# llama.cpp/example/infill
+
+This example shows how to use the infill mode with Code Llama models supporting infill mode.
+Currently the 7B and 13B models support infill mode.
+
+Infill supports most of the options available in the main example.
+
+For further information have a look at the main README.md in llama.cpp/example/main/README.md
+
+## Common Options
+
+In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
+
+-   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
+-   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
+-   `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
+
+## Input Prompts
+
+The `infill` program provides several ways to interact with the LLaMA models using input prompts:
+
+-   `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
+-   `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
+-   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
+
+## Interaction
+
+The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
+
+### Interaction Options
+
+-   `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
+-   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
+-   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
+
+### Example
+
+```bash
+./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
+```
diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp
new file mode 100644
index 000000000..9ec75ce42
--- /dev/null
+++ b/examples/infill/infill.cpp
@@ -0,0 +1,769 @@
+#include "common.h"
+
+#include "console.h"
+#include "llama.h"
+#include "build-info.h"
+#include "grammar-parser.h"
+
+#include <cassert>
+#include <cinttypes>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+#include <signal.h>
+#include <unistd.h>
+#elif defined (_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <signal.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static gpt_params               * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
+static bool is_interacting = false;
+
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = get_sortable_timestamp();
+
+    const bool success = create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: infill\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+static void sigint_handler(int signo) {
+    if (signo == SIGINT) {
+        if (!is_interacting) {
+            is_interacting = true;
+        } else {
+            console::cleanup();
+            printf("\n");
+            llama_print_timings(*g_ctx);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+            _exit(130);
+        }
+    }
+}
+#endif
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    g_params = &params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("infill", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
+    if (params.logits_all) {
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.embedding) {
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        params.n_ctx = 8;
+    }
+    if (params.instruct) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for instruct mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.antiprompt.empty()) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
+        printf("\n************\n");
+        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (params.random_prompt) {
+        printf("\n************\n");
+        printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+    if (!params.path_prompt_cache.empty()) {
+        printf("\n************\n");
+        printf("%s: infill does not support prompt caching\n", __func__);
+        printf("************\n\n");
+
+        return 0;
+    }
+
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    LOG("%s: llama backend init\n", __func__);
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+    llama_context * ctx_guidance = NULL;
+    g_model = &model;
+    g_ctx = &ctx;
+
+    // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (params.cfg_scale > 1.f) {
+        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
+        ctx_guidance = llama_new_context_with_model(model, lparams);
+    }
+
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
+    // print system information
+    {
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", get_system_info(params).c_str());
+    }
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
+
+    std::vector<llama_token> embd_inp;
+    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+    embd_inp = inp_pfx;
+    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+    embd_inp.push_back(llama_token_middle(ctx));
+
+    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
+    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
+    // Should not run without any tokens
+    if (embd_inp.empty()) {
+        embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+    }
+
+    // Tokenize negative prompt
+    std::vector<llama_token> guidance_inp;
+    int guidance_offset = 0;
+    int original_prompt_len = 0;
+    if (ctx_guidance) {
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
+
+        original_prompt_len = original_inp.size();
+        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
+    }
+
+    if ((int) embd_inp.size() > n_ctx - 4) {
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        return 1;
+    }
+
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
+        params.n_keep = (int)embd_inp.size();
+    }
+
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
+
+
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
+        params.interactive = true;
+    }
+
+    if (params.verbose_prompt) {
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+
+        if (ctx_guidance) {
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            for (int i = 0; i < (int) guidance_inp.size(); i++) {
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
+            }
+        }
+
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+            for (int i = 0; i < params.n_keep; i++) {
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            }
+            LOG_TEE("'\n");
+        }
+        LOG_TEE("\n");
+    }
+
+    if (params.interactive) {
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
+        struct sigaction sigint_action;
+        sigint_action.sa_handler = sigint_handler;
+        sigemptyset (&sigint_action.sa_mask);
+        sigint_action.sa_flags = 0;
+        sigaction(SIGINT, &sigint_action, NULL);
+#elif defined (_WIN32)
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
+            return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
+        };
+        SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
+#endif
+
+        LOG_TEE("%s: interactive mode on.\n", __func__);
+
+        if (params.input_prefix_bos) {
+            LOG_TEE("Input prefix with BOS\n");
+        }
+
+        if (!params.input_prefix.empty()) {
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+        }
+
+        if (!params.input_suffix.empty()) {
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+        }
+    }
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+            params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
+
+    struct llama_grammar * grammar = NULL;
+    grammar_parser::parse_state parsed_grammar;
+
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        LOG_TEE("%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        LOG_TEE("\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos(ctx));
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+    LOG_TEE("\n#####  Infill mode  #####\n\n");
+    if (params.infill) {
+        printf("\n************\n");
+        printf("no need to specify '--infill', always running infill\n");
+        printf("************\n\n");
+    }
+    if (params.interactive) {
+        const char *control_message;
+        if (params.multiline_input) {
+            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n";
+        } else {
+            control_message = " - Press Return to return control to LLaMa.\n"
+                              " - To return control without starting a new line, end your input with '/'.\n"
+                              " - If you want to submit another line, end your input with '\\'.\n";
+        }
+        LOG_TEE("== Running in interactive mode. ==\n");
+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
+#endif
+        LOG_TEE(       "%s\n", control_message);
+
+        is_interacting = params.interactive_first;
+    }
+
+    bool input_echo           = true;
+
+    int n_past             = 0;
+    int n_remain           = params.n_predict;
+    int n_consumed         = 0;
+    int n_past_guidance    = 0;
+
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+
+    // the first thing we will do is to output the prompt, so set color accordingly
+    console::set_display(console::prompt);
+
+    std::vector<llama_token> embd;
+    std::vector<llama_token> embd_guidance;
+
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    while (n_remain != 0 || params.interactive) {
+        // predict
+        if (!embd.empty()) {
+            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // --prompt or --file which uses the same value.
+            int max_embd_size = n_ctx - 4;
+
+            // Ensure the input doesn't exceed the context size by truncating embd if necessary.
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
+                console::set_display(console::error);
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                console::set_display(console::reset);
+                fflush(stdout);
+            }
+
+            // infinite text generation via context swapping
+            // if we run out of context:
+            // - take the n_keep first tokens from the original prompt (via n_past)
+            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                if (params.n_predict == -2) {
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    break;
+                }
+
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
+
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+            }
+
+            // evaluate tokens in batches
+            // embd is typically prepared beforehand to fit within a batch, but not always
+
+            if (ctx_guidance) {
+                int input_size = 0;
+                llama_token * input_buf = NULL;
+
+                if (n_past_guidance < (int) guidance_inp.size()) {
+                    // Guidance context should have the same data with these modifications:
+                    //
+                    // * Replace the initial prompt
+                    // * Shift everything by guidance_offset
+                    embd_guidance = guidance_inp;
+                    if (embd.begin() + original_prompt_len < embd.end()) {
+                        embd_guidance.insert(
+                            embd_guidance.end(),
+                            embd.begin() + original_prompt_len,
+                            embd.end()
+                        );
+                    }
+
+                    input_buf  = embd_guidance.data();
+                    input_size = embd_guidance.size();
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
+                } else {
+                    input_buf  = embd.data();
+                    input_size = embd.size();
+                }
+
+                for (int i = 0; i < input_size; i += params.n_batch) {
+                    int n_eval = std::min(input_size - i, params.n_batch);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
+                        return 1;
+                    }
+
+                    n_past_guidance += n_eval;
+                }
+            }
+
+            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
+                int n_eval = (int) embd.size() - i;
+                if (n_eval > params.n_batch) {
+                    n_eval = params.n_batch;
+                }
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
+                    return 1;
+                }
+
+                n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
+            }
+
+        }
+
+        embd.clear();
+        embd_guidance.clear();
+
+        if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
+
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
+
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
+
+            embd.push_back(id);
+
+            // echo this to console
+            input_echo = true;
+
+            // decrement remaining sampling budget
+            --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
+        } else {
+            // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        if (input_echo) {
+            for (auto id : embd) {
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                printf("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
+            }
+            fflush(stdout);
+        }
+        // reset color to default if we there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
+            console::set_display(console::reset);
+        }
+
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
+
+            // deal with eot token in infill mode
+            if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
+                if(is_interacting && !params.interactive_first) {
+                    // print an eot token
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+                }
+                fflush(stdout);
+                printf("\n");
+                console::set_display(console::user_input);
+                std::string buffer;
+                std::string line;
+                bool another_line=true;
+                // set a new prefix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line, if so we use the old input
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_prefix = buffer;
+                }
+                buffer.clear();
+                // set a new suffix via stdin
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+                // check if we got an empty line
+                if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
+                    params.input_suffix = buffer;
+                }
+                buffer.clear();
+                // done taking input, reset color
+                console::set_display(console::reset);
+                // tokenize new prefix and suffix
+                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
+                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
+                embd_inp = inp_pfx;
+                embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                embd_inp.push_back(llama_token_middle(ctx));
+                embd.clear();
+                embd_guidance.clear();
+                n_remain = params.n_predict;
+                n_past = 0;
+                n_consumed = 0;
+                // LOG_TEE("took new input\n");
+                is_interacting = false;
+            }
+            // deal with end of text token in interactive mode
+            else if (last_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
+                if (params.interactive) {
+
+                    is_interacting = true;
+                    printf("\n");
+                    console::set_display(console::user_input);
+                    fflush(stdout);
+               }
+            }
+
+            if (n_past > 0 && is_interacting && !params.interactive) {
+                LOG("waiting for user input\n");
+
+                if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
+                    embd_inp.push_back(llama_token_bos(ctx));
+                }
+
+                std::string buffer;
+                if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    buffer += params.input_prefix;
+                    printf("%s", buffer.c_str());
+                }
+
+                std::string line;
+                bool another_line = true;
+                do {
+                    another_line = console::readline(line, params.multiline_input);
+                    buffer += line;
+                } while (another_line);
+
+                // done taking input, reset color
+                console::set_display(console::reset);
+
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
+                    // append input suffix if any
+                    if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        buffer += params.input_suffix;
+                        printf("%s", params.input_suffix.c_str());
+                    }
+
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << llama_token_to_piece(ctx, token);
+                    }
+
+                    n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
+                }
+
+                input_echo = false; // do not echo this again
+            }
+
+            if (n_past > 0) {
+                if (is_interacting) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
+                is_interacting = false;
+            }
+        }
+
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
+            break;
+        }
+
+        // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
+            n_remain = params.n_predict;
+            is_interacting = true;
+        }
+    }
+    if (!params.interactive && n_remain <= 0) {
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
+        fflush(stdout);
+    }
+
+    llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
+    if (ctx_guidance) { llama_free(ctx_guidance); }
+    llama_free(ctx);
+    llama_free_model(model);
+
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
+    llama_backend_free();
+
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
+    return 0;
+}
+
diff --git a/examples/jeopardy/README.md b/examples/jeopardy/README.md
index 4c42e3cdb..ffa13cbf3 100644
--- a/examples/jeopardy/README.md
+++ b/examples/jeopardy/README.md
@@ -2,7 +2,7 @@
 
 This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
 
-The jeopardy test can be used to compare the fact knowledge of different models and compare them to eachother. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
+The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
 
 
 Step 1: Open jeopardy.sh and modify the following:
diff --git a/examples/llama-bench/README.md b/examples/llama-bench/README.md
new file mode 100644
index 000000000..d02824bfa
--- /dev/null
+++ b/examples/llama-bench/README.md
@@ -0,0 +1,271 @@
+# llama.cpp/example/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>            (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                (default: 512)
+  -n, --n-gen <n>                   (default: 128)
+  -b, --batch-size <n>              (default: 512)
+  --memory-f32 <0|1>                (default: 0)
+  -t, --threads <n>                 (default: 16)
+  -ngl N, --n-gpu-layers <n>        (default: 99)
+  -mg i, --main-gpu <i>             (default: 0)
+  -mmq, --mul-mat-q <0|1>           (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>
+  -r, --repetitions <n>             (default: 5)
+  -o, --output <csv|json|md|sql>    (default: md)
+  -v, --verbose                     (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform two types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  opencl INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
old mode 100755
new mode 100644
index 36057bfca..a04115c96
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -3,6 +3,9 @@
 #include <cassert>
 #include <chrono>
 #include <cinttypes>
+#include <clocale>
+#include <cmath>
+#include <cstdio>
 #include <cstring>
 #include <ctime>
 #include <iterator>
@@ -10,7 +13,6 @@
 #include <numeric>
 #include <regex>
 #include <sstream>
-#include <stdio.h>
 #include <string>
 #include <vector>
 
@@ -18,9 +20,7 @@
 #include "llama.h"
 #include "common.h"
 #include "build-info.h"
-#ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#endif
 
 // utils
 static uint64_t get_time_ns() {
@@ -74,14 +74,6 @@ static T stdev(const std::vector<T> & v) {
     return stdev;
 }
 
-static bool ggml_cpu_has_metal() {
-#if defined(GGML_USE_METAL)
-    return true;
-#else
-    return false;
-#endif
-}
-
 static std::string get_cpu_info() {
     std::string id;
 #ifdef __linux__
@@ -140,7 +132,6 @@ struct cmd_params {
     std::vector<int> n_gpu_layers;
     std::vector<int> main_gpu;
     std::vector<bool> mul_mat_q;
-    std::vector<bool> low_vram;
     std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
     bool verbose;
@@ -157,7 +148,6 @@ static const cmd_params cmd_params_defaults = {
     /* n_gpu_layers  */ {99},
     /* main_gpu      */ {0},
     /* mul_mat_q     */ {true},
-    /* low_vram      */ {false},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
     /* verbose       */ false,
@@ -165,26 +155,25 @@ static const cmd_params cmd_params_defaults = {
 };
 
 static void print_usage(int /* argc */, char ** argv) {
-    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help\n");
-    fprintf(stdout, "  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    fprintf(stdout, "  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    fprintf(stdout, "  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    fprintf(stdout, "  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    fprintf(stdout, "  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
-    fprintf(stdout, "  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    fprintf(stdout, "  -ngl N, --n-gpu-layers <n>        (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    fprintf(stdout, "  -mg i, --main-gpu <n>             (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    fprintf(stdout, "  -lv, --low-vram <0|1>             (default: %s)\n", join(cmd_params_defaults.low_vram, ",").c_str());
-    fprintf(stdout, "  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    fprintf(stdout, "  -ts, --tensor_split <ts0/ts1/..>               \n");
-    fprintf(stdout, "  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
-    fprintf(stdout, "  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
-    fprintf(stdout, "  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
-    fprintf(stdout, "\n");
-    fprintf(stdout, "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+    printf("usage: %s [options]\n", argv[0]);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help\n");
+    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
+    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
+    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("\n");
+    printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 
 }
 
@@ -263,13 +252,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
-        } else if (arg == "-lv" || arg == "--low-vram") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.low_vram.insert(params.low_vram.end(), p.begin(), p.end());
         } else if (arg == "-mmq" || arg == "--mul-mat-q") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -344,7 +326,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
-    if (params.low_vram.empty())     { params.low_vram = cmd_params_defaults.low_vram; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
@@ -361,21 +342,34 @@ struct cmd_params_instance {
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
-    llama_context_params to_llama_params() const {
-        llama_context_params lparams = llama_context_default_params();
-        lparams.n_ctx = n_prompt + n_gen;
-        lparams.n_batch = n_batch;
-        lparams.f16_kv = !f32_kv;
-        lparams.n_gpu_layers = n_gpu_layers;
-        lparams.main_gpu = main_gpu;
-        lparams.mul_mat_q = mul_mat_q;
-        lparams.low_vram = low_vram;
-        lparams.tensor_split = tensor_split.data();
+    llama_model_params to_llama_mparams() const {
+        llama_model_params mparams = llama_model_default_params();
 
-        return lparams;
+        mparams.n_gpu_layers = n_gpu_layers;
+        mparams.main_gpu = main_gpu;
+        mparams.tensor_split = tensor_split.data();
+
+        return mparams;
+    }
+
+    bool equal_mparams(const cmd_params_instance & other) const {
+        return model == other.model &&
+               n_gpu_layers == other.n_gpu_layers &&
+               main_gpu == other.main_gpu &&
+               tensor_split == other.tensor_split;
+    }
+
+    llama_context_params to_llama_cparams() const {
+        llama_context_params cparams = llama_context_default_params();
+
+        cparams.n_ctx = n_prompt + n_gen;
+        cparams.n_batch = n_batch;
+        cparams.f16_kv = !f32_kv;
+        cparams.mul_mat_q = mul_mat_q;
+
+        return cparams;
     }
 };
 
@@ -383,13 +377,12 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
     std::vector<cmd_params_instance> instances;
 
     for (const auto & m : params.model)
-    for (const auto & nb : params.n_batch)
-    for (const auto & fk : params.f32_kv)
     for (const auto & nl : params.n_gpu_layers)
     for (const auto & mg : params.main_gpu)
-    for (const auto & mmq : params.mul_mat_q)
-    for (const auto & lv : params.low_vram)
     for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nt : params.n_threads) {
         cmd_params_instance instance = {
             /* .model        = */ m,
@@ -401,7 +394,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
             /* .n_gpu_layers = */ nl,
             /* .main_gpu     = */ mg,
             /* .mul_mat_q    = */ mmq,
-            /* .low_vram     = */ lv,
             /* .tensor_split = */ ts,
         };
         instances.push_back(instance);
@@ -412,6 +404,56 @@ static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_p
 static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
     std::vector<cmd_params_instance> instances;
 
+#if 1
+    // this ordering minimizes the number of times that each model needs to be reloaded
+    for (const auto & m : params.model)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nt : params.n_threads) {
+        for (const auto & n_prompt : params.n_prompt) {
+            if (n_prompt == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_prompt,
+                /* .n_gen        = */ 0,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+
+        for (const auto & n_gen : params.n_gen) {
+            if (n_gen == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ 0,
+                /* .n_gen        = */ n_gen,
+                /* .n_batch      = */ nb,
+                /* .f32_kv       = */ fk,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .main_gpu     = */ mg,
+                /* .mul_mat_q    = */ mmq,
+                /* .tensor_split = */ ts,
+            };
+            instances.push_back(instance);
+        }
+    }
+#else
+    // this ordering separates the prompt and generation tests
     for (const auto & n_prompt : params.n_prompt) {
         if (n_prompt == 0) {
             continue;
@@ -427,6 +469,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
         auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
         instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
     }
+#endif
 
     return instances;
 }
@@ -443,13 +486,14 @@ struct test {
     static const std::string gpu_info;
     std::string model_filename;
     std::string model_type;
+    uint64_t model_size;
+    uint64_t model_n_params;
     int n_batch;
     int n_threads;
     bool f32_kv;
     int n_gpu_layers;
     int main_gpu;
     bool mul_mat_q;
-    bool low_vram;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
     int n_gen;
@@ -459,15 +503,16 @@ struct test {
     test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
         model_filename = inst.model;
         char buf[128];
-        llama_model_type(lmodel, buf, sizeof(buf));
+        llama_model_desc(lmodel, buf, sizeof(buf));
         model_type = buf;
+        model_size = llama_model_size(lmodel);
+        model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
         f32_kv = inst.f32_kv;
         n_gpu_layers = inst.n_gpu_layers;
         main_gpu = inst.main_gpu;
         mul_mat_q = inst.mul_mat_q;
-        low_vram = inst.low_vram;
         tensor_split = inst.tensor_split;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
@@ -504,7 +549,7 @@ struct test {
 
     static std::string get_backend() {
         if (cuda) {
-            return "CUDA";
+            return GGML_CUDA_NAME;
         }
         if (opencl) {
             return "OpenCL";
@@ -526,9 +571,9 @@ struct test {
             "build_commit", "build_number",
             "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
-            "model_filename", "model_type",
+            "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "f16_kv",
-            "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split",
+            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -540,13 +585,14 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+            field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
         if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "mul_mat_q" || field == "low_vram") {
+            field == "f16_kv" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -575,9 +621,9 @@ struct test {
             build_commit, std::to_string(build_number),
             std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
-            model_filename, model_type,
+            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
-            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str,
+            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -609,9 +655,9 @@ struct printer {
     virtual ~printer() {}
 
     FILE * fout;
-    virtual void print_header(const cmd_params & params) { (void) params; };
+    virtual void print_header(const cmd_params & params) { (void) params; }
     virtual void print_test(const test & t) = 0;
-    virtual void print_footer() { };
+    virtual void print_footer() { }
 };
 
 struct csv_printer : public printer {
@@ -711,8 +757,15 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 15;
+            return 16;
         }
+        if (field == "size" || field == "params") {
+            return 10;
+        }
+        if (field == "n_gpu_layers") {
+            return 3;
+        }
+
         int width = std::max((int)field.length(), 10);
 
         if (test::get_field_type(field) == test::STRING) {
@@ -721,9 +774,28 @@ struct markdown_printer : public printer {
         return width;
     }
 
+    static std::string get_field_display_name(const std::string & field) {
+        if (field == "n_gpu_layers") {
+            return "ngl";
+        }
+        if (field == "n_threads") {
+            return "threads";
+        }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
+        if (field == "tensor_split") {
+            return "ts";
+        }
+        return field;
+    }
+
     void print_header(const cmd_params & params) override {
         // select fields to print
-        fields = { "model", "backend" };
+        fields.push_back("model");
+        fields.push_back("size");
+        fields.push_back("params");
+        fields.push_back("backend");
         bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
         if (!is_cpu_backend) {
             fields.push_back("n_gpu_layers");
@@ -743,9 +815,6 @@ struct markdown_printer : public printer {
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
-        if (params.low_vram.size() > 1 || params.low_vram != cmd_params_defaults.low_vram) {
-            fields.push_back("low_vram");
-        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }
@@ -754,7 +823,7 @@ struct markdown_printer : public printer {
 
         fprintf(fout, "|");
         for (const auto & field : fields) {
-            fprintf(fout, " %*s |", get_field_width(field), field.c_str());
+            fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str());
         }
         fprintf(fout, "\n");
         fprintf(fout, "|");
@@ -771,12 +840,26 @@ struct markdown_printer : public printer {
         fprintf(fout, "|");
         for (const auto & field : fields) {
             std::string value;
+            char buf[128];
             if (field == "model") {
                 value = t.model_type;
+            } else if (field == "size") {
+                if (t.model_size < 1024*1024*1024) {
+                    snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
+                }
+                value = buf;
+            } else if (field == "params") {
+                if (t.model_n_params < 1000*1000*1000) {
+                    snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
+                }
+                value = buf;
             } else if (field == "backend") {
                 value = test::get_backend();
             } else if (field == "test") {
-                char buf[128];
                 if (t.n_prompt > 0 && t.n_gen == 0) {
                     snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
                 } else if (t.n_gen > 0 && t.n_prompt == 0) {
@@ -787,7 +870,6 @@ struct markdown_printer : public printer {
                 }
                 value = buf;
             } else if (field == "t/s") {
-                char buf[128];
                 snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
                 value = buf;
             } else if (vmap.find(field) != vmap.end()) {
@@ -853,27 +935,36 @@ struct sql_printer : public printer {
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
     std::vector<llama_token> tokens(n_batch, llama_token_bos(ctx));
     int n_processed = 0;
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        llama_eval(ctx, tokens.data(), n_tokens, n_past + n_processed, n_threads);
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_token token = llama_token_bos(ctx);
+
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     for (int i = 0; i < n_gen; i++) {
-        llama_eval(ctx, &token, 1, n_past + i, n_threads);
+        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
     }
 }
 
-static void llama_null_log_callback(enum llama_log_level level, const char * text, void * user_data) {
+static void llama_null_log_callback(enum ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) text;
     (void) user_data;
 }
 
 int main(int argc, char ** argv) {
+    // try to set locale for unicode characters in markdown
+    setlocale(LC_CTYPE, ".UTF-8");
+
 #if !defined(NDEBUG)
     fprintf(stderr, "warning: asserts enabled, performance may be affected\n");
 #endif
@@ -919,17 +1010,25 @@ int main(int argc, char ** argv) {
 
     std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
 
-    for (const auto & inst : params_instances) {
-        // TODO: keep the model between tests when possible
-        llama_context_params lparams = inst.to_llama_params();
+    llama_model * lmodel = nullptr;
+    const cmd_params_instance * prev_inst = nullptr;
 
-        llama_model * lmodel  = llama_load_model_from_file(inst.model.c_str(), lparams);
-        if (lmodel == NULL) {
-            fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
-            return 1;
+    for (const auto & inst : params_instances) {
+        // keep the same model between tests when possible
+        if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
+            if (lmodel) {
+                llama_free_model(lmodel);
+            }
+
+            lmodel = llama_load_model_from_file(inst.model.c_str(), inst.to_llama_mparams());
+            if (lmodel == NULL) {
+                fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, inst.model.c_str());
+                return 1;
+            }
+            prev_inst = &inst;
         }
 
-        llama_context * ctx = llama_new_context_with_model(lmodel, lparams);
+        llama_context * ctx = llama_new_context_with_model(lmodel, inst.to_llama_cparams());
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, inst.model.c_str());
             llama_free_model(lmodel);
@@ -938,10 +1037,19 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         // warmup run
-        test_gen(ctx, 1, 0, t.n_threads);
+        if (t.n_prompt > 0) {
+            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
+        }
+        if (t.n_gen > 0) {
+            test_gen(ctx, 1, 0, t.n_threads);
+        }
 
         for (int i = 0; i < params.reps; i++) {
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
+
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
@@ -958,9 +1066,10 @@ int main(int argc, char ** argv) {
         llama_print_timings(ctx);
 
         llama_free(ctx);
-        llama_free_model(lmodel);
     }
 
+    llama_free_model(lmodel);
+
     p->print_footer();
 
     llama_backend_free();
diff --git a/examples/llm.vim b/examples/llm.vim
index 594a28549..d580a3d00 100644
--- a/examples/llm.vim
+++ b/examples/llm.vim
@@ -8,7 +8,7 @@ function! Llm()
   let buffer_content = join(getline(1, '$'), "\n")
 
   " Create the JSON payload
-  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false}
+  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false}
   let json_payload.prompt = buffer_content
 
   " Define the curl command
@@ -25,3 +25,4 @@ function! Llm()
 endfunction
 
 command! Llm call Llm()
+noremap <F2> :Llm<CR>
diff --git a/examples/main-cmake-pkg/.gitignore b/examples/main-cmake-pkg/.gitignore
new file mode 100644
index 000000000..e32c11c7f
--- /dev/null
+++ b/examples/main-cmake-pkg/.gitignore
@@ -0,0 +1,51 @@
+# Prerequisites
+*.d
+
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+*.smod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+*.gguf
+
+*.log
+.DS_Store
+.build/
+.cache/
+.direnv/
+.envrc
+.swiftpm
+.venv
+.clang-tidy
+.vs/
+.vscode/
+
+build*/
+out/
+tmp/
+
diff --git a/examples/main-cmake-pkg/CMakeLists.txt b/examples/main-cmake-pkg/CMakeLists.txt
new file mode 100644
index 000000000..908131884
--- /dev/null
+++ b/examples/main-cmake-pkg/CMakeLists.txt
@@ -0,0 +1,46 @@
+cmake_minimum_required(VERSION 3.12)
+project("main-cmake-pkg" C CXX)
+set(TARGET main-cmake-pkg)
+
+find_package(Llama 0.0.1 REQUIRED)
+
+# Bake common functionality in with target. Because applications
+# using the relocatable Llama package should be outside of the
+# source tree, main-cmake-pkg pretends the dependencies are built-in.
+
+set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
+add_library(common OBJECT
+    ${_common_path}/common.h
+    ${_common_path}/common.cpp
+    ${_common_path}/console.h
+    ${_common_path}/console.cpp
+    ${_common_path}/grammar-parser.h
+    ${_common_path}/grammar-parser.cpp
+    )
+
+# WARNING: because build-info.h is auto-generated, it will only
+# be available after the user has built the llama.cpp sources.
+#
+configure_file(${_common_path}/../build-info.h
+    ${CMAKE_CURRENT_BINARY_DIR}/build-info.h
+    COPYONLY)
+
+target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR}
+    ${CMAKE_CURRENT_BINARY_DIR})
+
+# If the common project was part of "main-cmake-pkg" the transient
+# defines would automatically be attached. Because the common func-
+# tionality is separate, but dependent upon the defines, it must be
+# explicitly extracted from the "llama" target.
+#
+get_target_property(_llama_transient_defines llama
+    INTERFACE_COMPILE_DEFINITIONS)
+
+target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
+
+add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
+target_include_directories(${TARGET} PRIVATE ${_common_path})
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
diff --git a/examples/main-cmake-pkg/README.md b/examples/main-cmake-pkg/README.md
new file mode 100644
index 000000000..6d665f28f
--- /dev/null
+++ b/examples/main-cmake-pkg/README.md
@@ -0,0 +1,37 @@
+# llama.cpp/example/main-cmake-pkg
+
+This program builds the [main](../main) application using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+
+## Building
+
+Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
+
+### Considerations
+
+When hardware acceleration libraries are used (e.g. CUBlas, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
+
+### Build llama.cpp and install to C:\LlamaCPP directory
+
+In this case, CLBlast was already installed so the CMake package is referenced in `CMAKE_PREFIX_PATH`.
+
+```cmd
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/LlamaCPP
+```
+
+### Build main-cmake-pkg
+
+
+```cmd
+cd ..\examples\main-cmake-pkg
+mkdir build
+cd build
+cmake .. -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
+cmake --build . --config Release
+cmake --install . --prefix C:/MyLlamaApp
+```
diff --git a/examples/main/README.md b/examples/main/README.md
index d555afdcc..a9561c383 100644
--- a/examples/main/README.md
+++ b/examples/main/README.md
@@ -34,7 +34,7 @@ For an interactive experience, try this command:
 #### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " \
+./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \
 'User: Hi
 AI: Hello. I am an AI chatbot. Would you like to talk?
 User: Sure!
@@ -45,7 +45,7 @@ User:'
 #### Windows:
 
 ```powershell
-main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -e --prompt "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
+main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
 ```
 
 The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -144,7 +144,7 @@ The `--ctx-size` option allows you to set the size of the prompt context used by
 
 Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8.
 
-- `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
+-   `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model.
 
 ### Keep Prompt
 
@@ -262,7 +262,8 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Number of Threads
 
--   `-t N, --threads N`: Set the number of threads to use during computation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has (as opposed to the logical number of cores). Using the correct number of threads can greatly improve performance.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. In some systems, it is beneficial to use a higher number of threads during batch processing than during generation. If not specified, the number of threads used for batch processing will be the same as the number of threads used for generation.
 
 ### Mlock
 
@@ -274,7 +275,7 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop\_caches' as root.
+-   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 
@@ -302,10 +303,8 @@ These options provide extra functionality and customization when running the LLa
 
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
--   `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 0a22f3c25..775a5a201 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -1,9 +1,5 @@
-// Defines sigaction on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
 #include "common.h"
+
 #include "console.h"
 #include "llama.h"
 #include "build-info.h"
@@ -17,6 +13,7 @@
 #include <ctime>
 #include <fstream>
 #include <iostream>
+#include <sstream>
 #include <string>
 #include <vector>
 
@@ -36,18 +33,69 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static llama_context ** g_ctx;
+static llama_context           ** g_ctx;
+static llama_model             ** g_model;
+static gpt_params               * g_params;
+static std::vector<llama_token> * g_input_tokens;
+static std::ostringstream       * g_output_ss;
+static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = get_sortable_timestamp();
+
+    const bool success = create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: main\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-void sigint_handler(int signo) {
+static void sigint_handler(int signo) {
     if (signo == SIGINT) {
         if (!is_interacting) {
-            is_interacting=true;
+            is_interacting = true;
         } else {
             console::cleanup();
             printf("\n");
             llama_print_timings(*g_ctx);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
             _exit(130);
         }
     }
@@ -56,17 +104,27 @@ void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
+    g_params = &params;
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("main", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // TODO: Dump params ?
+    //LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
+
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
 
-    if (params.perplexity) {
+    if (params.logits_all) {
         printf("\n************\n");
         printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
         printf("************\n\n");
@@ -82,43 +140,44 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
-    if (params.rope_freq_base != 10000.0) {
-        fprintf(stderr, "%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
-    }
-
-    if (params.rope_freq_scale != 1.0) {
-        fprintf(stderr, "%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
-    }
-
-    if (params.n_ctx > 2048) {
-        // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
-        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
-    } else if (params.n_ctx < 8) {
-        fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+    if (params.n_ctx != 0 && params.n_ctx < 8) {
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    if (params.rope_freq_base != 0.0) {
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 0.0) {
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+    }
+
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    LOG("%s: llama backend init\n", __func__);
     llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
     llama_context * ctx_guidance = NULL;
+    g_model = &model;
     g_ctx = &ctx;
 
     // load the model and apply lora adapter, if any
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (params.cfg_scale > 1.f) {
         struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
@@ -126,80 +185,71 @@ int main(int argc, char ** argv) {
     }
 
     if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        LOG_TEE("%s: error: unable to load model\n", __func__);
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+    LOG("n_ctx: %d\n", n_ctx);
+
+    if (n_ctx > n_ctx_train) {
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
+    }
+
     // print system information
     {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
-    }
-
-    // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters
-    // uncomment the "used_mem" line in llama.cpp to see the results
-    if (params.mem_test) {
-        {
-            fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
-
-            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos(ctx));
-            llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
-        }
-
-        llama_print_timings(ctx);
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
-    }
-
-    // export the cgraph and exit
-    if (params.export_cgraph) {
-        llama_eval_export(ctx, "llama.ggml");
-        llama_free(ctx);
-        llama_free_model(model);
-
-        return 0;
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", get_system_info(params).c_str());
     }
 
     std::string path_session = params.path_prompt_cache;
     std::vector<llama_token> session_tokens;
 
     if (!path_session.empty()) {
-        fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
 
         // fopen to check for existing session
         FILE * fp = std::fopen(path_session.c_str(), "rb");
         if (fp != NULL) {
             std::fclose(fp);
 
-            session_tokens.resize(params.n_ctx);
+            session_tokens.resize(n_ctx);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
-                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
+                LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
             llama_set_rng_seed(ctx, params.seed);
 
-            fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
+            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
-            fprintf(stderr, "%s: session file does not exist, will create\n", __func__);
+            LOG_TEE("%s: session file does not exist, will create\n", __func__);
         }
     }
 
-    // tokenize the prompt
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    LOG("add_bos: %d\n", add_bos);
+
     std::vector<llama_token> embd_inp;
+
     if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
-        embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+        LOG("tokenize the prompt\n");
+        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
     } else {
+        LOG("use session tokens\n");
         embd_inp = session_tokens;
     }
 
+    LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
+
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(ctx));
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
     }
 
     // Tokenize negative prompt
@@ -207,24 +257,28 @@ int main(int argc, char ** argv) {
     int guidance_offset = 0;
     int original_prompt_len = 0;
     if (ctx_guidance) {
-        params.cfg_negative_prompt.insert(0, 1, ' ');
-        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, true);
+        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt));
+
+        guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos);
+        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
+
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
         original_prompt_len = original_inp.size();
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
+        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
+        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
     }
 
-    const int n_ctx = llama_n_ctx(ctx);
-
     if ((int) embd_inp.size() > n_ctx - 4) {
-        fprintf(stderr, "%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
     }
 
     // debug message about similarity of saved session, if applicable
     size_t n_matching_session_tokens = 0;
-    if (session_tokens.size()) {
+    if (!session_tokens.empty()) {
         for (llama_token id : session_tokens) {
             if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
                 break;
@@ -232,22 +286,27 @@ int main(int argc, char ** argv) {
             n_matching_session_tokens++;
         }
         if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
-            fprintf(stderr, "%s: using full prompt from session file\n", __func__);
+            LOG_TEE("%s: using full prompt from session file\n", __func__);
         } else if (n_matching_session_tokens >= embd_inp.size()) {
-            fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
+            LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
         } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-            fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+            LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         } else {
-            fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n",
+            LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
                 __func__, n_matching_session_tokens, embd_inp.size());
         }
     }
 
+    LOGLN(
+            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
+            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
+
     // if we will use the cache for the full prompt without reaching the end of the cache, force
     // reevaluation of the last token token to recalculate the cached logits
-    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() &&
-            session_tokens.size() > embd_inp.size()) {
+    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
+        LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
+
         session_tokens.resize(embd_inp.size() - 1);
     }
 
@@ -257,8 +316,11 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false);
+
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
@@ -272,30 +334,30 @@ int main(int argc, char ** argv) {
     }
 
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str());
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (ctx_guidance) {
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
-            fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
+            LOG_TEE("\n");
+            LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str());
+            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
             for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_str(ctx, guidance_inp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
             }
         }
 
         if (params.n_keep > 0) {
-        fprintf(stderr, "%s: static prompt based on n_keep: '", __func__);
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                fprintf(stderr, "%s", llama_token_to_str(ctx, embd_inp[i]).c_str());
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            fprintf(stderr, "'\n");
+            LOG_TEE("'\n");
         }
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
     }
 
     if (params.interactive) {
@@ -312,47 +374,48 @@ int main(int argc, char ** argv) {
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-        fprintf(stderr, "%s: interactive mode on.\n", __func__);
+        LOG_TEE("%s: interactive mode on.\n", __func__);
 
-        if (params.antiprompt.size()) {
-            for (auto antiprompt : params.antiprompt) {
-                fprintf(stderr, "Reverse prompt: '%s'\n", antiprompt.c_str());
+        if (!params.antiprompt.empty()) {
+            for (const auto & antiprompt : params.antiprompt) {
+                LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
             }
         }
 
         if (params.input_prefix_bos) {
-            fprintf(stderr, "Input prefix with BOS\n");
+            LOG_TEE("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
         }
 
         if (!params.input_suffix.empty()) {
-            fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
+    LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
             params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau);
-    fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-    fprintf(stderr, "\n\n");
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
 
+    struct llama_grammar * grammar = NULL;
     grammar_parser::parse_state parsed_grammar;
-    llama_grammar *             grammar = NULL;
+
     if (!params.grammar.empty()) {
         parsed_grammar = grammar_parser::parse(params.grammar.c_str());
         // will be empty (default) if there are parse errors
         if (parsed_grammar.rules.empty()) {
             return 1;
         }
-        fprintf(stderr, "%s: grammar:\n", __func__);
+        LOG_TEE("%s: grammar:\n", __func__);
         grammar_parser::print_grammar(stderr, parsed_grammar);
-        fprintf(stderr, "\n");
+        LOG_TEE("\n");
 
         {
             auto it = params.logit_bias.find(llama_token_eos(ctx));
             if (it != params.logit_bias.end() && it->second == -INFINITY) {
-                fprintf(stderr, "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+                LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
             }
         }
 
@@ -362,8 +425,8 @@ int main(int argc, char ** argv) {
     }
 
     // TODO: replace with ring-buffer
-    std::vector<llama_token> last_n_tokens(n_ctx);
-    std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
 
     if (params.interactive) {
         const char *control_message;
@@ -375,11 +438,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        fprintf(stderr, "== Running in interactive mode. ==\n"
+        LOG_TEE("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-               " - Press Ctrl+C to interject at any time.\n"
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-               "%s\n", control_message);
+        LOG_TEE(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -394,33 +457,37 @@ int main(int argc, char ** argv) {
     int n_session_consumed = 0;
     int n_past_guidance    = 0;
 
+    std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
+    std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
+    std::ostringstream output_ss;     g_output_ss     = &output_ss;
+
     // the first thing we will do is to output the prompt, so set color accordingly
     console::set_display(console::prompt);
 
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    // do one empty run to warm up the model
-    {
-        const std::vector<llama_token> tmp = { llama_token_bos(ctx), };
-        llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
-        llama_reset_timings(ctx);
-    }
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
-        if (embd.size() > 0) {
+        if (!embd.empty()) {
             // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
-            auto max_embd_size = n_ctx - 4;
+            int max_embd_size = n_ctx - 4;
+
             // Ensure the input doesn't exceed the context size by truncating embd if necessary.
-            if ((int)embd.size() > max_embd_size) {
-                auto skipped_tokens = embd.size() - max_embd_size;
+            if ((int) embd.size() > max_embd_size) {
+                const int skipped_tokens = (int) embd.size() - max_embd_size;
+                embd.resize(max_embd_size);
+
                 console::set_display(console::error);
-                printf("<<input too long: skipped %zu token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console::set_display(console::reset);
                 fflush(stdout);
-                embd.resize(max_embd_size);
             }
 
             // infinite text generation via context swapping
@@ -429,28 +496,31 @@ int main(int argc, char ** argv) {
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
                 if (params.n_predict == -2) {
-                    fprintf(stderr, "\n\n%s: context full, stopping generation\n", __func__);
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
                 }
 
-                const int n_left = n_past - params.n_keep;
-                // always keep the first token - BOS
-                n_past = std::max(1, params.n_keep);
-                n_past_guidance = std::max(1, params.n_keep + guidance_offset);
+                const int n_left    = n_past - params.n_keep - 1;
+                const int n_discard = n_left/2;
 
-                // insert n_left/2 tokens at the start of embd from last_n_tokens
-                embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                    n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                // stop saving session if we run out of context
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                n_past -= n_discard;
+
+                if (ctx_guidance) {
+                    n_past_guidance -= n_discard;
+                }
+
+                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                LOG("clear session path\n");
                 path_session.clear();
-
-                //printf("\n---\n");
-                //printf("resetting: '");
-                //for (int i = 0; i < (int) embd.size(); i++) {
-                //    printf("%s", llama_token_to_str(ctx, embd[i]));
-                //}
-                //printf("'\n");
-                //printf("\n---\n");
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -473,6 +543,9 @@ int main(int argc, char ** argv) {
                 if (i > 0) {
                     embd.erase(embd.begin(), embd.begin() + i);
                 }
+
+                // remove any "future" tokens that we might have inherited from the session from the KV cache
+                llama_kv_cache_tokens_rm(ctx, n_past, -1);
             }
 
             // evaluate tokens in batches
@@ -480,7 +553,7 @@ int main(int argc, char ** argv) {
 
             if (ctx_guidance) {
                 int input_size = 0;
-                llama_token* input_buf = NULL;
+                llama_token * input_buf = NULL;
 
                 if (n_past_guidance < (int) guidance_inp.size()) {
                     // Guidance context should have the same data with these modifications:
@@ -496,22 +569,19 @@ int main(int argc, char ** argv) {
                         );
                     }
 
-                    input_buf = embd_guidance.data();
+                    input_buf  = embd_guidance.data();
                     input_size = embd_guidance.size();
-                    //fprintf(stderr, "\n---------------------\n");
-                    //for (int i = 0; i < (int) embd_guidance.size(); i++) {
-                        //fprintf(stderr, "%s", llama_token_to_str(ctx, embd_guidance[i]));
-                    //}
-                    //fprintf(stderr, "\n---------------------\n");
+
+                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
                 } else {
-                    input_buf = embd.data();
+                    input_buf  = embd.data();
                     input_size = embd.size();
                 }
 
                 for (int i = 0; i < input_size; i += params.n_batch) {
                     int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) {
-                        fprintf(stderr, "%s : failed to eval\n", __func__);
+                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
+                        LOG_TEE("%s : failed to eval\n", __func__);
                         return 1;
                     }
 
@@ -524,14 +594,20 @@ int main(int argc, char ** argv) {
                 if (n_eval > params.n_batch) {
                     n_eval = params.n_batch;
                 }
-                if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
-                    fprintf(stderr, "%s : failed to eval\n", __func__);
+
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
+
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                    LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
+
                 n_past += n_eval;
+
+                LOG("n_past = %d\n", n_past);
             }
 
-            if (embd.size() > 0 && !path_session.empty()) {
+            if (!embd.empty() && !path_session.empty()) {
                 session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
                 n_session_consumed = session_tokens.size();
             }
@@ -541,101 +617,21 @@ int main(int argc, char ** argv) {
         embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            // out of user input, sample next token
-            const float   temp            = params.temp;
-            const int32_t top_k           = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
-            const float   top_p           = params.top_p;
-            const float   tfs_z           = params.tfs_z;
-            const float   typical_p       = params.typical_p;
-            const int32_t repeat_last_n   = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
-            const float   repeat_penalty  = params.repeat_penalty;
-            const float   alpha_presence  = params.presence_penalty;
-            const float   alpha_frequency = params.frequency_penalty;
-            const int     mirostat        = params.mirostat;
-            const float   mirostat_tau    = params.mirostat_tau;
-            const float   mirostat_eta    = params.mirostat_eta;
-            const bool    penalize_nl     = params.penalize_nl;
-
             // optionally save the session on first sample (for faster prompt loading next time)
             if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
                 llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+
+                LOG("saved session to %s\n", path_session.c_str());
             }
 
-            llama_token id = 0;
+            const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
 
-            {
-                auto logits  = llama_get_logits(ctx);
-                auto n_vocab = llama_n_vocab(ctx);
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
 
-                // Apply params.logit_bias map
-                for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-                    logits[it->first] += it->second;
-                }
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
 
-                std::vector<llama_token_data> candidates;
-                candidates.reserve(n_vocab);
-                for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                    candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-                }
-
-                llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-                if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
-                }
-
-                // Apply penalties
-                float nl_logit = logits[llama_token_nl(ctx)];
-                auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
-                llama_sample_repetition_penalty(ctx, &candidates_p,
-                    last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                    last_n_repeat, repeat_penalty);
-                llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-                    last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                    last_n_repeat, alpha_frequency, alpha_presence);
-                if (!penalize_nl) {
-                    logits[llama_token_nl(ctx)] = nl_logit;
-                }
-
-                if (grammar != NULL) {
-                    llama_sample_grammar(ctx, &candidates_p, grammar);
-                }
-
-                if (temp <= 0) {
-                    // Greedy sampling
-                    id = llama_sample_token_greedy(ctx, &candidates_p);
-                } else {
-                    if (mirostat == 1) {
-                        static float mirostat_mu = 2.0f * mirostat_tau;
-                        const int mirostat_m = 100;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-                    } else if (mirostat == 2) {
-                        static float mirostat_mu = 2.0f * mirostat_tau;
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                    } else {
-                        // Temperature sampling
-                        llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-                        llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
-                        llama_sample_typical(ctx, &candidates_p, typical_p, 1);
-                        llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                        llama_sample_temperature(ctx, &candidates_p, temp);
-                        id = llama_sample_token(ctx, &candidates_p);
-                    }
-                }
-                // printf("`%d`", candidates_p.size);
-
-                if (grammar != NULL) {
-                    llama_grammar_accept_token(ctx, grammar, id);
-                }
-
-                last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(id);
-            }
-
-            // add it to the context
             embd.push_back(id);
 
             // echo this to console
@@ -643,12 +639,15 @@ int main(int argc, char ** argv) {
 
             // decrement remaining sampling budget
             --n_remain;
+
+            LOG("n_remain: %d\n", n_remain);
         } else {
             // some user input remains from prompt or interaction, forward it to processing
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
-                last_n_tokens.erase(last_n_tokens.begin());
-                last_n_tokens.push_back(embd_inp[n_consumed]);
+                last_tokens.erase(last_tokens.begin());
+                last_tokens.push_back(embd_inp[n_consumed]);
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
                     break;
@@ -659,23 +658,30 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo) {
             for (auto id : embd) {
-                printf("%s", llama_token_to_str(ctx, id).c_str());
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                printf("%s", token_str.c_str());
+
+                if (embd.size() > 1) {
+                    input_tokens.push_back(id);
+                } else {
+                    output_tokens.push_back(id);
+                    output_ss << token_str;
+                }
             }
             fflush(stdout);
         }
-        // reset color to default if we there is no pending user input
-        if (input_echo && (int)embd_inp.size() == n_consumed) {
+        // reset color to default if there is no pending user input
+        if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
         }
 
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
-
             // check for reverse prompt
-            if (params.antiprompt.size()) {
+            if (!params.antiprompt.empty()) {
                 std::string last_output;
-                for (auto id : last_n_tokens) {
-                    last_output += llama_token_to_str(ctx, id);
+                for (auto id : last_tokens) {
+                    last_output += llama_token_to_piece(ctx, id);
                 }
 
                 is_antiprompt = false;
@@ -688,22 +694,26 @@ int main(int argc, char ** argv) {
                         ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
                         : 0;
 
-                    if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
+                    if (last_output.find(antiprompt, search_start_pos) != std::string::npos) {
                         if (params.interactive) {
                             is_interacting = true;
-                            console::set_display(console::user_input);
                         }
                         is_antiprompt = true;
-                        fflush(stdout);
                         break;
                     }
                 }
+
+                if (is_antiprompt) {
+                    LOG("found antiprompt: %s\n", last_output.c_str());
+                }
             }
 
             // deal with end of text token in interactive mode
-            if (last_n_tokens.back() == llama_token_eos(ctx)) {
+            if (last_tokens.back() == llama_token_eos(ctx)) {
+                LOG("found EOS token\n");
+
                 if (params.interactive) {
-                    if (params.antiprompt.size() != 0) {
+                    if (!params.antiprompt.empty()) {
                         // tokenize and inject first reverse prompt
                         const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
                         embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
@@ -712,28 +722,33 @@ int main(int argc, char ** argv) {
 
                     is_interacting = true;
                     printf("\n");
-                    console::set_display(console::user_input);
-                    fflush(stdout);
                 } else if (params.instruct) {
                     is_interacting = true;
                 }
             }
 
             if (n_past > 0 && is_interacting) {
+                LOG("waiting for user input\n");
+
                 if (params.instruct) {
                     printf("\n> ");
                 }
 
                 if (params.input_prefix_bos) {
+                    LOG("adding input prefix BOS token\n");
                     embd_inp.push_back(llama_token_bos(ctx));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     buffer += params.input_prefix;
                     printf("%s", buffer.c_str());
                 }
 
+                // color user input only
+                console::set_display(console::user_input);
+
                 std::string line;
                 bool another_line = true;
                 do {
@@ -749,25 +764,43 @@ int main(int argc, char ** argv) {
                 if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         buffer += params.input_suffix;
                         printf("%s", params.input_suffix.c_str());
                     }
 
+                    LOG("buffer: '%s'\n", buffer.c_str());
+
+                    const size_t original_size = embd_inp.size();
+
                     // instruct mode: insert instruction prefix
                     if (params.instruct && !is_antiprompt) {
+                        LOG("inserting instruction prefix\n");
                         n_consumed = embd_inp.size();
                         embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
                     }
 
-                    auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
+
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
                     // instruct mode: insert response suffix
                     if (params.instruct) {
+                        LOG("inserting instruction suffix\n");
                         embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
                     }
 
+                    for (size_t i = original_size; i < embd_inp.size(); ++i) {
+                        const llama_token token = embd_inp[i];
+                        output_tokens.push_back(token);
+                        output_ss << llama_token_to_piece(ctx, token);
+                    }
+
                     n_remain -= line_inp.size();
+                    LOG("n_remain: %d\n", n_remain);
+                } else {
+                    LOG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again
@@ -779,7 +812,7 @@ int main(int argc, char ** argv) {
                     if (grammar != NULL) {
                         llama_grammar_free(grammar);
 
-                        std::vector<const llama_grammar_element *> grammar_rules( parsed_grammar.c_rules());
+                        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
                         grammar = llama_grammar_init(
                             grammar_rules.data(), grammar_rules.size(),
                             parsed_grammar.symbol_ids.at("root"));
@@ -791,23 +824,26 @@ int main(int argc, char ** argv) {
 
         // end of text token
         if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) {
-            fprintf(stderr, " [end of text]\n");
+            LOG_TEE(" [end of text]\n");
             break;
         }
 
         // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
-        if (params.interactive && n_remain <= 0 && params.n_predict != -1) {
+        // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size).
+        if (params.interactive && n_remain <= 0 && params.n_predict >= 0) {
             n_remain = params.n_predict;
             is_interacting = true;
         }
     }
 
     if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
-        fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
+        LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
         llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
     llama_print_timings(ctx);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
+
     if (ctx_guidance) { llama_free(ctx_guidance); }
     llama_free(ctx);
     llama_free_model(model);
@@ -817,5 +853,9 @@ int main(int argc, char ** argv) {
     }
     llama_backend_free();
 
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
     return 0;
 }
diff --git a/examples/make-ggml.py b/examples/make-ggml.py
index 6a34eeac5..c73485ebf 100755
--- a/examples/make-ggml.py
+++ b/examples/make-ggml.py
@@ -1,22 +1,25 @@
 #!/usr/bin/env python3
 """
-This script converts Hugging Face llama models to GGML and quantizes them.
+This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them.
 
 Usage:
-python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
 
 Arguments:
-- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.
 - --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
 - --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
 - --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
 - --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
 
-Quant types:
+Old quant types (some base model types require these):
 - Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
 - Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
 - Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
 - Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+
+New quant types (recommended):
 - Q2_K: smallest, extreme quality loss - not recommended
 - Q3_K: alias for Q3_K_M
 - Q3_K_S: very small, very high quality loss
@@ -40,9 +43,7 @@ import argparse
 import os
 from huggingface_hub import snapshot_download
 
-def main(model, outname, outdir, quants, keep_fp16):
-    ggml_version = "v3"
-
+def main(model, model_type, outname, outdir, quants, keep_fp16):
     if not os.path.isdir(model):
         print(f"Model not found at {model}. Downloading...")
         try:
@@ -63,17 +64,20 @@ def main(model, outname, outdir, quants, keep_fp16):
     print("Building llama.cpp")
     subprocess.run(f"cd .. && make quantize", shell=True, check=True)
 
-    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+    fp16 = f"{outdir}/{outname}.gguf.fp16.bin"
 
-    print(f"Making unquantised GGML at {fp16}")
+    print(f"Making unquantised GGUF at {fp16}")
     if not os.path.isfile(fp16):
-        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+        if model_type != "llama":
+            subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True)
+        else:
+            subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
     else:
         print(f"Unquantised GGML already exists at: {fp16}")
 
     print("Making quants")
     for type in quants:
-        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        outfile = f"{outdir}/{outname}.gguf.{type}.bin"
         print(f"Making {type} : {outfile}")
         subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
 
@@ -81,8 +85,9 @@ def main(model, outname, outdir, quants, keep_fp16):
         os.remove(fp16)
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
-    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.')
     parser.add_argument('--outname', default=None, help='Output model(s) name')
     parser.add_argument('--outdir', default=None, help='Output directory')
     parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
@@ -90,4 +95,4 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)
+    main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16)
diff --git a/examples/parallel/CMakeLists.txt b/examples/parallel/CMakeLists.txt
new file mode 100644
index 000000000..0bbf89eae
--- /dev/null
+++ b/examples/parallel/CMakeLists.txt
@@ -0,0 +1,8 @@
+set(TARGET parallel)
+add_executable(${TARGET} parallel.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()
diff --git a/examples/parallel/README.md b/examples/parallel/README.md
new file mode 100644
index 000000000..4d0fe5cef
--- /dev/null
+++ b/examples/parallel/README.md
@@ -0,0 +1,3 @@
+# llama.cpp/example/parallel
+
+Simplified simluation for serving incoming requests in parallel
diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp
new file mode 100644
index 000000000..721888da7
--- /dev/null
+++ b/examples/parallel/parallel.cpp
@@ -0,0 +1,426 @@
+// A basic application simulating a server with multiple clients.
+// The clients submite requests to the server and they are processed in parallel.
+
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+#include <ctime>
+
+// trim whitespace from the beginning and end of a string
+static std::string trim(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+
+    while (start < end && isspace(str[start])) {
+        start += 1;
+    }
+
+    while (end > start && isspace(str[end - 1])) {
+        end -= 1;
+    }
+
+    return str.substr(start, end - start);
+}
+
+static std::string k_system =
+R"(Transcript of a never ending dialog, where the User interacts with an Assistant.
+The Assistant is helpful, kind, honest, good at writing, and never fails to answer the User's requests immediately and with precision.
+
+User: Recommend a nice restaurant in the area.
+Assistant: I recommend the restaurant "The Golden Duck". It is a 5 star restaurant with a great view of the city. The food is delicious and the service is excellent. The prices are reasonable and the portions are generous. The restaurant is located at 123 Main Street, New York, NY 10001. The phone number is (212) 555-1234. The hours are Monday through Friday from 11:00 am to 10:00 pm. The restaurant is closed on Saturdays and Sundays.
+User: Who is Richard Feynman?
+Assistant: Richard Feynman was an American physicist who is best known for his work in quantum mechanics and particle physics. He was awarded the Nobel Prize in Physics in 1965 for his contributions to the development of quantum electrodynamics. He was a popular lecturer and author, and he wrote several books, including "Surely You're Joking, Mr. Feynman!" and "What Do You Care What Other People Think?".
+User:)";
+
+static std::vector<std::string> k_prompts = {
+    "What is the meaning of life?",
+    "Tell me an interesting fact about llamas.",
+    "What is the best way to cook a steak?",
+    "Are you familiar with the Special Theory of Relativity and can you explain it to me?",
+    "Recommend some interesting books to read.",
+    "What is the best way to learn a new language?",
+    "How to get a job at Google?",
+    "If you could have any superpower, what would it be?",
+    "I want to learn how to play the piano.",
+};
+
+struct client {
+    int32_t id = 0;
+
+    llama_seq_id seq_id = -1;
+
+    llama_token sampled;
+
+    int64_t t_start_prompt;
+    int64_t t_start_gen;
+
+    int32_t n_prompt  = 0;
+    int32_t n_decoded = 0;
+    int32_t i_batch   = -1;
+
+    std::string input;
+    std::string prompt;
+    std::string response;
+
+    std::vector<llama_token> tokens_prev;
+};
+
+static void print_date_time() {
+    std::time_t current_time = std::time(nullptr);
+    std::tm* local_time = std::localtime(&current_time);
+    char buffer[80];
+    strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
+
+    printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
+}
+
+// Define a split string function to ...
+static std::vector<std::string> split_string(const std::string& input, char delimiter) {
+    std::vector<std::string> tokens;
+    std::istringstream stream(input);
+    std::string token;
+    while (std::getline(stream, token, delimiter)) {
+        tokens.push_back(token);
+    }
+    return tokens;
+}
+
+int main(int argc, char ** argv) {
+    srand(1234);
+
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    // number of simultaneous "clients" to simulate
+    const int32_t n_clients = params.n_parallel;
+
+    // requests to simulate
+    const int32_t n_seq = params.n_sequences;
+
+    // insert new requests as soon as the previous one is done
+    const bool cont_batching = params.cont_batching;
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("parallel", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the target model
+    params.logits_all = true;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+
+    // load the prompts from an external file if there are any
+    if (params.prompt.empty()) {
+        printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
+    } else {
+        // Output each line of the input params.prompts vector and copy to k_prompts
+        int index = 0;
+        printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
+
+        std::vector<std::string> prompts = split_string(params.prompt, '\n');
+        for (const auto& prompt : prompts) {
+            k_prompts.resize(index + 1);
+            k_prompts[index] = prompt;
+            index++;
+            printf("%3d prompt: %s\n", index, prompt.c_str());
+        }
+    }
+
+    fprintf(stderr, "\n\n");
+    fflush(stderr);
+
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_vocab = llama_n_vocab(model);
+
+    std::vector<client> clients(n_clients);
+    for (size_t i = 0; i < clients.size(); ++i) {
+        auto & client = clients[i];
+        client.id = i;
+        client.tokens_prev.resize(std::max(256, params.n_predict));
+        std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    std::vector<llama_token> tokens_system;
+    tokens_system = ::llama_tokenize(ctx, k_system, true);
+    const int32_t n_tokens_system = tokens_system.size();
+
+    llama_seq_id g_seq_id = 0;
+
+    // the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
+    // users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
+    llama_batch batch = llama_batch_init(params.n_ctx, 0);
+
+    int32_t n_total_prompt = 0;
+    int32_t n_total_gen    = 0;
+    int32_t n_cache_miss   = 0;
+
+    const auto t_main_start = ggml_time_us();
+
+    LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
+    LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    LOG_TEE("\n");
+
+    {
+        LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
+
+        batch.n_tokens = n_tokens_system;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            batch.token[i]  = tokens_system[i];
+            batch.pos[i]    = i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        // assign the system KV cache to all parallel sequences
+        for (int32_t i = 1; i < n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
+        }
+
+        LOG_TEE("\n");
+    }
+
+    LOG_TEE("Processing requests ...\n\n");
+
+    while (true) {
+        batch.n_tokens = 0;
+
+        // decode any currently ongoing sequences
+        for (auto & client : clients) {
+            if (client.seq_id == -1) {
+                continue;
+            }
+
+            batch.token [batch.n_tokens] = client.sampled;
+            batch.pos   [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
+            batch.seq_id[batch.n_tokens] = client.id;
+            batch.logits[batch.n_tokens] = true;
+
+            client.n_decoded += 1;
+            client.i_batch = batch.n_tokens;
+
+            batch.n_tokens += 1;
+        }
+
+        if (batch.n_tokens == 0) {
+            // all sequences have ended - clear the entire KV cache
+            for (int i = 0; i < n_clients; ++i) {
+                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
+            }
+
+            LOG_TEE("%s: clearing the KV cache\n", __func__);
+        }
+
+        // insert new sequences for decoding
+        if (cont_batching || batch.n_tokens == 0) {
+            for (auto & client : clients) {
+                if (client.seq_id == -1 && g_seq_id < n_seq) {
+                    client.seq_id = g_seq_id;
+
+                    client.t_start_prompt = ggml_time_us();
+                    client.t_start_gen    = 0;
+
+                    client.input    = k_prompts[rand() % k_prompts.size()];
+                    client.prompt   = client.input + "\nAssistant:";
+                    client.response = "";
+
+                    std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
+
+                    // do not prepend BOS because we have a system prompt!
+                    std::vector<llama_token> tokens_prompt;
+                    tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
+
+                    for (size_t i = 0; i < tokens_prompt.size(); ++i) {
+                        batch.token [batch.n_tokens] = tokens_prompt[i];
+                        batch.pos   [batch.n_tokens] = i + n_tokens_system;
+                        batch.seq_id[batch.n_tokens] = client.id;
+                        batch.logits[batch.n_tokens] = false;
+                        batch.n_tokens += 1;
+                    }
+
+                    // extract the logits only for the last token
+                    if (batch.n_tokens > 0) {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    client.n_prompt  = tokens_prompt.size();
+                    client.n_decoded = 0;
+                    client.i_batch   = batch.n_tokens - 1;
+
+                    LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+
+                    g_seq_id += 1;
+
+                    // insert new requests one-by-one
+                    //if (cont_batching) {
+                    //    break;
+                    //}
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0) {
+            break;
+        }
+
+        // process in chunks of params.n_batch
+        int32_t n_batch = params.n_batch;
+
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+            // experiment: process in powers of 2
+            //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
+            //    n_batch /= 2;
+            //    i -= n_batch;
+            //    continue;
+            //}
+
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            llama_batch batch_view = {
+                n_tokens,
+                batch.token  + i,
+                nullptr,
+                batch.pos    + i,
+                batch.seq_id + i,
+                batch.logits + i,
+                0, 0, 0, // unused
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+            if (ret != 0) {
+                if (n_batch == 1 || ret < 0) {
+                    // if you get here, it means the KV cache is full - try increasing it via the context size
+                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    return 1;
+                }
+
+                LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+
+                n_cache_miss += 1;
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+
+                continue;
+            }
+
+            LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
+
+            for (auto & client : clients) {
+                if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
+                    continue;
+                }
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
+                //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
+
+                const llama_token id = llama_sample_token(ctx, NULL, NULL, params, client.tokens_prev, candidates, client.i_batch - i);
+
+                if (client.n_decoded == 1) {
+                    // start measuring generation time after the first token to make sure all concurrent clients
+                    // have their prompt already processed
+                    client.t_start_gen = ggml_time_us();
+                }
+
+                // remember which tokens were sampled - used for repetition penalties during sampling
+                client.tokens_prev.erase(client.tokens_prev.begin());
+                client.tokens_prev.push_back(id);
+
+                const std::string token_str = llama_token_to_piece(ctx, id);
+                client.response += token_str;
+                client.sampled = id;
+
+                //printf("client %d, seq %d, token %d, pos %d, batch %d: %s\n",
+                //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
+
+                if (client.n_decoded > 2 &&
+                        (id == llama_token_eos(ctx) ||
+                         (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
+                         client.response.find("User:") != std::string::npos ||
+                         client.response.find('\n') != std::string::npos)) {
+                    // basic reverse prompt
+                    const size_t pos = client.response.find("User:");
+                    if (pos != std::string::npos) {
+                        client.response = client.response.substr(0, pos);
+                    }
+
+                    // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
+
+                    const auto t_main_end = ggml_time_us();
+
+                    LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput:    %s\n\033[35mResponse: %s\033[0m\n\n",
+                            client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
+                            (t_main_end - client.t_start_prompt) / 1e6,
+                            (double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
+                            n_cache_miss,
+                            ::trim(client.input).c_str(),
+                            ::trim(client.response).c_str());
+
+                    n_total_prompt += client.n_prompt;
+                    n_total_gen    += client.n_decoded;
+
+                    client.seq_id = -1;
+                }
+
+                client.i_batch = -1;
+            }
+        }
+    }
+
+    const auto t_main_end = ggml_time_us();
+
+    print_date_time();
+
+    LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
+    if (params.prompt_file.empty()) {
+        params.prompt_file = "used built-in defaults";
+    }
+    LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
+    LOG_TEE("Model and path used:  \033[32m%s\033[0m\n\n", params.model.c_str());
+
+    LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt              ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total gen tokens:    %6d, speed: %5.2f t/s\n", n_total_gen,    (double) (n_total_gen                 ) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Total speed (AVG):   %6s  speed: %5.2f t/s\n", "",             (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
+    LOG_TEE("Cache misses:        %6d\n", n_cache_miss);
+
+    LOG_TEE("\n");
+
+    llama_print_timings(ctx);
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}
diff --git a/examples/perplexity/README.md b/examples/perplexity/README.md
index eacfb17c6..50e1af011 100644
--- a/examples/perplexity/README.md
+++ b/examples/perplexity/README.md
@@ -1,3 +1,21 @@
 # perplexity
 
 TODO
+
+## Llama 2 70B Scorechart
+Quantization | Model size (GiB) | Perplexity | Delta to fp16
+-- | -- | -- | --
+Q4_0 | 36.20 | 3.5550 | 3.61%
+Q4_1 | 40.20 | 3.5125 | 2.37%
+Q5_0 | 44.20 | 3.4744 | 1.26%
+Q2_K | 27.27 | 3.7339 | 8.82%
+Q3_K_S | 27.86 | 3.7019 | 7.89%
+Q3_K_M | 30.83 | 3.5932 | 4.72%
+Q3_K_L | 33.67 | 3.5617 | 3.80%
+Q4_K_S | 36.39 | 3.4852 | 1.57%
+Q4_K_M | 38.54 | 3.4725 | 1.20%
+Q5_K_S | 44.20 | 3.4483 | 0.50%
+Q5_K_M | 45.41 | 3.4451 | 0.40%
+Q6_K | 52.70 | 3.4367 | 0.16%
+fp16 | 128.5 | 3.4313 | -
+
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index e89725efc..7d0038bd4 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -1,20 +1,88 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <cmath>
+#include <cstdio>
+#include <cstring>
 #include <ctime>
 #include <sstream>
-#include <cstring>
+#include <thread>
+#include <mutex>
+#include <vector>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-std::vector<float> softmax(const std::vector<float>& logits) {
+struct results_perplexity {
+    std::vector<llama_token> tokens;
+    double                   ppl_value;
+    std::vector<float>       logits;
+    std::vector<float>       probs;
+};
+
+struct results_log_softmax {
+    double log_softmax;
+    float  logit;
+    float  prob;
+};
+
+static void write_logfile(
+    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const struct results_perplexity & results
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    if (params.hellaswag) {
+        fprintf(stderr, "%s: warning: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
+        return;
+    }
+
+    const std::string timestamp = get_sortable_timestamp();
+
+    const bool success = create_directory_with_parents(params.logdir);
+    if (!success) {
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: main\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Perplexity Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    dump_vector_float_yaml(logfile, "logits", results.logits);
+    fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
+    dump_vector_float_yaml(logfile, "probs", results.probs);
+
+    llama_dump_timing_info_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
+static std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
-    for (float v : logits) max_logit = std::max(max_logit, v);
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
     double sum_exp = 0.0;
     for (size_t i = 0; i < logits.size(); i++) {
         // Subtract the maximum logit value from the current logit value for numerical stability
@@ -23,37 +91,106 @@ std::vector<float> softmax(const std::vector<float>& logits) {
         sum_exp += exp_logit;
         probs[i] = exp_logit;
     }
-    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
     return probs;
 }
 
-void perplexity_v2(llama_context * ctx, const gpt_params & params) {
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
+}
 
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
+        double local_nll  = 0;
+        double local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const double v = -results.log_softmax;
+            local_nll += v;
+            local_nll2 += v*v;
+
+            logit_history[i] = results.logit;
+            prob_history[i]  = results.prob;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+}
+
+static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
 
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
+    std::vector<float> logit_history;
+    std::vector<float> prob_history;
+
+    logit_history.resize(tokens.size());
+    prob_history.resize(tokens.size());
+
     if (params.ppl_stride <= 0) {
         fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
-        return;
+        return {tokens, -1, logit_history, prob_history};
     }
-    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int calc_chunk = params.n_ctx;
+    const int calc_chunk = n_ctx;
 
     fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
 
     if (int(tokens.size()) <= calc_chunk) {
         fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
-                tokens.size(), params.n_ctx, params.ppl_stride);
-        return;
+                tokens.size(), n_ctx, params.ppl_stride);
+        return {tokens, -1, logit_history, prob_history};
     }
 
     const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1)  / params.ppl_stride;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
@@ -72,21 +209,24 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) {
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
             //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 //fprintf(stderr, "%s : failed to eval\n", __func__);
-                return;
+                return {tokens, -1, logit_history, prob_history};
             }
 
             // save original token and restore it after eval
             const auto token_org = tokens[batch_start];
 
             // add BOS token for the first batch of each chunk
-            if (j == 0) {
+            if (add_bos && j == 0) {
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
@@ -112,7 +252,7 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) {
         }
 
         //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
-        for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) {
+        for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
 
             // Calculate probability of next token, given the previous ones.
             const std::vector<float> tok_logits(
@@ -120,6 +260,8 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) {
                 logits.begin() + (j + 1) * n_vocab);
 
             const float prob = softmax(tok_logits)[tokens[start + j + 1]];
+            logit_history[start + j + 1] = tok_logits[tokens[start + j + 1]];
+            prob_history[start + j + 1]  = prob;
 
             nll += -std::log(prob);
             ++count;
@@ -133,42 +275,72 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) {
         fflush(stdout);
     }
     printf("\n");
+
+    return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-void perplexity(llama_context * ctx, const gpt_params & params) {
-
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     if (params.ppl_stride > 0) {
-        perplexity_v2(ctx, params);
-        return;
+        return perplexity_v2(ctx, params);
     }
 
     // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
-    auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int n_chunk_max = tokens.size() / params.n_ctx;
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = is_spm;
+    const int n_ctx = llama_n_ctx(ctx);
+
+    auto tim1 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return {std::move(tokens), 0., {}, {}};
+    }
+
+    std::vector<float> logit_history;
+    logit_history.resize(tokens.size());
+
+    std::vector<float> prob_history;
+    prob_history.resize(tokens.size());
+
+    const int n_chunk_max = tokens.size() / n_ctx;
 
     const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
     const int n_batch = params.n_batch;
 
     int count = 0;
     double nll = 0.0;
+    double nll2 = 0.0;
 
     fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
-    for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * params.n_ctx;
-        const int end   = start + params.n_ctx;
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
-        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
 
         std::vector<float> logits;
 
         const auto t_start = std::chrono::high_resolution_clock::now();
 
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -177,19 +349,19 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             const auto token_org = tokens[batch_start];
 
             // add BOS token for the first batch of each chunk
-            if (j == 0) {
+            if (add_bos && j == 0) {
                 tokens[batch_start] = llama_token_bos(ctx);
             }
 
-            if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
-                return;
+                return {tokens, -1, logit_history, prob_history};
             }
 
             // restore the original token in case it was set to BOS
             tokens[batch_start] = token_org;
 
-            const auto batch_logits = llama_get_logits(ctx);
+            const auto * batch_logits = llama_get_logits(ctx);
             logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
         }
 
@@ -218,37 +390,48 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         // Example, we have a context window of 512, we will compute perplexity for each of the
         // last 256 tokens.  Then, we split the input up into context window size chunks to
         // process the entire prompt.
-        for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
-            // Calculate probability of next token, given the previous ones.
-            const std::vector<float> tok_logits(
-                logits.begin() + (j + 0) * n_vocab,
-                logits.begin() + (j + 1) * n_vocab);
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                       workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        count += n_ctx - first - 1;
 
-            const float prob = softmax(tok_logits)[tokens[start + j + 1]];
-
-            nll += -std::log(prob);
-            ++count;
-        }
         // perplexity is e^(average negative log-likelihood)
         if (params.ppl_output_type == 0) {
             printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
         } else {
-            printf("%8d  %.4lf\n", i*params.n_ctx, std::exp(nll / count));
+            double av = nll/count;
+            double av2 = nll2/count - av*av;
+            if (av2 > 0) av2 = sqrt(av2/(count-1));
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
     }
     printf("\n");
+
+    nll2 /= count;
+    nll /= count;
+    const double ppl = exp(nll);
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
+
+    return {tokens, ppl, logit_history, prob_history};
 }
 
-std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vector<int>& tokens, int n_past, int n_batch,
-        int n_vocab, int n_thread) {
+static std::vector<float> hellaswag_evaluate_tokens(
+    llama_context * ctx, std::vector<int> & tokens, int n_past, int n_batch, int n_vocab
+) {
     std::vector<float> result;
     result.reserve(tokens.size() * n_vocab);
     size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch;
     for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) {
         size_t n_tokens = tokens.size() - i_chunk * n_batch;
         n_tokens = std::min(n_tokens, size_t(n_batch));
-        if (llama_eval(ctx, tokens.data() + i_chunk * n_batch, n_tokens, n_past, n_thread)) {
+        if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0))) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return {};
         }
@@ -261,7 +444,7 @@ std::vector<float> hellaswag_evaluate_tokens(llama_context * ctx, const std::vec
     return result;
 }
 
-void hellaswag_score(llama_context * ctx, const gpt_params & params) {
+static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     // Calculates hellaswag score (acc_norm) from prompt
     //
     // Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
@@ -295,8 +478,11 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     size_t hs_task_count = prompt_lines.size()/6;
     fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
 
+    const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
+    fprintf(stderr, "================================= is_spm = %d\n", is_spm);
+
     // This is needed as usual for LLaMA models
-    bool prepend_bos = true;
+    const bool add_bos = is_spm;
 
     // Number of tasks to use when computing the score
     if ( params.hellaswag_tasks < hs_task_count  ) {
@@ -334,7 +520,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         hs_data[i].context = prompt_lines[idx*6];
         hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
         for (size_t j=0; j < 4; j++) {
-            hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j];
+            hs_data[i].ending[j] = prompt_lines[idx*6+2+j];
         }
 
         // Delete the selected random example from the prompt
@@ -347,24 +533,36 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     printf("\ntask\tacc_norm\n");
 
     double acc = 0.0f;
-    const int n_vocab = llama_n_vocab(ctx);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
+
+    std::vector<std::vector<int>> ending_tokens(4);
 
     std::vector<float> tok_logits(n_vocab);
 
     for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) {
-
         // Tokenize the context to count tokens
-        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos);
+        std::vector<int> context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos);
         size_t context_size = context_embd.size();
 
+        for (int i = 0; i < 4; ++i) {
+            ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + " " + hs_data[task_idx].ending[i], add_bos);
+            for (int k = 0; k < int(context_size); ++k) {
+                if (ending_tokens[i][k] != context_embd[k]) {
+                    fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k);
+                    break;
+                }
+            }
+        }
+
         // Do the 1st ending
         // In this case we include the context when evaluating
-        auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], prepend_bos);
+        //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos);
+        auto query_embd = ending_tokens[0];
         auto query_size = query_embd.size();
-        //printf("First query: %d\n",(int)query_size);
 
         // Stop if query wont fit the ctx window
-        if (query_size > (size_t)params.n_ctx) {
+        if (query_size > (size_t)n_ctx) {
             fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
             return;
         }
@@ -374,7 +572,10 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             query_embd.resize(32);
         }
 
-        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab, params.n_threads);
+        // clear the KV cache
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
+
+        auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab);
         if (logits.empty()) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return;
@@ -406,11 +607,12 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) {
 
             // Tokenize the query
-            query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false);
+            query_embd.resize(ending_tokens[ending_idx].size() - context_size);
+            std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int));
             query_size = query_embd.size();
 
             // Stop if query wont fit the ctx window
-            if (context_size + query_size > (size_t)params.n_ctx) {
+            if (context_size + query_size > (size_t)n_ctx) {
                 fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size);
                 return;
             }
@@ -422,7 +624,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             //}
 
             // Evaluate the query
-            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab, params.n_threads);
+            logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab);
             if (logits.empty()) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return;
@@ -480,11 +682,11 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     params.n_batch = 512;
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
-    params.perplexity = true;
+    params.logits_all = true;
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     if (params.ppl_stride > 0) {
@@ -493,12 +695,7 @@ int main(int argc, char ** argv) {
         params.n_ctx += params.ppl_stride/2;
     }
 
-    if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
-    }
-
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -523,20 +720,28 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
-                params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
+    struct results_perplexity results;
     if (params.hellaswag) {
         hellaswag_score(ctx, params);
     } else {
-        perplexity(ctx, params);
+        results = perplexity(ctx, params);
     }
 
     llama_print_timings(ctx);
+    write_logfile(ctx, params, model, results);
+
     llama_free(ctx);
     llama_free_model(model);
 
diff --git a/examples/quantize-stats/CMakeLists.txt b/examples/quantize-stats/CMakeLists.txt
index c5c394058..db182e263 100644
--- a/examples/quantize-stats/CMakeLists.txt
+++ b/examples/quantize-stats/CMakeLists.txt
@@ -2,4 +2,5 @@ set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp
index 06ce18f09..dd76b1cee 100644
--- a/examples/quantize-stats/quantize-stats.cpp
+++ b/examples/quantize-stats/quantize-stats.cpp
@@ -1,7 +1,7 @@
-#include "ggml.h"
-#include "build-info.h"
-
 #define LLAMA_API_INTERNAL
+#include "build-info.h"
+#include "common.h"
+#include "ggml.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -34,8 +34,8 @@ struct quantize_stats_params {
     std::vector<enum ggml_type> include_types;
 };
 
-const size_t HISTOGRAM_BUCKETS = 150;
-const double HISTOGRAM_RANGE = 0.03;
+constexpr size_t HISTOGRAM_BUCKETS = 150;
+constexpr double HISTOGRAM_RANGE = 0.03;
 
 struct error_stats {
     size_t num_samples;
@@ -44,8 +44,7 @@ struct error_stats {
     uint64_t error_histogram[HISTOGRAM_BUCKETS];
 };
 
-
-void quantize_stats_print_usage(int /*argc*/, char ** argv) {
+static void quantize_stats_print_usage(int /*argc*/, char ** argv) {
     quantize_stats_params params;
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
@@ -71,7 +70,7 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
 }
 
 // Check if a layer is included/excluded by command line
-bool layer_included(const quantize_stats_params params, const std::string & layer) {
+static bool layer_included(const quantize_stats_params & params, const std::string & layer) {
     for (const auto& excluded : params.exclude_layers) {
         if (std::regex_search(layer, std::regex(excluded))) {
             return false;
@@ -86,7 +85,7 @@ bool layer_included(const quantize_stats_params params, const std::string & laye
 }
 
 // Update error statistics given vectors with the before/after result of quantization
-void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
+static void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) {
     for (int64_t i = 0; i < nelements; i++) {
         double diff = input[i] - output[i];
         stats.total_error += diff * diff;
@@ -96,14 +95,14 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
     stats.num_samples += nelements;
 }
 
-void combine_error_stats(error_stats & into, const error_stats & from) {
+static void combine_error_stats(error_stats & into, const error_stats & from) {
     into.num_samples += from.num_samples;
     into.total_error += from.total_error;
     if (from.max_error > into.max_error) into.max_error = from.max_error;
     for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
 }
 
-double find_quantile(const error_stats & stats, double quantile) {
+static double find_quantile(const error_stats & stats, double quantile) {
     double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
 
     double accum = 0;
@@ -116,7 +115,7 @@ double find_quantile(const error_stats & stats, double quantile) {
     return INFINITY;
 }
 
-void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
+static void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) {
     double rmse = sqrt(stats.total_error / (double) stats.num_samples);
     double median = find_quantile(stats, .5);
     double pct95 = find_quantile(stats, .95);
@@ -143,17 +142,10 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
-void test_roundtrip_on_chunk(
-        const ggml_tensor * layer,
-        int64_t offset,
-        int64_t chunk_size,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        float * input_scratch,
-        char * quantized_scratch,
-        float * output_scratch,
-        error_stats & stats) {
-
+static void test_roundtrip_on_chunk(
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
+    float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
+) {
     if (layer->type == GGML_TYPE_F16) {
         for (int i = 0; i < chunk_size; i++) {
             input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
@@ -174,18 +166,11 @@ void test_roundtrip_on_chunk(
 
 
 // Run quantization function for a single layer and update error stats
-void test_roundtrip_on_layer(
-        std::string & name,
-        bool print_layer_stats,
-        const ggml_type_traits_t & qfns,
-        bool use_reference,
-        const ggml_tensor * layer,
-        std::vector<float> & input_scratch,
-        std::vector<char> & quantized_scratch,
-        std::vector<float> & output_scratch,
-        error_stats & total_error,
-        int max_thread = 0) {
-
+static void test_roundtrip_on_layer(
+    std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
+    const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
+    std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
+) {
     assert(tensor_is_contiguous(layer));
     error_stats layer_error {};
     uint64_t nelements = ggml_nelements(layer);
@@ -314,7 +299,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     // load the model
     fprintf(stderr, "Loading model\n");
@@ -324,21 +309,22 @@ int main(int argc, char ** argv) {
     llama_context * ctx;
 
     {
-        auto lparams = llama_context_default_params();
+        auto mparams = llama_model_default_params();
+        mparams.use_mlock  = false;
 
-        lparams.n_ctx      = 256;
-        lparams.seed       = 1;
-        lparams.f16_kv     = false;
-        lparams.use_mlock  = false;
-
-        model = llama_load_model_from_file(params.model.c_str(), lparams);
+        model = llama_load_model_from_file(params.model.c_str(), mparams);
 
         if (model == NULL) {
             fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
             return 1;
         }
 
-        ctx = llama_new_context_with_model(model, lparams);
+        auto cparams = llama_context_default_params();
+        cparams.n_ctx      = 256;
+        cparams.seed       = 1;
+        cparams.f16_kv     = false;
+
+        ctx = llama_new_context_with_model(model, cparams);
 
         if (ctx == NULL) {
             fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
diff --git a/examples/quantize/CMakeLists.txt b/examples/quantize/CMakeLists.txt
index 47d0be72e..4a8eed544 100644
--- a/examples/quantize/CMakeLists.txt
+++ b/examples/quantize/CMakeLists.txt
@@ -2,6 +2,7 @@ set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)
diff --git a/examples/quantize/README.md b/examples/quantize/README.md
index f349e913e..c8b9a27a0 100644
--- a/examples/quantize/README.md
+++ b/examples/quantize/README.md
@@ -1,3 +1,44 @@
 # quantize
 
 TODO
+
+## Llama 2 7B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.35
+Q3_K_S | 3.50
+Q3_K_M | 3.91
+Q3_K_L | 4.27
+Q4_K_S | 4.58
+Q4_K_M | 4.84
+Q5_K_S | 5.52
+Q5_K_M | 5.68
+Q6_K | 6.56
+
+## Llama 2 13B
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.34
+Q3_K_S | 3.48
+Q3_K_M | 3.89
+Q3_K_L | 4.26
+Q4_K_S | 4.56
+Q4_K_M | 4.83
+Q5_K_S | 5.51
+Q5_K_M | 5.67
+Q6_K | 6.56
+
+# Llama 2 70B
+
+Quantization | Bits per Weight (BPW)
+-- | --
+Q2_K | 3.40
+Q3_K_S | 3.47
+Q3_K_M | 3.85
+Q3_K_L | 4.19
+Q4_K_S | 4.53
+Q4_K_M | 4.80
+Q5_K_S | 5.50
+Q5_K_M | 5.65
+Q6_K | 6.56
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp
index d172f645a..c7dd0d894 100644
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -1,5 +1,5 @@
 #include "build-info.h"
-
+#include "common.h"
 #include "llama.h"
 
 #include <cstdio>
@@ -35,10 +35,12 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },
+    // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
+    { "COPY",   LLAMA_FTYPE_ALL_F32,       "only copy tensors, no quantizing", },
 };
 
 
-bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
+static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
     std::string ftype_str;
 
     for (auto ch : ftype_str_in) {
@@ -70,13 +72,19 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
 // usage:
 //  ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
 //
-void usage(const char * executable) {
-    fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
-    fprintf(stderr, "  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
-    fprintf(stderr, "  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
-    fprintf(stderr, "\nAllowed quantization types:\n");
+[[noreturn]]
+static void usage(const char * executable) {
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
+    printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
+    printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
-        printf("  %2d  or  %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
+        if (it.name != "COPY") {
+            printf("  %2d  or  ", it.ftype);
+        } else {
+            printf("          ");
+        }
+        printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }
@@ -100,7 +108,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if (argc - arg_idx < 3) {
+    if (argc - arg_idx < 2) {
         usage(argv[0]);
     }
 
@@ -114,13 +122,16 @@ int main(int argc, char ** argv) {
     std::string ftype_str;
     if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
         std::string fpath;
-        const size_t pos = fname_inp.find_last_of('/');
+        const size_t pos = fname_inp.find_last_of("/\\");
         if (pos != std::string::npos) {
             fpath = fname_inp.substr(0, pos + 1);
         }
         // export as [inp path]/ggml-model-[ftype].gguf
         fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
         arg_idx++;
+        if (ftype_str == "COPY") {
+            params.only_copy = true;
+        }
     }
     else {
         fname_out = argv[arg_idx];
@@ -134,6 +145,9 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
             return 1;
         }
+        if (ftype_str == "COPY") {
+           params.only_copy = true;
+        }
         arg_idx++;
     }
 
@@ -148,7 +162,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
     if (params.nthread > 0) {
diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp
index 3db61b754..acc6dbdfd 100644
--- a/examples/save-load-state/save-load-state.cpp
+++ b/examples/save-load-state/save-load-state.cpp
@@ -1,6 +1,6 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 
 #include <vector>
 #include <cstdio>
@@ -13,38 +13,32 @@ int main(int argc, char ** argv) {
     params.repeat_last_n = 64;
     params.prompt = "The quick brown fox";
 
-    if (gpt_params_parse(argc, argv, params) == false) {
+    if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    print_build_info();
 
     if (params.n_predict < 0) {
         params.n_predict = 16;
     }
 
-    auto lparams = llama_context_default_params();
-
-    lparams.n_ctx     = params.n_ctx;
-    lparams.seed      = params.seed;
-    lparams.f16_kv    = params.memory_f16;
-    lparams.use_mmap  = params.use_mmap;
-    lparams.use_mlock = params.use_mlock;
-
     auto n_past = 0;
     auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
 
     // init
-    auto model = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model;
+    llama_context * ctx;
+
+    std::tie(model, ctx) = llama_init_from_gpt_params( params );
     if (model == nullptr) {
         return 1;
     }
-    auto ctx = llama_new_context_with_model(model, lparams);
     if (ctx == nullptr) {
         llama_free_model(model);
         return 1;
     }
-    auto tokens = llama_tokenize(ctx, params.prompt.c_str(), true);
+    auto tokens = llama_tokenize(ctx, params.prompt, true);
     auto n_prompt_tokens = tokens.size();
     if (n_prompt_tokens < 1) {
         fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
@@ -54,7 +48,7 @@ int main(int argc, char ** argv) {
     }
 
     // evaluate prompt
-    llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
+    llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
 
     last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
     n_past += n_prompt_tokens;
@@ -78,8 +72,8 @@ int main(int argc, char ** argv) {
     printf("\n%s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
+        auto * logits = llama_get_logits(ctx);
+        auto n_vocab = llama_n_vocab(model);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -87,11 +81,11 @@ int main(int argc, char ** argv) {
         }
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx, &candidates_p);
-        auto next_token_str = llama_token_to_str(ctx, next_token);
+        auto next_token_str = llama_token_to_piece(ctx, next_token);
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx);
             llama_free_model(model);
@@ -106,7 +100,7 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
 
     // make new context
-    auto ctx2 = llama_new_context_with_model(model, lparams);
+    auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
     // Load state (rng, logits, embedding and kv_cache) from file
     {
@@ -138,8 +132,8 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto logits = llama_get_logits(ctx2);
-        auto n_vocab = llama_n_vocab(ctx2);
+        auto * logits = llama_get_logits(ctx2);
+        auto n_vocab = llama_n_vocab(model);
         std::vector<llama_token_data> candidates;
         candidates.reserve(n_vocab);
         for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@@ -147,11 +141,11 @@ int main(int argc, char ** argv) {
         }
         llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
         auto next_token = llama_sample_token(ctx2, &candidates_p);
-        auto next_token_str = llama_token_to_str(ctx2, next_token);
+        auto next_token_str = llama_token_to_piece(ctx2, next_token);
         last_n_tokens_data.push_back(next_token);
 
         printf("%s", next_token_str.c_str());
-        if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
+        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
diff --git a/examples/server/README.md b/examples/server/README.md
index 77997f98d..8a079ae26 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -4,14 +4,14 @@ This example demonstrates a simple HTTP API server and a simple web front end to
 
 Command line options:
 
--   `--threads N`, `-t N`: Set the number of threads to use during computation.
+-   `--threads N`, `-t N`: Set the number of threads to use during generation.
+-   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
 -   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 -   `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 -   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
 -   `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 -   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
@@ -77,34 +77,31 @@ You need to have [Node.js](https://nodejs.org/en) installed.
 ```bash
 mkdir llama-client
 cd llama-client
-npm init
-npm install axios
 ```
 
 Create a index.js file and put inside this:
 
 ```javascript
-const axios = require("axios");
-
 const prompt = `Building a website can be done in 10 simple steps:`;
 
 async function Test() {
-    let result = await axios.post("http://127.0.0.1:8080/completion", {
-        prompt,
-        n_predict: 512,
-    });
-
-    // the response is received until completion finish
-    console.log(result.data.content);
+    let response = await fetch("http://127.0.0.1:8080/completion", {
+        method: 'POST',
+        body: JSON.stringify({
+            prompt,
+            n_predict: 512,
+        })
+    })
+    console.log((await response.json()).content)
 }
 
-Test();
+Test()
 ```
 
 And run it:
 
 ```bash
-node .
+node index.js
 ```
 
 ## API Endpoints
@@ -117,9 +114,9 @@ node .
 
     `top_k`: Limit the next token selection to the K most probable tokens (default: 40).
 
-    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
+    `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
 
-    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: 128, -1 = infinity).
+    `n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
 
     `n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
     By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
@@ -159,6 +156,8 @@ node .
 
     `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
 
+    `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
+
 -   **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
@@ -167,12 +166,28 @@ node .
 
     Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
+-   **POST** `/detokenize`: Convert tokens to text.
+
+    *Options:*
+
+    `tokens`: Set the tokens to detokenize.
+
 -   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 
     *Options:*
 
     `content`: Set the text to process.
 
+    **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+
+    *Options:*
+
+    `input_prefix`: Set the prefix of the code to infill.
+
+    `input_suffix`: Set the suffix of the code to infill.
+
+    It also accepts all the options of `/completion` except `stream` and `prompt`.
+
 ## More examples
 
 ### Interactive mode
diff --git a/examples/server/api_like_OAI.py b/examples/server/api_like_OAI.py
index 79af1768b..14d2dcf65 100755
--- a/examples/server/api_like_OAI.py
+++ b/examples/server/api_like_OAI.py
@@ -31,8 +31,6 @@ def is_present(json, key):
         return False
     return True
 
-
-
 #convert chat to prompt
 def convert_chat(messages):
     prompt = "" + args.chat_prompt.replace("\\n", "\n")
diff --git a/examples/server/index.html.hpp b/examples/server/index.html.hpp
index 76851d9c6..f30232929 100644
--- a/examples/server/index.html.hpp
+++ b/examples/server/index.html.hpp
@@ -152,1547 +152,2187 @@ unsigned char index_html[] = {
   0x6f, 0x70, 0x65, 0x6e, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30,
   0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72,
-  0x65, 0x61, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70,
-  0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d,
-  0x67, 0x72, 0x6f, 0x77, 0x3a, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x31, 0x30,
-  0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x70, 0x72, 0x65, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70,
-  0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73,
+  0x65, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70,
+  0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x2e, 0x33, 0x65,
+  0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72,
+  0x64, 0x65, 0x72, 0x2d, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x3a, 0x20,
+  0x31, 0x70, 0x78, 0x20, 0x73, 0x6f, 0x6c, 0x69, 0x64, 0x20, 0x23, 0x63,
+  0x63, 0x63, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e,
+  0x3a, 0x20, 0x61, 0x62, 0x73, 0x6f, 0x6c, 0x75, 0x74, 0x65, 0x3b, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72,
   0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20,
-  0x23, 0x32, 0x32, 0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x64, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a,
-  0x20, 0x6d, 0x6f, 0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e,
-  0x67, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33,
-  0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f,
-  0x72, 0x64, 0x65, 0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a,
-  0x20, 0x33, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65,
-  0x74, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20,
-  0x30, 0x2e, 0x35, 0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20,
-  0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72,
-  0x2c, 0x20, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c,
-  0x69, 0x67, 0x6e, 0x3a, 0x20, 0x63, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b,
+  0x77, 0x68, 0x69, 0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x2e,
+  0x32, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62,
+  0x6f, 0x78, 0x2d, 0x73, 0x68, 0x61, 0x64, 0x6f, 0x77, 0x3a, 0x20, 0x30,
+  0x20, 0x30, 0x20, 0x31, 0x30, 0x70, 0x78, 0x20, 0x72, 0x67, 0x62, 0x61,
+  0x28, 0x30, 0x2c, 0x20, 0x30, 0x2c, 0x20, 0x30, 0x2c, 0x20, 0x30, 0x2e,
+  0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64,
+  0x69, 0x6e, 0x67, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, 0x67, 0x72, 0x6f,
+  0x77, 0x3a, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b,
   0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x73, 0x69, 0x7a, 0x65,
-  0x3a, 0x20, 0x38, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x38, 0x38, 0x38,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x3c, 0x2f,
-  0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c, 0x73,
-  0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
-  0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x68, 0x2c,
-  0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65, 0x66, 0x66,
-  0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65,
-  0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x75,
-  0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75, 0x73,
-  0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73, 0x65,
-  0x52, 0x65, 0x66, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66, 0x72,
-  0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e, 0x6a,
-  0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70,
-  0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20,
-  0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d,
-  0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x3b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20,
-  0x7b, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76,
-  0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d,
-  0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68, 0x65,
-  0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61,
-  0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69,
-  0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d,
-  0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73,
-  0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61, 0x74,
-  0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20,
-  0x75, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, 0x65, 0x6e, 0x64,
-  0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, 0x74, 0x2e, 0x20,
-  0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x73,
-  0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f,
-  0x77, 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b,
-  0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c,
-  0x6e, 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d,
-  0x5c, 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22,
-  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74,
-  0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a,
-  0x20, 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20,
-  0x7b, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22,
-  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e,
-  0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20,
-  0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75,
-  0x73, 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65,
-  0x64, 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61,
-  0x74, 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f,
-  0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c,
-  0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61,
-  0x62, 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x2c,
-  0x20, 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78,
-  0x74, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61,
-  0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, 0x2f,
-  0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61,
-  0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
-  0x6f, 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, 0x2f,
-  0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, 0x65,
-  0x20, 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a,
-  0x20, 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30,
+  0x70, 0x72, 0x65, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79,
+  0x3a, 0x20, 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e,
+  0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x32, 0x32,
+  0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c,
+  0x6f, 0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x64, 0x65,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e,
+  0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, 0x20, 0x6d, 0x6f,
+  0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20,
+  0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65,
+  0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x33, 0x70,
+  0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x35,
+  0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f,
+  0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x66,
+  0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c, 0x69, 0x67, 0x6e,
+  0x3a, 0x20, 0x63, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6f,
+  0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x38,
+  0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x38, 0x38, 0x38, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40,
+  0x6b, 0x65, 0x79, 0x66, 0x72, 0x61, 0x6d, 0x65, 0x73, 0x20, 0x6c, 0x6f,
+  0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, 0x70,
+  0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x30, 0x25,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62,
+  0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f,
+  0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x30, 0x25, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x31, 0x30, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f,
+  0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e,
+  0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f,
+  0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d,
+  0x31, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x30, 0x30,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f,
+  0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d,
+  0x32, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x66, 0x66,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b,
+  0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a,
+  0x20, 0x35, 0x30, 0x25, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f,
+  0x75, 0x6e, 0x64, 0x2d, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x61, 0x72, 0x2d, 0x67, 0x72, 0x61, 0x64, 0x69, 0x65,
+  0x6e, 0x74, 0x28, 0x39, 0x30, 0x64, 0x65, 0x67, 0x2c, 0x20, 0x76, 0x61,
+  0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d,
+  0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x2c, 0x20, 0x76, 0x61,
+  0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d,
+  0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x29, 0x2c, 0x20, 0x76, 0x61,
+  0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d,
+  0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69,
+  0x6f, 0x6e, 0x3a, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d,
+  0x62, 0x67, 0x2d, 0x77, 0x69, 0x70, 0x65, 0x20, 0x32, 0x73, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x61, 0x72, 0x20, 0x69, 0x6e, 0x66, 0x69, 0x6e, 0x69,
+  0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x40, 0x6d, 0x65, 0x64, 0x69, 0x61, 0x20, 0x28, 0x70,
+  0x72, 0x65, 0x66, 0x65, 0x72, 0x73, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72,
+  0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x65, 0x3a, 0x20, 0x64, 0x61, 0x72,
+  0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
+  0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64,
+  0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x3a,
+  0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x30, 0x30, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f,
+  0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d,
+  0x32, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x66, 0x66,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72,
+  0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67,
+  0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a,
+  0x20, 0x62, 0x6c, 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x3c, 0x2f, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20,
+  0x3c, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65,
+  0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20,
+  0x68, 0x2c, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65,
+  0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75,
+  0x74, 0x65, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c,
+  0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20,
+  0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75,
+  0x73, 0x65, 0x52, 0x65, 0x66, 0x2c, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f,
+  0x6e, 0x65, 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e,
+  0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74,
+  0x20, 0x7b, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e,
+  0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f,
+  0x6d, 0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68,
+  0x65, 0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d,
+  0x61, 0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x73, 0x73,
+  0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c,
+  0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69,
+  0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e,
+  0x20, 0x55, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x4c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, 0x65, 0x6e,
+  0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, 0x74, 0x2e,
+  0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, 0x20, 0x68, 0x65,
+  0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, 0x6e, 0x64, 0x2c,
+  0x20, 0x68, 0x6f, 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, 0x67, 0x6f, 0x6f,
+  0x64, 0x20, 0x61, 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, 0x69, 0x6e, 0x67,
+  0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, 0x65, 0x72, 0x20,
+  0x66, 0x61, 0x69, 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x61, 0x6e, 0x73,
+  0x77, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, 0x65, 0x71, 0x75,
+  0x65, 0x73, 0x74, 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, 0x64, 0x69, 0x61,
+  0x74, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x77, 0x69, 0x74,
+  0x68, 0x20, 0x70, 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
+  0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, 0x70, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, 0x6e, 0x7b, 0x7b,
+  0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, 0x5c, 0x6e, 0x7b,
+  0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79,
+  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b,
+  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, 0x7b, 0x7b, 0x6d,
+  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
+  0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, 0x22, 0x63, 0x68,
+  0x61, 0x74, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x22,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72,
+  0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
+  0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63,
+  0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72,
+  0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73,
+  0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c, 0x20, 0x2f, 0x2f,
+  0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65,
+  0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x2c, 0x20, 0x2d, 0x31,
+  0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20, 0x73,
+  0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79,
+  0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31,
+  0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65,
+  0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f,
+  0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x3c, 0x3d,
+  0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, 0x65, 0x20, 0x76, 0x6f,
+  0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e,
+  0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20,
+  0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x3a, 0x20, 0x31, 0x2e,
+  0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20,
+  0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70,
+  0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e,
+  0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x65, 0x73, 0x65,
+  0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a,
+  0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30,
   0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x3a,
-  0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65,
+  0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a,
+  0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30,
   0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61,
-  0x6c, 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f,
-  0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62,
-  0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72,
-  0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c,
-  0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20,
-  0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c,
-  0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, 0x65,
-  0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c,
-  0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20,
-  0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c,
-  0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72,
-  0x6f, 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f,
-  0x20, 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61,
-  0x75, 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, 0x72,
-  0x67, 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74,
-  0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x2c,
-  0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, 0x67,
-  0x20, 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, 0x2c,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, 0x20, 0x53,
-  0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73,
-  0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20,
-  0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65,
-  0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65,
-  0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61,
-  0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73,
-  0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61,
-  0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x22, 0x6c, 0x6c, 0x61,
-  0x6d, 0x61, 0x63, 0x70, 0x70, 0x5f, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
-  0x5f, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61,
-  0x67, 0x65, 0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75,
-  0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c,
-  0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74,
-  0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65,
-  0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74,
-  0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67,
-  0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f,
-  0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f,
-  0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b,
-  0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20,
-  0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69,
-  0x66, 0x79, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f,
-  0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f,
-  0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x52,
-  0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20,
-  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74,
-  0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65,
-  0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72,
-  0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b,
-  0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74,
-  0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f,
-  0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f,
-  0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a,
-  0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69,
-  0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53,
-  0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74,
-  0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f,
+  0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2f,
+  0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69,
+  0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x3a, 0x20,
+  0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f,
+  0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x2c, 0x20, 0x2f, 0x2f,
+  0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, 0x67, 0x20, 0x72, 0x61,
+  0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61,
+  0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x3a,
+  0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f,
+  0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f,
+  0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e,
+  0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73,
+  0x20, 0x69, 0x6e, 0x20, 0x62, 0x6f, 0x72, 0x77, 0x73, 0x65, 0x72, 0x20,
+  0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65,
+  0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f,
   0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65,
-  0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20,
-  0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74,
-  0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72,
-  0x73, 0x65, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f,
-  0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72,
-  0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41,
-  0x73, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67,
-  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c,
+  0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x63, 0x70, 0x70, 0x5f, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c,
+  0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65,
+  0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63,
+  0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73,
+  0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61,
+  0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74,
+  0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
   0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e,
-  0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61,
+  0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61,
   0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74,
   0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27,
-  0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74,
-  0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c,
-  0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65,
-  0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x74, 0x65,
-  0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
-  0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20,
-  0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69,
-  0x6e, 0x67, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72,
-  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20,
-  0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x7d, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c,
-  0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e,
-  0x61, 0x6c, 0x28, 0x7b, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x27,
-  0x27, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a,
-  0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20,
-  0x7b, 0x7d, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20,
-  0x7b, 0x7d, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x69, 0x6d,
-  0x70, 0x6f, 0x72, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x6c, 0x79,
-  0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74,
-  0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65,
-  0x72, 0x65, 0x20, 0x61, 0x72, 0x65, 0x20, 0x61, 0x6e, 0x79, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64,
-  0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x61, 0x72,
-  0x65, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, 0x20, 0x69, 0x6e, 0x20,
-  0x6f, 0x6e, 0x65, 0x20, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x6e, 0x20, 0x66, 0x6f, 0x72,
-  0x6d, 0x20, 0x6f, 0x66, 0x20, 0x7b, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x20, 0x22,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x64, 0x61, 0x74, 0x61,
-  0x22, 0x20, 0x7d, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x7b, 0x20, 0x22, 0x73,
-  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x74, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x22, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20,
-  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
-  0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6d, 0x70, 0x6f,
-  0x72, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x6d, 0x70,
-  0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x73, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73,
-  0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61,
-  0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27,
-  0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
-  0x20, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65,
-  0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20,
-  0x77, 0x65, 0x72, 0x65, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73,
-  0x66, 0x75, 0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65,
-  0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x50,
-  0x72, 0x6f, 0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61,
-  0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x69,
-  0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
-  0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72,
-  0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
-  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76,
-  0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
-  0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x73, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x2f, 0x2f, 0x6f, 0x76, 0x65, 0x72, 0x72, 0x69, 0x64, 0x65,
-  0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x2e, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x7b,
-  0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65,
-  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f,
-  0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44,
-  0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63,
-  0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65,
-  0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20,
+  0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53,
+  0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79,
+  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61,
+  0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65,
+  0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77,
+  0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72,
+  0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28,
+  0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67,
+  0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79,
+  0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67,
+  0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61,
+  0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65,
+  0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63,
+  0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65,
+  0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f,
+  0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d,
+  0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61,
+  0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65,
+  0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61,
+  0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
+  0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x73,
-  0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x73, 0x20, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x65, 0x64, 0x2e,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6e, 0x69,
-  0x74, 0x69, 0x61, 0x6c, 0x69, 0x7a, 0x69, 0x6e, 0x67, 0x20, 0x4c, 0x6f,
-  0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x61,
-  0x6e, 0x64, 0x20, 0x73, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65,
-  0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x20, 0x3d, 0x20, 0x7b, 0x20, 0x22, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c,
-  0x74, 0x22, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
-  0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61,
-  0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72,
-  0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73,
-  0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
-  0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72,
-  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e,
-  0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61,
-  0x75, 0x6c, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f,
-  0x67, 0x28, 0x27, 0x52, 0x65, 0x73, 0x65, 0x74, 0x69, 0x6e, 0x67, 0x20,
-  0x74, 0x68, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x74, 0x6f,
-  0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65,
-  0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
-  0x20, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65,
-  0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64,
-  0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55,
-  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
-  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5b, 0x27, 0x64, 0x65, 0x66, 0x61,
-  0x75, 0x6c, 0x74, 0x27, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69,
-  0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x74, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73,
-  0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
-  0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69,
-  0x6f, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61,
-  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
-  0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
+  0x6e, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65,
+  0x28, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
   0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20,
-  0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75,
-  0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65,
-  0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73,
-  0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28,
-  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72,
-  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c,
-  0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73,
-  0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f,
-  0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70,
-  0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28,
-  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
-  0x20, 0x67, 0x65, 0x74, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76,
-  0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64,
-  0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74,
-  0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67,
+  0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x52,
+  0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63,
+  0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65,
+  0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f,
+  0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72,
+  0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27,
+  0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73,
+  0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63,
+  0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73,
+  0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
+  0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67,
+  0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69,
+  0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63,
+  0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c,
+  0x28, 0x7b, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c,
+  0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b,
+  0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d,
+  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d,
+  0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f,
+  0x72, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73,
+  0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65, 0x72, 0x65,
+  0x20, 0x61, 0x72, 0x65, 0x20, 0x61, 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73,
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x61, 0x72, 0x65, 0x20,
+  0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e,
+  0x65, 0x20, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x69, 0x6e, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20,
+  0x6f, 0x66, 0x20, 0x7b, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20,
+  0x7d, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74,
+  0x74, 0x69, 0x6e, 0x67, 0x73, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x73, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
+  0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74,
+  0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72,
+  0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
   0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f,
   0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61,
   0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73,
   0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
-  0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, 0x61, 0x73, 0x74, 0x55,
-  0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29,
-  0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x77, 0x65,
+  0x72, 0x65, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x66, 0x75,
+  0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f,
+  0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65,
+  0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20,
+  0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67,
+  0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65,
+  0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65,
+  0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64,
+  0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x69, 0x6d,
+  0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x73, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x6f, 0x76, 0x65, 0x72, 0x72, 0x69, 0x64, 0x65, 0x20, 0x64,
+  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61,
+  0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64,
+  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x73,
+  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73,
+  0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74,
+  0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74,
+  0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28,
+  0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55,
+  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x73, 0x61, 0x76,
+  0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73,
+  0x20, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c,
+  0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6e, 0x69, 0x74, 0x69,
+  0x61, 0x6c, 0x69, 0x7a, 0x69, 0x6e, 0x67, 0x20, 0x4c, 0x6f, 0x63, 0x61,
+  0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x61, 0x6e, 0x64,
+  0x20, 0x73, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61,
+  0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61,
+  0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
+  0x20, 0x7b, 0x20, 0x22, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x22,
+  0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a,
+  0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65,
+  0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d,
+  0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72,
+  0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c,
+  0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75,
+  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52,
+  0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c,
+  0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28,
-  0x27, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e,
-  0x64, 0x2c, 0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67,
-  0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x27, 0x52, 0x65, 0x73, 0x65, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x68,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x64,
+  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55,
+  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d,
+  0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74,
+  0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x20, 0x3d, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65,
+  0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x5b, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c,
+  0x74, 0x27, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e,
+  0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x74, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e,
+  0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73,
+  0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65,
+  0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
+  0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74,
+  0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28,
   0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72,
   0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65,
-  0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e,
-  0x6c, 0x6f, 0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f,
-  0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x75, 0x73,
-  0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f,
-  0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c,
-  0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x77, 0x61, 0x73, 0x20, 0x66, 0x6f,
-  0x75, 0x6e, 0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x6f, 0x61, 0x64,
-  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c,
-  0x74, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20,
   0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75,
-  0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x70, 0x70,
-  0x6c, 0x79, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65,
-  0x20, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, 0x6c, 0x20, 0x64, 0x61,
-  0x74, 0x61, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63,
-  0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
-  0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28,
-  0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d,
-  0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73,
-  0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x65, 0x6c, 0x65,
-  0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f,
-  0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x54, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, 0x75, 0x74, 0x6f, 0x73,
-  0x61, 0x76, 0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63,
-  0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61,
-  0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75,
-  0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e,
-  0x27, 0x74, 0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, 0x74, 0x6f, 0x20, 0x73,
-  0x61, 0x76, 0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x65, 0x66,
-  0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20,
-  0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x6e, 0x65, 0x77,
-  0x20, 0x6f, 0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27,
-  0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x2d, 0x27, 0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f,
-  0x77, 0x28, 0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67,
-  0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
+  0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79,
+  0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67,
+  0x65, 0x74, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64,
+  0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73,
+  0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d,
+  0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61,
+  0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73,
+  0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72,
+  0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c,
+  0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65,
+  0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41,
+  0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c,
+  0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65,
+  0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c,
+  0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f,
+  0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61,
+  0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e,
+  0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61,
+  0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73,
+  0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x20, 0x77, 0x61, 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e,
+  0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73,
+  0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65,
+  0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
+  0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c,
+  0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79,
+  0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x69,
+  0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61,
+  0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75,
+  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41,
+  0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65,
+  0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f,
+  0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61,
+  0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c,
+  0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74,
+  0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20,
+  0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76,
+  0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65,
+  0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
+  0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
+  0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74,
+  0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76,
+  0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75,
+  0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c,
+  0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x63, 0x72,
+  0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f,
+  0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
   0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65,
-  0x27, 0x3a, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74,
+  0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73,
+  0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27,
+  0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28,
+  0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74,
+  0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a,
+  0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27,
+  0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e,
+  0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c,
+  0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e,
+  0x67, 0x20, 0x61, 0x73, 0x20, 0x27, 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77,
+  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65,
+  0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68,
+  0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x73,
+  0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67,
+  0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f,
+  0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65,
+  0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f,
+  0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20,
+  0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b,
+  0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41,
+  0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73,
+  0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f,
+  0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44,
+  0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63,
+  0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c,
+  0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73,
+  0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74,
   0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69,
   0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e,
   0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72,
   0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x53, 0x61, 0x76,
-  0x69, 0x6e, 0x67, 0x20, 0x61, 0x73, 0x20, 0x27, 0x20, 0x2b, 0x20, 0x6e,
-  0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61,
-  0x6d, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x69, 0x6e, 0x20,
-  0x74, 0x68, 0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65,
-  0x20, 0x73, 0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72,
-  0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46,
-  0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75,
-  0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x6e, 0x65, 0x77,
-  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e,
-  0x64, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, 0x74, 0x20, 0x62, 0x61,
-  0x63, 0x6b, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65,
-  0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61,
-  0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74,
-  0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61,
-  0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65,
-  0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a,
-  0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74,
-  0x27, 0x2c, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a,
-  0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65,
-  0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64,
-  0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69,
-  0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d,
-  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x43, 0x68,
-  0x65, 0x63, 0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61,
-  0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73,
-  0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73,
-  0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f,
-  0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75,
-  0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, 0x44, 0x3a, 0x20, 0x53,
-  0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73,
-  0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20,
-  0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65,
-  0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65,
-  0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61,
-  0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74,
-  0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c,
-  0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
-  0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c,
-  0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
-  0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75,
-  0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x20, 0x29,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
-  0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x3d,
-  0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29,
-  0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63,
-  0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20,
-  0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69,
-  0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28,
-  0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20,
-  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73,
-  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70,
-  0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
-  0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74,
-  0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74,
-  0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73,
-  0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74,
-  0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65,
-  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e,
-  0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20,
-  0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74,
-  0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
-  0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28,
-  0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65,
-  0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a,
-  0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f,
-  0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69,
-  0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
-  0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
-  0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20,
-  0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e,
-  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75,
-  0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67,
-  0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75,
-  0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
-  0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
-  0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
-  0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f,
-  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73,
-  0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28,
-  0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63,
-  0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73,
-  0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d,
-  0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x43, 0x68, 0x65, 0x63,
+  0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74,
+  0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20,
+  0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72,
+  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64,
+  0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f,
+  0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f,
+  0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e,
+  0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73,
+  0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73,
+  0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67,
+  0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61,
+  0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28,
+  0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28,
+  0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61,
+  0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67,
+  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20,
+  0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
+  0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e,
+  0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x68, 0x61, 0x73, 0x20, 0x74, 0x68, 0x65, 0x20, 0x75, 0x73, 0x65,
+  0x72, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x61, 0x20,
+  0x63, 0x68, 0x61, 0x74, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72,
+  0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74,
+  0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73,
   0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79,
-  0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69,
-  0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x28,
-  0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e,
-  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f,
-  0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20,
-  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x7d, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22,
-  0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c,
-  0x65, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65,
-  0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65,
-  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
-  0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20,
-  0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73,
-  0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22,
-  0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28,
-  0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29,
-  0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77,
-  0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
-  0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c,
-  0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b,
-  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a,
-  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75,
-  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74,
-  0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x64,
-  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
-  0x20, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x20, 0x6c, 0x65, 0x61, 0x64,
-  0x69, 0x6e, 0x67, 0x20, 0x77, 0x68, 0x69, 0x74, 0x65, 0x73, 0x70, 0x61,
-  0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67,
-  0x65, 0x20, 0x3d, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d,
-  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61,
-  0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x22,
-  0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65,
+  0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61,
+  0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74,
+  0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72,
+  0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e,
+  0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70,
+  0x6c, 0x61, 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20,
+  0x3d, 0x20, 0x28, 0x73, 0x74, 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72,
+  0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65,
+  0x74, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d,
+  0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
+  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72,
+  0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74,
+  0x72, 0x69, 0x6e, 0x67, 0x28, 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65,
+  0x70, 0x6c, 0x61, 0x63, 0x65, 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b,
+  0x5c, 0x7b, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f,
+  0x67, 0x2c, 0x20, 0x28, 0x5f, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79,
+  0x5d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73,
+  0x65, 0x72, 0x76, 0x65, 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61,
+  0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c,
+  0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61,
+  0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e,
+  0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62,
+  0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
+  0x72, 0x28, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73,
-  0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x63, 0x68,
-  0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65,
-  0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x5d, 0x5d, 0x29,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
-  0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67,
-  0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
-  0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27,
-  0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65,
-  0x73, 0x73, 0x61, 0x67, 0x65, 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73,
-  0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64,
-  0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69,
-  0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e,
-  0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c,
-  0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
-  0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f,
-  0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70,
-  0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e,
-  0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70,
-  0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72,
-  0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74,
-  0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c,
-  0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a,
+  0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74,
+  0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b,
+  0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20,
+  0x6d, 0x73, 0x67, 0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69,
+  0x73, 0x74, 0x6f, 0x72, 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69,
+  0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61,
+  0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74,
+  0x4d, 0x61, 0x70, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x28, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68,
+  0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65,
+  0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79,
+  0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74,
+  0x61, 0x29, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20,
+  0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29,
+  0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c,
+  0x73, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, 0x29,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68,
+  0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73,
+  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74,
+  0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
+  0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e,
+  0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28,
+  0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29,
+  0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22,
+  0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61,
+  0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68,
+  0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61,
+  0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e,
+  0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c,
+  0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d,
+  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67,
+  0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75,
+  0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68,
+  0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24,
+  0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72,
+  0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61,
+  0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f,
+  0x72, 0x79, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72,
+  0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74,
+  0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
-  0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20,
-  0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28,
-  0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43,
+  0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69,
+  0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20,
+  0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61,
+  0x67, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20,
+  0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29,
+  0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72,
+  0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65,
+  0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d,
+  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68,
+  0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63,
+  0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b,
+  0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20,
+  0x5b, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c,
+  0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e,
+  0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61,
+  0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x20, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d,
+  0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
+  0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74,
+  0x69, 0x6f, 0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49,
+  0x6e, 0x70, 0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69,
+  0x67, 0x6e, 0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74,
+  0x6f, 0x70, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e,
+  0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75,
+  0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f,
+  0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65,
   0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
   0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f,
   0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x22,
-  0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73,
-  0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x68,
-  0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, 0x20, 0x26,
-  0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x68, 0x69,
-  0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69,
-  0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
-  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72,
-  0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x3d, 0x24,
-  0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20,
-  0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20,
-  0x72, 0x6f, 0x77, 0x73, 0x3d, 0x32, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79,
-  0x70, 0x72, 0x65, 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65,
-  0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x20, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61,
-  0x67, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
-  0x3d, 0x24, 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65,
-  0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x7d, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68,
-  0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, 0x73,
-  0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x22,
-  0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63,
-  0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, 0x22,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79,
-  0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, 0x20,
-  0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x21,
-  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x7d, 0x20, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c,
-  0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75,
-  0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b,
-  0x3d, 0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73,
-  0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65,
-  0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70,
-  0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62,
-  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63,
-  0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52,
-  0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3d, 0x20, 0x28, 0x70,
-  0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d,
-  0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65,
-  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
-  0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20,
-  0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45,
-  0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
-  0x20, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x62,
-  0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x28, 0x69, 0x66, 0x20, 0x6e, 0x65,
-  0x65, 0x64, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x65,
-  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e,
-  0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x70,
-  0x61, 0x72, 0x65, 0x6e, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
-  0x20, 0x28, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20,
-  0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c,
-  0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, 0x70,
-  0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c,
-  0x54, 0x6f, 0x70, 0x20, 0x2b, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74,
-  0x2e, 0x6f, 0x66, 0x66, 0x73, 0x65, 0x74, 0x48, 0x65, 0x69, 0x67, 0x68,
-  0x74, 0x20, 0x2b, 0x20, 0x33, 0x30, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72,
-  0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f,
-  0x28, 0x30, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73,
-  0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x29,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73,
-  0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61,
-  0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73,
-  0x65, 0x72, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x29, 0x20, 0x3d, 0x3e,
-  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
-  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c,
-  0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x7d,
-  0x3e, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72,
-  0x29, 0x7d, 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e,
-  0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e,
-  0x69, 0x73, 0x68, 0x7d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x6d, 0x73, 0x67,
-  0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74,
-  0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d,
-  0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24,
-  0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24,
-  0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c,
-  0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69,
-  0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x60,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66,
-  0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72,
-  0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20,
-  0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65,
-  0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69,
-  0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65,
-  0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d,
-  0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65,
-  0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d,
-  0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
-  0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a,
-  0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74,
-  0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74,
-  0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c,
-  0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
-  0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61,
-  0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64,
-  0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74,
-  0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c,
-  0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65,
-  0x5d, 0x3a, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f,
-  0x72, 0x28, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74,
-  0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x29, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72,
-  0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68,
-  0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72,
-  0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27,
-  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d,
-  0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d,
-  0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d,
-  0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61,
-  0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65,
-  0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74,
-  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
-  0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53,
-  0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72,
-  0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
-  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20,
-  0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65,
-  0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
-  0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72,
-  0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d,
-  0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x28, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53,
-  0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64,
-  0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
-  0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c, 0x27, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63,
-  0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x28, 0x7b, 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20,
-  0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d,
-  0x3a, 0x20, 0x69, 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
-  0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74,
-  0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
-  0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61,
-  0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72,
-  0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72,
-  0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68,
-  0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60,
-  0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c,
-  0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73,
-  0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74,
+  0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62,
+  0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73,
+  0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73,
+  0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d,
+  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
   0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64,
-  0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20,
-  0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61,
-  0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a,
+  0x74, 0x20, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69,
+  0x74, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29,
+  0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e,
+  0x77, 0x68, 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33,
+  0x20, 0x26, 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73,
+  0x68, 0x69, 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62,
+  0x6d, 0x69, 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c,
+  0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66,
+  0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74,
+  0x3d, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65,
+  0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x4e, 0x61,
+  0x6d, 0x65, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74,
+  0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20,
+  0x22, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x22, 0x20, 0x3a, 0x20,
+  0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x6e, 0x69,
+  0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
-  0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c,
-  0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70,
-  0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64,
-  0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d,
-  0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20,
-  0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22,
-  0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65,
-  0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b,
-  0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20,
-  0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c,
-  0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61,
-  0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f,
-  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69,
-  0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65,
-  0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c,
-  0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68,
-  0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61,
-  0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e,
-  0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65,
-  0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
-  0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24,
-  0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d,
-  0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78,
-  0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61,
-  0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22,
-  0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75,
-  0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61,
-  0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a,
+  0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, 0x65, 0x73, 0x73, 0x3d,
+  0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69,
+  0x74, 0x73, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65,
+  0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20,
+  0x73, 0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e,
+  0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x32, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78,
+  0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d,
+  0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x22,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20,
+  0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74,
+  0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74,
+  0x79, 0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22,
+  0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, 0x2f,
+  0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74,
+  0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d,
+  0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61,
+  0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x21, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74,
+  0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20,
+  0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65,
+  0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f,
+  0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66,
+  0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f,
+  0x67, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63,
+  0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69,
+  0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66,
+  0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28,
+  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x63, 0x72, 0x6f, 0x6c,
+  0x6c, 0x20, 0x74, 0x6f, 0x20, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20,
+  0x28, 0x69, 0x66, 0x20, 0x6e, 0x65, 0x65, 0x64, 0x65, 0x64, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72,
+  0x72, 0x65, 0x6e, 0x74, 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x45,
+  0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x65,
+  0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74,
+  0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68,
+  0x74, 0x20, 0x3c, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e,
+  0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, 0x70, 0x20, 0x2b, 0x20,
+  0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x6f, 0x66, 0x66, 0x73, 0x65,
+  0x74, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x2b, 0x20, 0x33, 0x30,
+  0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63,
+  0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, 0x28, 0x30, 0x2c, 0x20, 0x70, 0x61,
+  0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48,
+  0x65, 0x69, 0x67, 0x68, 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d,
+  0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20,
+  0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, 0x64, 0x61,
+  0x74, 0x61, 0x5d, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65,
+  0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x41, 0x72, 0x72, 0x61,
+  0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61,
+  0x74, 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73,
+  0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x69, 0x73, 0x41, 0x72,
+  0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x68, 0x74,
+  0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62,
+  0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74,
+  0x61, 0x3d, 0x24, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e,
+  0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20,
+  0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61,
+  0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3f, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20,
+  0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29,
+  0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c,
+  0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d,
+  0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x4d, 0x61, 0x72,
+  0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x7d, 0x20, 0x74, 0x65,
+  0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x60,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
+  0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65,
+  0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c,
+  0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d,
+  0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24,
+  0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70,
+  0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
+  0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e,
+  0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72,
+  0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e,
+  0x65, 0x72, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x73, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68,
+  0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69,
+  0x6f, 0x6e, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d,
+  0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73,
+  0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73,
+  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b,
+  0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61,
+  0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75,
+  0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46,
+  0x6c, 0x6f, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c,
+  0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e,
+  0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65,
+  0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x49, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c,
+  0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e,
+  0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e,
+  0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46,
+  0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x29, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f,
+  0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f,
+  0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61,
+  0x6c, 0x28, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65,
+  0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53,
+  0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64,
+  0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e,
+  0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72,
+  0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20,
+  0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a,
+  0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61,
+  0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
+  0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68,
+  0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70,
+  0x61, 0x72, 0x73, 0x65, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61,
+  0x72, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65,
+  0x72, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53,
+  0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74,
+  0x65, 0x72, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a,
+  0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f,
+  0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c,
+  0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65,
+  0x28, 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20,
+  0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x2e, 0x2e, 0x2e, 0x61,
+  0x63, 0x63, 0x2c, 0x20, 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69,
+  0x6d, 0x28, 0x29, 0x5d, 0x3a, 0x20, 0x69, 0x7d, 0x29, 0x2c, 0x20, 0x7b,
+  0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76,
+  0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c,
+  0x20, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f,
+  0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d,
+  0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63,
+  0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65,
+  0x72, 0x74, 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20,
+  0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46,
+  0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e,
+  0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70,
+  0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, 0x3e,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66,
+  0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22,
+  0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65,
+  0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65,
+  0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69,
+  0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d,
+  0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24,
+  0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d,
+  0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75,
   0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
   0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60,
   0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75,
-  0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52,
-  0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d,
-  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66,
-  0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44,
-  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70,
-  0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
-  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f,
-  0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
-  0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72,
-  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27,
-  0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
-  0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62,
-  0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65,
-  0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72,
-  0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f,
-  0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b,
-  0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
-  0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74,
-  0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61,
-  0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74,
-  0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f,
-  0x73, 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
-  0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63,
-  0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61,
-  0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73,
-  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f,
-  0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49,
+  0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b,
+  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20,
+  0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e,
   0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74,
-  0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64,
-  0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65,
-  0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d,
+  0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b,
+  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65,
+  0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79,
+  0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69,
+  0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20,
+  0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22,
+  0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d,
+  0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61,
+  0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22,
+  0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e,
+  0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61,
+  0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d,
+  0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c,
+  0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28,
+  0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72,
+  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65,
+  0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e,
+  0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42,
+  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64,
+  0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20,
+  0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d,
+  0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64,
+  0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e,
+  0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74,
+  0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62,
+  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63,
+  0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70,
+  0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52,
+  0x65, 0x73, 0x65, 0x74, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20,
+  0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74,
+  0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66,
+  0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d,
+  0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65,
+  0x72, 0x79, 0x20, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65,
+  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61,
+  0x76, 0x65, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x2c, 0x20, 0x5b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68,
+  0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64,
+  0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72,
+  0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65,
+  0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66,
+  0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65,
+  0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f,
+  0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65,
+  0x61, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74,
+  0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d,
+  0x70, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24,
+  0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20,
+  0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70,
+  0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53,
+  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65,
+  0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20,
+  0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61,
+  0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65,
+  0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x70,
-  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70,
-  0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20,
+  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
+  0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22,
+  0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d,
+  0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20,
+  0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70,
+  0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d,
+  0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79,
-  0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61,
-  0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70,
-  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73,
-  0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
-  0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69,
-  0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66,
-  0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73,
-  0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69,
-  0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20,
-  0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55,
-  0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61,
-  0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75,
-  0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74,
-  0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72,
-  0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73,
-  0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e,
-  0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65,
-  0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a,
+  0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22,
+  0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70,
+  0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78,
+  0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61,
+  0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b,
+  0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69,
+  0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74,
+  0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c,
+  0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73,
+  0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74,
+  0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64,
+  0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61,
+  0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24,
+  0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d,
+  0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69,
+  0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74,
+  0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76,
   0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66,
-  0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74,
-  0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74,
-  0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e,
-  0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73,
-  0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68,
-  0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
-  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64,
-  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74,
-  0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a,
+  0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f,
+  0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c,
+  0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74,
+  0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22,
+  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61,
+  0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54,
+  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74,
+  0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d,
+  0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69,
+  0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74,
+  0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62,
-  0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74,
-  0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c,
+  0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66,
+  0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65,
+  0x22, 0x3e, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c,
   0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78,
-  0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65,
-  0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
-  0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f,
-  0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
-  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
-  0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68,
-  0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62,
-  0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61,
-  0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70,
-  0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22,
-  0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c,
-  0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22,
-  0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54,
-  0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f,
-  0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
-  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73,
-  0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69,
-  0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22,
-  0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72,
-  0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65,
-  0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61,
-  0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61,
-  0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68,
-  0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67,
-  0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20,
-  0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65,
-  0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24,
-  0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
-  0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20,
-  0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70,
-  0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72,
+  0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d,
+  0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c,
+  0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55,
+  0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63,
+  0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61,
+  0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f,
+  0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64,
+  0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79,
+  0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61,
+  0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64,
+  0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c,
+  0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20,
+  0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c,
+  0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70,
+  0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47,
+  0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63,
+  0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65,
+  0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74,
+  0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, 0x75,
+  0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63,
+  0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a,
+  0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61,
+  0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72,
+  0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d,
+  0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
-  0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22,
-  0x70, 0x72, 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20,
-  0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d,
-  0x22, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70,
-  0x31, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70,
-  0x33, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
-  0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d,
-  0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61,
-  0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f,
+  0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73,
+  0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74,
+  0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22,
   0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20,
-  0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e,
-  0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b,
-  0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53,
-  0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72,
-  0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53,
-  0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62,
-  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61,
-  0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b,
-  0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61,
-  0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63,
-  0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a,
-  0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20,
-  0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e,
-  0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
-  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65,
-  0x64, 0x69, 0x63, 0x74, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c,
-  0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61,
-  0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72,
-  0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a,
-  0x20, 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30,
-  0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74,
-  0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c,
-  0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c,
-  0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d,
-  0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x7d, 0x29, 0x7d, 0x0a,
+  0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64,
+  0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72,
+  0x65, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20,
+  0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d,
+  0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74,
+  0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e,
+  0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x7d, 0x29, 0x7d, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64,
-  0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65,
-  0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61,
-  0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c,
-  0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d,
-  0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d,
-  0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70,
-  0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65,
-  0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f,
-  0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24,
-  0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c,
-  0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69,
-  0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73,
-  0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a,
-  0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34,
-  0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e,
-  0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74,
-  0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74,
-  0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d, 0x29, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24,
-  0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c,
-  0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b,
-  0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20,
-  0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69,
-  0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a,
-  0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x7d,
-  0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69,
-  0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20,
-  0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c,
-  0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31,
-  0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30,
-  0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70,
-  0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30,
+  0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65,
+  0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20,
+  0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69,
+  0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65,
+  0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75,
+  0x72, 0x65, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30,
   0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20,
   0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69,
-  0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69,
-  0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79,
-  0x3e, 0x4d, 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e,
-  0x73, 0x3c, 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c,
-  0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c,
+  0x2e, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65,
+  0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46,
+  0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a,
+  0x20, 0x22, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72,
+  0x65, 0x70, 0x65, 0x61, 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e,
+  0x63, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e,
+  0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c,
+  0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65,
+  0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c,
+  0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c,
+  0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70,
+  0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d,
+  0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c,
+  0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43,
+  0x6f, 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f,
+  0x6b, 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e,
+  0x61, 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a,
+  0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20,
+  0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65,
+  0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22,
+  0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65,
+  0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d,
+  0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c,
   0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54,
-  0x46, 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20,
-  0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e,
-  0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66,
-  0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20,
-  0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a,
-  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
-  0x65, 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x7d, 0x29, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c,
-  0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54,
-  0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d,
+  0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e,
+  0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30,
+  0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e,
+  0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22,
+  0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f,
+  0x70, 0x5f, 0x6b, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f,
+  0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73,
+  0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61,
+  0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a,
+  0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20,
+  0x22, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65,
+  0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x7d, 0x29,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64,
+  0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d,
+  0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70,
+  0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61,
+  0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65,
+  0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f,
+  0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74,
+  0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c,
+  0x3a, 0x20, 0x22, 0x54, 0x46, 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d,
   0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e,
   0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a,
-  0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22,
-  0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31,
-  0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79,
-  0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c,
-  0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50,
-  0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61,
-  0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31,
-  0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30,
-  0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, 0x65,
-  0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74,
-  0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e,
-  0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70,
-  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
-  0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e,
-  0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b,
-  0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b,
-  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x46, 0x72, 0x65, 0x71,
-  0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74,
-  0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30,
-  0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20,
-  0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, 0x72, 0x65, 0x71, 0x75,
-  0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79,
-  0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30,
-  0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61,
-  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x66,
-  0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e,
-  0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69,
-  0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20,
-  0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74,
-  0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65,
-  0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c,
-  0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22,
-  0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d,
-  0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76,
-  0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22, 0x20, 0x63, 0x68, 0x65,
-  0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f,
-  0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f,
-  0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64,
-  0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74,
-  0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f,
-  0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e,
-  0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d,
-  0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65,
-  0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x63, 0x68,
-  0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61,
-  0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72,
-  0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20,
-  0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70,
-  0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e,
-  0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74,
-  0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c,
-  0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65,
-  0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d,
-  0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22,
-  0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63,
-  0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72,
-  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69,
-  0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d,
-  0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75,
-  0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49,
-  0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73,
-  0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65,
-  0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a,
+  0x20, 0x22, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74,
+  0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x7d,
+  0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74,
+  0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c,
+  0x3a, 0x20, 0x22, 0x54, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50,
+  0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c,
+  0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e,
+  0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61,
+  0x6c, 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20,
+  0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a,
+  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x7d,
+  0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74,
+  0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c,
+  0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20,
+  0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61,
+  0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a,
+  0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20,
+  0x22, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65,
+  0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70,
+  0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65,
+  0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65,
   0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22,
-  0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75,
-  0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30,
-  0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20,
-  0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73,
-  0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74,
-  0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61,
-  0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74,
-  0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64,
-  0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69,
-  0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c,
-  0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d,
-  0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d,
-  0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74,
-  0x5f, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a,
+  0x46, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65,
+  0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a,
+  0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30,
+  0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66,
+  0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e,
+  0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a,
   0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65,
   0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c,
-  0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f,
-  0x65, 0x74, 0x61, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65,
-  0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69,
-  0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61,
-  0x6e, 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20,
-  0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a,
-  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61,
-  0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20,
-  0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
-  0x73, 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28,
-  0x2f, 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b,
-  0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
-  0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67,
-  0x2c, 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20,
+  0x75, 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79,
+  0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c,
+  0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22,
+  0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62,
+  0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79,
+  0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e,
+  0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61,
+  0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22,
+  0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
+  0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20,
+  0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24,
+  0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20,
+  0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61,
+  0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61,
+  0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74,
+  0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20,
+  0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74,
+  0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31,
+  0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d,
+  0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d,
+  0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69,
+  0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c,
+  0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20,
+  0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22,
+  0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73,
+  0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22,
+  0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24,
+  0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d,
+  0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74,
+  0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d,
+  0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f,
+  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61,
+  0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65,
+  0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74,
+  0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20,
+  0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30,
+  0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d,
+  0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x22,
+  0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31,
+  0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72,
+  0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69,
+  0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x7d, 0x29,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46,
+  0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a,
+  0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x65,
+  0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e,
+  0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c,
+  0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f,
+  0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x73,
+  0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76,
+  0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73,
+  0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x7d, 0x29, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b,
+  0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x53, 0x68, 0x6f, 0x77,
+  0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69,
+  0x65, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30,
+  0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61,
+  0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73,
+  0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e,
+  0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20,
+  0x28, 0x70, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d,
+  0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28,
+  0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70,
+  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x67, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68,
+  0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a,
+  0x20, 0x70, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x60, 0x72, 0x67, 0x62, 0x61, 0x28,
+  0x24, 0x7b, 0x72, 0x7d, 0x2c, 0x24, 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c,
+  0x30, 0x2e, 0x33, 0x29, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65,
+  0x73, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29,
+  0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d,
+  0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20,
+  0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70,
+  0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73,
+  0x20, 0x7d, 0x20, 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f,
+  0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c,
+  0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70,
+  0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73,
+  0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20,
+  0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70,
+  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61,
+  0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e,
+  0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x4e, 0x6f, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x62, 0x79, 0x74, 0x65,
+  0x20, 0x70, 0x61, 0x69, 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70,
+  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61,
+  0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72,
+  0x74, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x62, 0x79, 0x74, 0x65,
+  0x3a, 0x20, 0x5c, 0x5c, 0x27, 0x29, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69,
+  0x74, 0x44, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70,
+  0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61,
+  0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70,
+  0x28, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f,
+  0x62, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72,
+  0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a,
+  0x20, 0x5b, 0x70, 0x72, 0x6f, 0x62, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50,
+  0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73,
+  0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c,
+  0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f,
+  0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30,
+  0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20,
+  0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70,
+  0x20, 0x3d, 0x3e, 0x20, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74,
+  0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f,
+  0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f,
+  0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20,
+  0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28,
+  0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20,
+  0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65,
+  0x6e, 0x74, 0x27, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76,
+  0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d,
+  0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c,
+  0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65,
+  0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e,
+  0x6d, 0x61, 0x70, 0x28, 0x28, 0x70, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65,
+  0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65,
+  0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x69, 0x74,
+  0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, 0x6f, 0x62, 0x3a, 0x20,
+  0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x7d, 0x60, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d,
+  0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x27, 0x30, 0x2e,
+  0x33, 0x65, 0x6d, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64,
+  0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x2e, 0x74, 0x6f, 0x6b,
+  0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43,
+  0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29,
+  0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72,
+  0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x2e,
+  0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x7d, 0x3a, 0x20, 0x3c, 0x2f,
+  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x4d, 0x61, 0x74, 0x68,
+  0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f,
+  0x62, 0x20, 0x2a, 0x20, 0x31, 0x30, 0x30, 0x29, 0x7d, 0x25, 0x3c, 0x2f,
+  0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x7d,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
+  0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c,
+  0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x20,
+  0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x62, 0x61,
+  0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f,
+  0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x7d, 0x7d,
+  0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c,
+  0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76,
+  0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e,
+  0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, 0x3f, 0x20, 0x68, 0x74, 0x6d, 0x6c,
+  0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x60, 0x20, 0x3a, 0x20, 0x6d,
+  0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, 0x6e, 0x73, 0x20, 0x6d, 0x61,
+  0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61,
+  0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63,
+  0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77,
+  0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x64, 0x20,
+  0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x74, 0x65, 0x78,
+  0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72,
+  0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x26, 0x2f, 0x67, 0x2c,
+  0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61,
-  0x63, 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67,
+  0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x6c,
   0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e,
-  0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24,
-  0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24,
-  0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63,
-  0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29,
-  0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74,
-  0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72,
-  0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28,
-  0x2f, 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67,
-  0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24,
-  0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65,
-  0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a,
-  0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d,
-  0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c,
-  0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f,
-  0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c,
-  0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28,
-  0x2f, 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c,
-  0x73, 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67,
-  0x2c, 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64,
-  0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c,
-  0x2f, 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3e,
+  0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, 0x74, 0x3b, 0x27, 0x29, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70,
+  0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, 0x7b, 0x31, 0x2c, 0x36,
+  0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x69, 0x6d, 0x2c,
+  0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x68, 0x33,
+  0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a,
+  0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x5c, 0x2a, 0x2f,
+  0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e,
+  0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27,
+  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72,
+  0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x5f, 0x28, 0x2e,
+  0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73,
+  0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74,
+  0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65,
-  0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c,
-  0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f,
-  0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65,
-  0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c,
-  0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74,
-  0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e,
-  0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49,
-  0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b,
-  0x20, 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20,
-  0x7d, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
-  0x74, 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72,
-  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20,
-  0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20,
-  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
-  0x21, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e,
-  0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x2f,
+  0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f,
+  0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f,
+  0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27,
+  0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27,
+  0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72,
+  0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x60, 0x60, 0x2e,
+  0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, 0x5c, 0x53, 0x5d, 0x2a,
+  0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x70,
+  0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c,
+  0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, 0x70, 0x72, 0x65, 0x3e,
+  0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
+  0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x28, 0x2e,
+  0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x63, 0x6f,
+  0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e,
+  0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e,
+  0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x6e, 0x2f,
+  0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e,
+  0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73,
+  0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75,
+  0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48,
+  0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x5f, 0x5f, 0x68, 0x74,
+  0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, 0x7d, 0x20, 0x2f, 0x3e,
+  0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x6f, 0x64,
+  0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+  0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60,
+  0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61,
+  0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74,
+  0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x74, 0x6f,
+  0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78,
+  0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, 0x20, 0x70, 0x65, 0x72, 0x20,
+  0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61,
+  0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f,
+  0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x2e, 0x74,
+  0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, 0x29, 0x7d, 0x20, 0x74,
+  0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x73, 0x65,
+  0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c,
+  0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x69, 0x6d,
+  0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28,
+  0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65,
+  0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20,
+  0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28,
+  0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27,
+  0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78,
+  0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e,
+  0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66,
+  0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f,
+  0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65,
+  0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20,
+  0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65,
+  0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75,
+  0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, 0x62, 0x75, 0x74, 0x74,
+  0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e,
+  0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, 0x6e, 0x64, 0x69, 0x6e,
+  0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, 0x65, 0x63, 0x74, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61,
+  0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a,
+  0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x62, 0x6f, 0x74,
+  0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77,
+  0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, 0x7d, 0x70, 0x78, 0x60,
+  0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72,
+  0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x20, 0x2b, 0x20, 0x77,
+  0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c,
+  0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c,
+  0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e,
+  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65,
+  0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65,
+  0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52,
+  0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26,
+  0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65,
+  0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f,
+  0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74,
+  0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x20, 0x26, 0x26, 0x20,
+  0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63,
+  0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61,
+  0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61,
+  0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65,
+  0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61,
+  0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66,
+  0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63,
+  0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65,
+  0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27,
+  0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20,
+  0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f,
+  0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
-  0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e,
-  0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65,
+  0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65,
+  0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27,
+  0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20,
+  0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f,
+  0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20,
+  0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x73, 0x74, 0x79, 0x6c,
+  0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x73, 0x74,
+  0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x62,
+  0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x7d, 0x20, 0x6f, 0x6e,
+  0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x74, 0x6f, 0x67, 0x67,
+  0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x3e, 0x24,
+  0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64,
+  0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x69, 0x73,
+  0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x26,
+  0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72,
+  0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x3d, 0x22, 0x23,
+  0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69,
+  0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x70, 0x6f,
+  0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, 0x6f, 0x70, 0x6f, 0x76,
+  0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73,
+  0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e,
+  0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66,
+  0x74, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e,
+  0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70,
+  0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72,
+  0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61,
+  0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x3a, 0x20, 0x70,
+  0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c,
+  0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69,
+  0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x64, 0x65, 0x76,
+  0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63,
+  0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2f, 0x62, 0x6c, 0x6f,
+  0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, 0x2f, 0x73, 0x72, 0x63,
+  0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74,
+  0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x20,
+  0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x66,
+  0x20, 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, 0x61, 0x6e, 0x74, 0x73,
+  0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x67, 0x69,
+  0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, 0x73, 0x65, 0x6c, 0x65,
+  0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c,
+  0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d,
+  0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74,
+  0x44, 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x70, 0x72,
+  0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x6c, 0x65, 0x74, 0x20,
+  0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x69, 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d,
+  0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72,
+  0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65,
+  0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e,
+  0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64,
+  0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69,
+  0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x74,
+  0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72,
+  0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73,
+  0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72,
+  0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69,
+  0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65,
+  0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70,
+  0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d,
+  0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65,
+  0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x66, 0x61,
+  0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75,
+  0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66,
+  0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74,
+  0x65, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65,
+  0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e,
+  0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65,
+  0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e,
+  0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x43, 0x68,
+  0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d,
+  0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x6e,
+  0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20,
+  0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74,
+  0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, 0x64, 0x6f, 0x63, 0x75,
+  0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65,
+  0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29,
+  0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28,
+  0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d,
+  0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75,
+  0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, 0x6e, 0x20, 0x75, 0x70,
+  0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x69, 0x66,
+  0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x62, 0x61, 0x73, 0x65,
+  0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70,
+  0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74,
+  0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e,
+  0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e,
+  0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69,
+  0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e,
+  0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73,
+  0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
+  0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d,
+  0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c,
+  0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72,
+  0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x2c, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69,
+  0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69,
+  0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69,
+  0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x74,
+  0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e,
+  0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20,
+  0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d,
+  0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72,
+  0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74,
+  0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x68, 0x6f, 0x77,
+  0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f,
+  0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20,
+  0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50,
+  0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x2c, 0x20,
+  0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74,
+  0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x29,
+  0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x69, 0x67,
+  0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x63, 0x6f, 0x6d, 0x70,
+  0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72,
+  0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, 0x74, 0x73, 0x20, 0x66,
+  0x69, 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x20, 0x69,
+  0x66, 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, 0x73, 0x74, 0x73, 0x2e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x64,
+  0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x64, 0x69, 0x74,
+  0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72,
+  0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, 0x79, 0x2e, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72,
+  0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x20, 0x65, 0x78, 0x74,
+  0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65,
+  0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67,
+  0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, 0x6f, 0x6e, 0x74, 0x65,
+  0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64,
+  0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65,
+  0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x68,
+  0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75,
+  0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66,
+  0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28,
+  0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20,
   0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68,
   0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d,
-  0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65,
-  0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70,
-  0x65, 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e,
-  0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73,
-  0x20, 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20,
-  0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73,
-  0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69,
-  0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63,
-  0x6f, 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28,
-  0x32, 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70,
-  0x65, 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e,
-  0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28, 0x70, 0x72,
-  0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d,
-  0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74,
-  0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65,
-  0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
-  0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65,
-  0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69,
-  0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72,
-  0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20,
-  0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, 0x43, 0x6f,
-  0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, 0x2f, 0x3e,
-  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c,
-  0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69,
-  0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, 0x74, 0x65,
-  0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67,
-  0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73,
-  0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74,
-  0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f,
-  0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f,
-  0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70,
+  0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72,
   0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64,
-  0x20, 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d,
-  0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74,
-  0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72,
-  0x67, 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
-  0x63, 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63,
-  0x70, 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c,
-  0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70,
-  0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22,
-  0x3e, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e,
-  0x2e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
-  0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72,
-  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
-  0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60,
-  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
-  0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70,
-  0x70, 0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74,
-  0x2e, 0x62, 0x6f, 0x64, 0x79, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f,
-  0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65,
-  0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a,
-  0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68,
-  0x74, 0x6d, 0x6c, 0x3e, 0x0a
+  0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65, 0x61,
+  0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69, 0x64,
+  0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74,
+  0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, 0x43,
+  0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6e,
+  0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, 0x2f, 0x3e, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f,
+  0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f,
+  0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, 0x74, 0x65, 0x22,
+  0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x49, 0x6e, 0x70, 0x75, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65,
+  0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65,
+  0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64,
+  0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+  0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20,
+  0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22,
+  0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68,
+  0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67,
+  0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63,
+  0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70,
+  0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61,
+  0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73,
+  0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e,
+  0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e,
+  0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64,
+  0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70,
+  0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e,
+  0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f,
+  0x72, 0x28, 0x27, 0x23, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65,
+  0x72, 0x27, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63,
+  0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64,
+  0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20,
+  0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e,
+  0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69,
+  0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64,
+  0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f,
+  0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e,
+  0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, 0x6d, 0x6c, 0x3e, 0x0a
 };
-unsigned int index_html_len = 20333;
+unsigned int index_html_len = 28018;
diff --git a/examples/server/public/index.html b/examples/server/public/index.html
index 5eedb0b28..1bf2a8b3a 100644
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@@ -102,6 +102,17 @@
       padding: 0.5em;
     }
 
+    .prob-set {
+      padding: 0.3em;
+      border-bottom: 1px solid #ccc;
+    }
+
+    .popover-content {
+      position: absolute;
+      background-color: white;
+      padding: 0.2em;
+      box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);
+    }
 
     textarea {
       padding: 5px;
@@ -133,11 +144,39 @@
       font-size: 80%;
       color: #888;
     }
+
+
+    @keyframes loading-bg-wipe {
+      0% {
+        background-position: 0%;
+      }
+      100% {
+        background-position: 100%;
+      }
+    }
+
+    .loading {
+      --loading-color-1: #eeeeee00;
+      --loading-color-2: #eeeeeeff;
+      background-size: 50% 100%;
+      background-image: linear-gradient(90deg, var(--loading-color-1), var(--loading-color-2), var(--loading-color-1));
+      animation: loading-bg-wipe 2s linear infinite;
+    }
+
+    @media (prefers-color-scheme: dark) {
+      .loading {
+        --loading-color-1: #22222200;
+        --loading-color-2: #222222ff;
+      }
+      .popover-content {
+        background-color: black;
+      }
+    }
   </style>
 
   <script type="module">
     import {
-      html, h, signal, effect, computed, render, useSignal, useEffect, useRef
+      html, h, signal, effect, computed, render, useSignal, useEffect, useRef, Component
     } from '/index.js';
 
     import { llama } from '/completion.js';
@@ -168,6 +207,7 @@
       mirostat_tau: 5, // target entropy
       mirostat_eta: 0.1, // learning rate
       grammar: '',
+      n_probs: 0, // no completion_probabilities
     })
 
     /* START: Support for storing prompt templates and parameters in borwser LocalStorage */
@@ -303,7 +343,10 @@
     const llamaStats = signal(null)
     const controller = signal(null)
 
-    const generating = computed(() => controller.value == null )
+    // currently generating a completion?
+    const generating = computed(() => controller.value != null)
+
+    // has the user started a chat?
     const chatStarted = computed(() => session.value.transcript.length > 0)
 
     const transcriptUpdate = (transcript) => {
@@ -334,10 +377,21 @@
 
       const prompt = template(session.value.template, {
         message: msg,
-        history: session.value.transcript.flatMap(([name, message]) => template(session.value.historyTemplate, {name, message})).join("\n"),
+        history: session.value.transcript.flatMap(
+          ([name, data]) =>
+            template(
+              session.value.historyTemplate,
+              {
+                name,
+                message: Array.isArray(data) ?
+                  data.map(msg => msg.content).join('').replace(/^\s/, '') :
+                  data,
+              }
+            )
+        ).join("\n"),
       });
 
-      let currentMessage = '';
+      const currentMessages = [];
       const history = session.value.transcript
 
       const llamaParams = {
@@ -347,15 +401,19 @@
 
       for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
         const data = chunk.data;
-        currentMessage += data.content;
-
-        // remove leading whitespace
-        currentMessage = currentMessage.replace(/^\s+/, "")
-
-        transcriptUpdate([...history, ["{{char}}", currentMessage]])
 
         if (data.stop) {
-          console.log("Completion finished: '", currentMessage, "', summary: ", data);
+          while (
+            currentMessages.length > 0 &&
+            currentMessages[currentMessages.length - 1].content.match(/\n$/) != null
+          ) {
+            currentMessages.pop();
+          }
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
+          console.log("Completion finished: '", currentMessages.map(msg => msg.content).join(''), "', summary: ", data);
+        } else {
+          currentMessages.push(data);
+          transcriptUpdate([...history, ["{{char}}", currentMessages]])
         }
 
         if (data.timings) {
@@ -397,11 +455,19 @@
       return html`
         <form onsubmit=${submit}>
           <div>
-            <textarea type="text" rows=2 onkeypress=${enterSubmits} value="${message}" oninput=${(e) => message.value = e.target.value} placeholder="Say something..."/>
+            <textarea
+               className=${generating.value ? "loading" : null}
+               oninput=${(e) => message.value = e.target.value}
+               onkeypress=${enterSubmits}
+               placeholder="Say something..."
+               rows=2
+               type="text"
+               value="${message}"
+            />
           </div>
           <div class="right">
-            <button type="submit" disabled=${!generating.value} >Send</button>
-            <button onclick=${stop} disabled=${generating}>Stop</button>
+            <button type="submit" disabled=${generating.value}>Send</button>
+            <button onclick=${stop} disabled=${!generating.value}>Stop</button>
             <button onclick=${reset}>Reset</button>
           </div>
         </form>
@@ -420,8 +486,18 @@
         }
       }, [messages])
 
-      const chatLine = ([user, msg]) => {
-        return html`<p key=${msg}><strong>${template(user)}:</strong> <${Markdownish} text=${template(msg)} /></p>`
+      const chatLine = ([user, data], index) => {
+        let message
+        const isArrayMessage = Array.isArray(data)
+        if (params.value.n_probs > 0 && isArrayMessage) {
+          message = html`<${Probabilities} data=${data} />`
+        } else {
+          const text = isArrayMessage ?
+            data.map(msg => msg.content).join('').replace(/^\s+/, '') :
+            data;
+          message = html`<${Markdownish} text=${template(text)} />`
+        }
+        return html`<p key=${index}><strong>${template(user)}:</strong> ${message}</p>`
       };
 
       return html`
@@ -568,10 +644,71 @@
               ${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})}
               ${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})}
             </fieldset>
+            <fieldset>
+              ${IntField({label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs})}
+            </fieldset>
           </details>
         </form>
       `
     }
+
+    const probColor = (p) => {
+      const r = Math.floor(192 * (1 - p));
+      const g = Math.floor(192 * p);
+      return `rgba(${r},${g},0,0.3)`;
+    }
+
+    const Probabilities = (params) => {
+      return params.data.map(msg => {
+        const { completion_probabilities } = msg;
+        if (
+          !completion_probabilities ||
+          completion_probabilities.length === 0
+        ) return msg.content
+
+        if (completion_probabilities.length > 1) {
+          // Not for byte pair
+          if (completion_probabilities[0].content.startsWith('byte: \\')) return msg.content
+
+          const splitData = completion_probabilities.map(prob => ({
+            content: prob.content,
+            completion_probabilities: [prob]
+          }))
+          return html`<${Probabilities} data=${splitData} />`
+        }
+
+        const { probs, content } = completion_probabilities[0]
+        const found = probs.find(p => p.tok_str === msg.content)
+        const pColor = found ? probColor(found.prob) : 'transparent'
+
+        const popoverChildren = html`
+          <div class="prob-set">
+            ${probs.map((p, index) => {
+              return html`
+                <div
+                  key=${index}
+                  title=${`prob: ${p.prob}`}
+                  style=${{
+                    padding: '0.3em',
+                    backgroundColor: p.tok_str === content ? probColor(p.prob) : 'transparent'
+                  }}
+                >
+                  <span>${p.tok_str}: </span>
+                  <span>${Math.floor(p.prob * 100)}%</span>
+                </div>
+              `
+            })}
+          </div>
+        `
+
+        return html`
+          <${Popover} style=${{ backgroundColor: pColor }} popoverChildren=${popoverChildren}>
+            ${msg.content.match(/\n/gim) ? html`<br />` : msg.content}
+          </>
+        `
+      });
+    }
+
     // poor mans markdown replacement
     const Markdownish = (params) => {
       const md = params.text
@@ -600,10 +737,121 @@
       `
     }
 
+    // simple popover impl
+    const Popover = (props) => {
+      const isOpen = useSignal(false);
+      const position = useSignal({ top: '0px', left: '0px' });
+      const buttonRef = useRef(null);
+      const popoverRef = useRef(null);
+
+      const togglePopover = () => {
+        if (buttonRef.current) {
+          const rect = buttonRef.current.getBoundingClientRect();
+          position.value = {
+            top: `${rect.bottom + window.scrollY}px`,
+            left: `${rect.left + window.scrollX}px`,
+          };
+        }
+        isOpen.value = !isOpen.value;
+      };
+
+      const handleClickOutside = (event) => {
+        if (popoverRef.current && !popoverRef.current.contains(event.target) && !buttonRef.current.contains(event.target)) {
+          isOpen.value = false;
+        }
+      };
+
+      useEffect(() => {
+        document.addEventListener('mousedown', handleClickOutside);
+        return () => {
+          document.removeEventListener('mousedown', handleClickOutside);
+        };
+      }, []);
+
+      return html`
+        <span style=${props.style} ref=${buttonRef} onClick=${togglePopover}>${props.children}</span>
+        ${isOpen.value && html`
+          <${Portal} into="#portal">
+            <div
+              ref=${popoverRef}
+              class="popover-content"
+              style=${{
+                top: position.value.top,
+                left: position.value.left,
+              }}
+            >
+              ${props.popoverChildren}
+            </div>
+          </${Portal}>
+        `}
+      `;
+    };
+
+    // Source: preact-portal (https://github.com/developit/preact-portal/blob/master/src/preact-portal.js)
+    /** Redirect rendering of descendants into the given CSS selector */
+    class Portal extends Component {
+      componentDidUpdate(props) {
+        for (let i in props) {
+          if (props[i] !== this.props[i]) {
+            return setTimeout(this.renderLayer);
+          }
+        }
+      }
+
+      componentDidMount() {
+        this.isMounted = true;
+        this.renderLayer = this.renderLayer.bind(this);
+        this.renderLayer();
+      }
+
+      componentWillUnmount() {
+        this.renderLayer(false);
+        this.isMounted = false;
+        if (this.remote && this.remote.parentNode) this.remote.parentNode.removeChild(this.remote);
+      }
+
+      findNode(node) {
+        return typeof node === 'string' ? document.querySelector(node) : node;
+      }
+
+      renderLayer(show = true) {
+        if (!this.isMounted) return;
+
+        // clean up old node if moving bases:
+        if (this.props.into !== this.intoPointer) {
+          this.intoPointer = this.props.into;
+          if (this.into && this.remote) {
+            this.remote = render(html`<${PortalProxy} />`, this.into, this.remote);
+          }
+          this.into = this.findNode(this.props.into);
+        }
+
+        this.remote = render(html`
+          <${PortalProxy} context=${this.context}>
+            ${show && this.props.children || null}
+          </${PortalProxy}>
+        `, this.into, this.remote);
+      }
+
+      render() {
+        return null;
+      }
+    }
+    // high-order component that renders its first child if it exists.
+    // used as a conditional rendering proxy.
+    class PortalProxy extends Component {
+      getChildContext() {
+        return this.props.context;
+      }
+      render({ children }) {
+        return children || null;
+      }
+    }
+
     function App(props) {
 
       return html`
-        <div id="container">
+        <div>
           <header>
             <h1>llama.cpp</h1>
           </header>
@@ -624,11 +872,13 @@
       `;
     }
 
-    render(h(App), document.body);
+    render(h(App), document.querySelector('#container'));
   </script>
 </head>
 
 <body>
+  <div id="container"></div>
+  <div id="portal"></div>
 </body>
 
 </html>
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 1e6d10c1d..c53a64867 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -17,6 +17,8 @@
 #include "completion.js.hpp"
 #include "json-schema-to-grammar.mjs.hpp"
 
+#include <cstddef>
+
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
 #endif
@@ -94,7 +96,7 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
     std::string ret;
     for (; begin != end; ++begin)
     {
-        ret += llama_token_to_str(ctx, *begin);
+        ret += llama_token_to_piece(ctx, *begin);
     }
     return ret;
 }
@@ -116,16 +118,17 @@ static void server_log(const char *level, const char *function, int line,
     }
 
     const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
-    fprintf(stdout, "%.*s\n", (int)str.size(), str.data());
+    printf("%.*s\n", (int)str.size(), str.data());
     fflush(stdout);
 }
 
 // format incomplete utf-8 multibyte character for output
 static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
 {
-    std::string out = token == -1 ? "" : llama_token_to_str(ctx, token);
-    // if first bit is 1, meaning it's a partial character
-    if (out.size() > 0 && (out[0] & 0x80) == 0x80)
+    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
+    // if the size is 1 and first bit is 1, meaning it's a partial character
+    //   (size > 1 meaning it's already a known token)
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
     {
         std::stringstream ss;
         ss << std::hex << (out[0] & 0xff);
@@ -136,7 +139,7 @@ static std::string tokens_to_output_formatted_string(const llama_context *ctx, c
 }
 
 // convert a vector of completion_token_output to json
-static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> probs)
+static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> & probs)
 {
     json out = json::array();
     for (const auto &prob : probs)
@@ -197,6 +200,7 @@ struct llama_server_context
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
     gpt_params params;
+    int n_ctx;
 
     grammar_parser::parse_state parsed_grammar;
     llama_grammar *grammar = nullptr;
@@ -236,7 +240,7 @@ struct llama_server_context
         num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
-        generated_text.reserve(params.n_ctx);
+        generated_text.reserve(n_ctx);
         generated_token_probs.clear();
         truncated = false;
         stopped_eos = false;
@@ -262,13 +266,13 @@ struct llama_server_context
             LOG_ERROR("unable to load model", {{"model", params_.model}});
             return false;
         }
-
-        last_n_tokens.resize(params.n_ctx);
+        n_ctx = llama_n_ctx(ctx);
+        last_n_tokens.resize(n_ctx);
         std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
         return true;
     }
 
-    std::vector<llama_token> tokenize(json json_prompt, bool add_bos)
+    std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
     {
         // If `add_bos` is true, we only add BOS, when json_prompt is a string,
         // or the first element of the json_prompt array is a string.
@@ -285,7 +289,6 @@ struct llama_server_context
                     std::vector<llama_token> p;
                     if (first)
                     {
-                        s.insert(0, 1, ' '); // add a space if it's the first
                         p = ::llama_tokenize(ctx, s, add_bos);
                         first = false;
                     }
@@ -308,7 +311,6 @@ struct llama_server_context
         else
         {
             auto s = json_prompt.template get<std::string>();
-            s.insert(0, 1, ' '); // always add a first space
             prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
         }
 
@@ -340,9 +342,15 @@ struct llama_server_context
         return true;
     }
 
-    void loadPrompt()
+    void loadInfill()
     {
-        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+        auto prefix_tokens = tokenize(params.input_prefix, true);  // always add BOS
+        auto suffix_tokens = tokenize(params.input_suffix, true);  // always add BOS
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+        prefix_tokens.push_back(llama_token_middle(ctx));
+        auto prompt_tokens = prefix_tokens;
 
         num_prompt_tokens = prompt_tokens.size();
 
@@ -355,6 +363,8 @@ struct llama_server_context
         // if input prompt is too big, truncate like normal
         if (num_prompt_tokens >= (size_t)params.n_ctx)
         {
+            printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
+            // todo we probably want to cut from both sides
             const int n_left = (params.n_ctx - params.n_keep) / 2;
             std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
             const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
@@ -380,6 +390,66 @@ struct llama_server_context
 
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            printf("we have to evaluate at least 1 token to generate logits\n");
+            n_past--;
+        }
+
+        LOG_VERBOSE("prompt ingested", {
+                                           {"n_past", n_past},
+                                           {"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
+                                           {"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
+                                       });
+
+        has_next_token = true;
+    }
+    void loadPrompt()
+    {
+        auto prompt_tokens = tokenize(prompt, true);  // always add BOS
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t)n_ctx)
+        {
+            const int n_left = (n_ctx - params.n_keep) / 2;
+            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
+            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
+            std::copy(prompt_tokens.end() - n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+
+            LOG_VERBOSE("input truncated", {
+                                               {"n_ctx", n_ctx},
+                                               {"n_keep", params.n_keep},
+                                               {"n_left", n_left},
+                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                                           });
+
+            truncated = true;
+            prompt_tokens = new_tokens;
+        }
+        else
+        {
+            const size_t ps = num_prompt_tokens;
+            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
+            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         embd = prompt_tokens;
         if (n_past == num_prompt_tokens)
         {
@@ -408,37 +478,47 @@ struct llama_server_context
         completion_token_output result;
         result.tok = -1;
 
-        if (embd.size() >= (size_t)params.n_ctx)
+        if (embd.size() >= (size_t)n_ctx)
         {
-            // Reset context
-            const int n_left = (params.n_ctx - params.n_keep) / 2;
+            // Shift context
+
+            const int n_left    = n_past - params.n_keep - 1;
+            const int n_discard = n_left/2;
+
+            llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+            llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+            for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++)
+            {
+                embd[i - n_discard] = embd[i];
+            }
+            embd.resize(embd.size() - n_discard);
+
+            n_past -= n_discard;
 
-            std::vector<llama_token> new_tokens(embd.begin(), embd.begin() + params.n_keep);
-            new_tokens.insert(new_tokens.end(), embd.end() - n_left, embd.end());
-            embd = new_tokens;
-            n_past = params.n_keep;
             truncated = true;
             LOG_VERBOSE("input truncated", {
-                                               {"n_ctx", params.n_ctx},
+                                               {"n_ctx", n_ctx},
                                                {"n_keep", params.n_keep},
                                                {"n_left", n_left},
-                                               {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
                                            });
         }
 
+        bool tg = true;
         while (n_past < embd.size())
         {
             int n_eval = (int)embd.size() - n_past;
+            tg = n_eval == 1;
             if (n_eval > params.n_batch)
             {
                 n_eval = params.n_batch;
             }
-            if (llama_eval(ctx, &embd[n_past], n_eval, n_past, params.n_threads))
+
+            if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0)))
             {
                 LOG_ERROR("failed to eval", {
                                                 {"n_eval", n_eval},
                                                 {"n_past", n_past},
-                                                {"n_threads", params.n_threads},
                                                 {"embd", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
                                             });
                 has_next_token = false;
@@ -454,98 +534,20 @@ struct llama_server_context
             return result;
         }
 
-        // out of user input, sample next token
-        const float temp = params.temp;
-        const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k;
-        const float top_p = params.top_p;
-        const float tfs_z = params.tfs_z;
-        const float typical_p = params.typical_p;
-        const int32_t repeat_last_n = params.repeat_last_n < 0 ? params.n_ctx : params.repeat_last_n;
-        const float repeat_penalty = params.repeat_penalty;
-        const float alpha_presence = params.presence_penalty;
-        const float alpha_frequency = params.frequency_penalty;
-        const int mirostat = params.mirostat;
-        const float mirostat_tau = params.mirostat_tau;
-        const float mirostat_eta = params.mirostat_eta;
-        const bool penalize_nl = params.penalize_nl;
-        const int32_t n_probs = params.n_probs;
-
         {
-            auto *logits = llama_get_logits(ctx);
-            auto n_vocab = llama_n_vocab(ctx);
-
-            // Apply params.logit_bias map
-            for (const auto &it : params.logit_bias)
-            {
-                logits[it.first] += it.second;
-            }
-
+            // out of user input, sample next token
             std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++)
+            candidates.reserve(llama_n_vocab(model));
+
+            result.tok = llama_sample_token(ctx, NULL, grammar, params, last_n_tokens, candidates);
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            const int32_t n_probs = params.n_probs;
+            if (params.temp <= 0 && n_probs > 0)
             {
-                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-            }
-
-            llama_token_data_array candidates_p = {candidates.data(), candidates.size(), false};
-
-            // Apply penalties
-            float nl_logit = logits[llama_token_nl(ctx)];
-            auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), params.n_ctx);
-            llama_sample_repetition_penalty(ctx, &candidates_p,
-                                            last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                            last_n_repeat, repeat_penalty);
-            llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
-                                                          last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
-                                                          last_n_repeat, alpha_frequency, alpha_presence);
-            if (!penalize_nl)
-            {
-                logits[llama_token_nl(ctx)] = nl_logit;
-            }
-
-            if (grammar != nullptr) {
-                llama_sample_grammar(ctx, &candidates_p, grammar);
-            }
-
-            if (temp <= 0)
-            {
-                // Greedy sampling
-                result.tok = llama_sample_token_greedy(ctx, &candidates_p);
-                if (n_probs > 0)
-                {
-                    llama_sample_softmax(ctx, &candidates_p);
-                }
-            }
-            else
-            {
-                if (mirostat == 1)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    const int mirostat_m = 100;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-                }
-                else if (mirostat == 2)
-                {
-                    static float mirostat_mu = 2.0f * mirostat_tau;
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-                }
-                else
-                {
-                    // Temperature sampling
-                    size_t min_keep = std::max(1, n_probs);
-                    llama_sample_top_k(ctx, &candidates_p, top_k, min_keep);
-                    llama_sample_tail_free(ctx, &candidates_p, tfs_z, min_keep);
-                    llama_sample_typical(ctx, &candidates_p, typical_p, min_keep);
-                    llama_sample_top_p(ctx, &candidates_p, top_p, min_keep);
-                    llama_sample_temperature(ctx, &candidates_p, temp);
-                    result.tok = llama_sample_token(ctx, &candidates_p);
-                }
-            }
-
-            if (grammar != nullptr) {
-                llama_grammar_accept_token(ctx, grammar, result.tok);
+                // For llama_sample_token_greedy we need to sort candidates
+                llama_sample_softmax(ctx, &candidates_p);
             }
 
             for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
@@ -555,7 +557,9 @@ struct llama_server_context
 
             last_n_tokens.erase(last_n_tokens.begin());
             last_n_tokens.push_back(result.tok);
-            num_tokens_predicted++;
+            if (tg) {
+                num_tokens_predicted++;
+            }
         }
 
         // add it to the context
@@ -565,7 +569,7 @@ struct llama_server_context
 
         if (!embd.empty() && embd.back() == llama_token_eos(ctx))
         {
-            // stopping_word = llama_token_to_str(ctx, embd.back());
+            // stopping_word = llama_token_to_piece(ctx, embd.back());
             has_next_token = false;
             stopped_eos = true;
             LOG_VERBOSE("eos token found", {});
@@ -610,9 +614,9 @@ struct llama_server_context
 
     completion_token_output doCompletion()
     {
-        const completion_token_output token_with_probs = nextToken();
+        auto token_with_probs = nextToken();
 
-        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(ctx, token_with_probs.tok);
+        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
         generated_text += token_text;
 
         if (params.n_probs > 0)
@@ -676,7 +680,7 @@ struct llama_server_context
 
     std::vector<float> getEmbedding()
     {
-        static const int n_embd = llama_n_embd(ctx);
+        static const int n_embd = llama_n_embd(model);
         if (!params.embedding)
         {
             LOG_WARNING("embedding disabled", {
@@ -693,50 +697,49 @@ struct llama_server_context
 static void server_print_usage(const char *argv0, const gpt_params &params,
                                const server_params &sparams)
 {
-    fprintf(stdout, "usage: %s [options]\n", argv0);
-    fprintf(stdout, "\n");
-    fprintf(stdout, "options:\n");
-    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stdout, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    printf("usage: %s [options]\n", argv0);
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  --rope-freq-base N    RoPE base frequency (default: loaded from model)\n");
+    printf("  --rope-freq-scale N   RoPE frequency scaling factor (default: loaded from model)\n");
+    printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported())
     {
-        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        printf("  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported())
     {
-        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
-    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
-    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
+    printf("  -ngl N, --n-gpu-layers N\n");
+    printf("                        number of layers to store in VRAM\n");
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
-    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stdout, "  -a ALIAS, --alias ALIAS\n");
-    fprintf(stdout, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stdout, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    fprintf(stdout, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
-    fprintf(stdout, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    fprintf(stdout, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    fprintf(stdout, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    fprintf(stdout, "\n");
+    printf("  -m FNAME, --model FNAME\n");
+    printf("                        model path (default: %s)\n", params.model.c_str());
+    printf("  -a ALIAS, --alias ALIAS\n");
+    printf("                        set an alias for the model, will be added as `model` field in completion response\n");
+    printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    printf("  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    printf("  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    printf("\n");
 }
 
 static void server_params_parse(int argc, char **argv, server_params &sparams,
@@ -904,14 +907,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
 #else
             LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {});
-#endif // GGML_USE_CUBLAS
-        }
-        else if (arg == "--low-vram" || arg == "-lv")
-        {
-#ifdef GGML_USE_CUBLAS
-            params.low_vram = true;
-#else
-            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--no-mul-mat-q" || arg == "-nommq")
@@ -942,7 +937,23 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            params.lora_adapter = argv[i];
+            params.lora_adapter.push_back(std::make_tuple(argv[i], 1.0f));
+            params.use_mmap = false;
+        }
+        else if (arg == "--lora-scaled")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            const char * lora_adapter = argv[i];
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            params.lora_adapter.push_back(std::make_tuple(lora_adapter, std::stof(argv[i])));
             params.use_mmap = false;
         }
         else if (arg == "--lora-base")
@@ -1001,7 +1012,7 @@ static json format_generation_settings(llama_server_context &llama)
                             eos_bias->second < 0.0f && std::isinf(eos_bias->second);
 
     return json{
-        {"n_ctx", llama.params.n_ctx},
+        {"n_ctx", llama.n_ctx},
         {"model", llama.params.model_alias},
         {"seed", llama.params.seed},
         {"temp", llama.params.temp},
@@ -1039,8 +1050,6 @@ static json format_timings(llama_server_context &llama)
 {
     const auto timings = llama_get_timings(llama.ctx);
 
-    assert(timings.n_eval == llama.num_tokens_predicted);
-
     return json{
         {"prompt_n", timings.n_p_eval},
         {"prompt_ms", timings.t_p_eval_ms},
@@ -1082,8 +1091,9 @@ static json format_final_response(llama_server_context &llama, const std::string
     return res;
 }
 
-static json format_partial_response(llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs)
-{
+static json format_partial_response(
+    llama_server_context &llama, const std::string &content, const std::vector<completion_token_output> &probs
+) {
     json res = json{
         {"content", content},
         {"stop", false},
@@ -1103,6 +1113,12 @@ static json format_tokenizer_response(const std::vector<llama_token> &tokens)
         {"tokens", tokens}};
 }
 
+static json format_detokenized_response(std::string content)
+{
+    return json{
+        {"content", content}};
+}
+
 template <typename T>
 static T json_value(const json &body, const std::string &key, const T &default_value)
 {
@@ -1154,7 +1170,7 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     const auto &logit_bias = body.find("logit_bias");
     if (logit_bias != body.end() && logit_bias->is_array())
     {
-        const int n_vocab = llama_n_vocab(llama.ctx);
+        const int n_vocab = llama_n_vocab(llama.model);
         for (const auto &el : *logit_bias)
         {
             if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
@@ -1191,6 +1207,27 @@ static void parse_options_completion(const json &body, llama_server_context &lla
     LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
 }
 
+static void parse_options_infill(const json &body, llama_server_context &llama)
+{
+    if (body.count("input_prefix") != 0)
+    {
+        llama.params.input_prefix = body["input_prefix"];
+    }
+    else
+    {
+        llama.params.input_prefix = "";
+    }
+    if (body.count("input_suffix") != 0)
+    {
+        llama.params.input_suffix = body["input_suffix"];
+    }
+    else
+    {
+        llama.params.input_suffix = "";
+    }
+    parse_options_completion(body, llama);
+}
+
 static void log_server_request(const Request &req, const Response &res)
 {
     LOG_INFO("request", {
@@ -1208,6 +1245,63 @@ static void log_server_request(const Request &req, const Response &res)
                            });
 }
 
+static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+static void beam_search_callback(void *callback_data, llama_beams_state beams_state) {
+    auto & llama = *static_cast<llama_server_context*>(callback_data);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        llama.generated_token_probs.resize(llama.generated_token_probs.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
+        std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
+        printf("%zu", n);
+    }
+    fflush(stdout);
+#if 0 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams:\n";
+    for (size_t i=0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{state.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+
+struct token_translator {
+    llama_context * ctx;
+    std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
+    std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); }
+};
+
+static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama)
+{
+    auto & gtps = llama.generated_token_probs;
+    auto translator = token_translator{llama.ctx};
+    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
+    const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
+    if (llama.generated_text.capacity() < llama.generated_text.size() + len) {
+        llama.generated_text.reserve(llama.generated_text.size() + len);
+    }
+    for (const completion_token_output & cto : gtps) {
+        llama.generated_text += translator(cto);
+    }
+}
+
 int main(int argc, char **argv)
 {
     // own arguments required by this example
@@ -1230,6 +1324,7 @@ int main(int argc, char **argv)
                             {"commit", BUILD_COMMIT}});
     LOG_INFO("system info", {
                                 {"n_threads", params.n_threads},
+                                {"n_threads_batch", params.n_threads_batch},
                                 {"total_threads", std::thread::hardware_concurrency()},
                                 {"system_info", llama_print_system_info()},
                             });
@@ -1290,25 +1385,39 @@ int main(int argc, char **argv)
         llama.beginCompletion();
 
         if (!llama.stream) {
-            size_t stop_pos = std::string::npos;
+            if (llama.params.n_beams) {
+                // Fill llama.generated_token_probs vector with final beam.
+                llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
+                                  llama.n_past, llama.n_remain);
+                // Translate llama.generated_token_probs to llama.generated_text.
+                append_to_generated_text_from_generated_token_probs(llama);
+            } else {
+                size_t stop_pos = std::string::npos;
 
-            while (llama.has_next_token) {
-                const completion_token_output token_with_probs = llama.doCompletion();
-                const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
+                while (llama.has_next_token) {
+                    const completion_token_output token_with_probs = llama.doCompletion();
+                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok);
 
-                stop_pos = llama.findStoppingStrings(llama.generated_text,
-                    token_text.size(), STOP_FULL);
+                    stop_pos = llama.findStoppingStrings(llama.generated_text,
+                        token_text.size(), STOP_FULL);
+                }
+
+                if (stop_pos == std::string::npos) {
+                    stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
+                }
+                if (stop_pos != std::string::npos) {
+                    llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
+                        llama.generated_text.end());
+                }
             }
 
-            if (stop_pos == std::string::npos) {
-                stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
-            }
-            if (stop_pos != std::string::npos) {
-                llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
-                    llama.generated_text.end());
+            auto probs = llama.generated_token_probs;
+            if (llama.params.n_probs > 0 && llama.stopped_word) {
+                const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
+                probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
             }
 
-            const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
+            const json data = format_final_response(llama, llama.generated_text, probs);
 
             llama_print_timings(llama.ctx);
 
@@ -1321,27 +1430,157 @@ int main(int argc, char **argv)
 
                 while (llama.has_next_token) {
                     const completion_token_output token_with_probs = llama.doCompletion();
-                    const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok);
-                    if (llama.multibyte_pending > 0) {
+                    if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
                         continue;
                     }
+                    const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
 
                     size_t pos = std::min(sent_count, llama.generated_text.size());
 
                     const std::string str_test = llama.generated_text.substr(pos);
+                    bool is_stop_full = false;
                     size_t stop_pos =
                         llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
                     if (stop_pos != std::string::npos) {
+                        is_stop_full = true;
                         llama.generated_text.erase(
                             llama.generated_text.begin() + pos + stop_pos,
                             llama.generated_text.end());
                         pos = std::min(sent_count, llama.generated_text.size());
                     } else {
+                        is_stop_full = false;
                         stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
                             STOP_PARTIAL);
                     }
 
-                    const std::string to_send = llama.generated_text.substr(pos, stop_pos);
+                    if (
+                        stop_pos == std::string::npos ||
+                        // Send rest of the text if we are at the end of the generation
+                        (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                    ) {
+                        const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
+                        sent_count += to_send.size();
+
+                        std::vector<completion_token_output> probs_output = {};
+
+                        if (llama.params.n_probs > 0) {
+                            const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                            size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                            size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                            if (probs_pos < probs_stop_pos) {
+                                probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                            }
+                            sent_token_probs_index = probs_stop_pos;
+                        }
+
+                        const json data = format_partial_response(llama, to_send, probs_output);
+
+                        const std::string str =
+                            "data: " +
+                            data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                            "\n\n";
+
+                        LOG_VERBOSE("data stream", {
+                            { "to_send", str }
+                        });
+
+                        if (!sink.write(str.data(), str.size())) {
+                            LOG_VERBOSE("stream closed", {});
+                            llama_print_timings(llama.ctx);
+                            return false;
+                        }
+                    }
+
+                    if (!llama.has_next_token) {
+                        // Generation is done, send extra information.
+                        const json data = format_final_response(
+                            llama,
+                            "",
+                            std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                        );
+
+                        const std::string str =
+                            "data: " +
+                            data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                            "\n\n";
+
+                        LOG_VERBOSE("data stream", {
+                            { "to_send", str }
+                        });
+
+                        if (!sink.write(str.data(), str.size())) {
+                            LOG_VERBOSE("stream closed", {});
+                            llama_print_timings(llama.ctx);
+                            return false;
+                        }
+                    }
+                }
+
+                llama_print_timings(llama.ctx);
+                sink.done();
+                return true;
+            };
+            const auto on_complete = [&](bool) {
+                llama.mutex.unlock();
+            };
+            lock.release();
+            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+        } });
+
+    svr.Post("/infill", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        llama.rewind();
+
+        llama_reset_timings(llama.ctx);
+
+        parse_options_infill(json::parse(req.body), llama);
+
+        if (!llama.loadGrammar())
+        {
+            res.status = 400;
+            return;
+        }
+        llama.loadInfill();
+        llama.beginCompletion();
+        const auto chunked_content_provider = [&](size_t, DataSink & sink) {
+            size_t sent_count = 0;
+            size_t sent_token_probs_index = 0;
+
+            while (llama.has_next_token) {
+                const completion_token_output token_with_probs = llama.doCompletion();
+                if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
+                    continue;
+                }
+                const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+                size_t pos = std::min(sent_count, llama.generated_text.size());
+
+                const std::string str_test = llama.generated_text.substr(pos);
+                bool is_stop_full = false;
+                size_t stop_pos =
+                    llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+                if (stop_pos != std::string::npos) {
+                    is_stop_full = true;
+                    llama.generated_text.erase(
+                        llama.generated_text.begin() + pos + stop_pos,
+                        llama.generated_text.end());
+                    pos = std::min(sent_count, llama.generated_text.size());
+                } else {
+                    is_stop_full = false;
+                    stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                        STOP_PARTIAL);
+                }
+
+                if (
+                    stop_pos == std::string::npos ||
+                    // Send rest of the text if we are at the end of the generation
+                    (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+                ) {
+                    const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
                     sent_count += to_send.size();
 
                     std::vector<completion_token_output> probs_output = {};
@@ -1356,10 +1595,7 @@ int main(int argc, char **argv)
                         sent_token_probs_index = probs_stop_pos;
                     }
 
-                    const json data = llama.has_next_token
-                                          ? format_partial_response(llama, to_send, probs_output)
-                                          // Generation is done, send extra information.
-                                          : format_final_response(llama, to_send, llama.generated_token_probs);
+                    const json data = format_partial_response(llama, to_send, probs_output);
 
                     const std::string str =
                         "data: " +
@@ -1377,16 +1613,41 @@ int main(int argc, char **argv)
                     }
                 }
 
-                llama_print_timings(llama.ctx);
-                sink.done();
-                return true;
-            };
-            const auto on_complete = [&](bool) {
-                llama.mutex.unlock();
-            };
-            lock.release();
-            res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
-        } });
+                if (!llama.has_next_token) {
+                    // Generation is done, send extra information.
+                    const json data = format_final_response(
+                        llama,
+                        "",
+                        std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
+                    );
+
+                    const std::string str =
+                        "data: " +
+                        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+                        "\n\n";
+
+                    LOG_VERBOSE("data stream", {
+                        { "to_send", str }
+                    });
+
+                    if (!sink.write(str.data(), str.size())) {
+                        LOG_VERBOSE("stream closed", {});
+                        llama_print_timings(llama.ctx);
+                        return false;
+                    }
+                }
+            }
+
+            llama_print_timings(llama.ctx);
+            sink.done();
+            return true;
+        };
+        const auto on_complete = [&](bool) {
+            llama.mutex.unlock();
+        };
+        lock.release();
+        res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
+        });
 
     svr.Get("/model.json", [&llama](const Request &, Response &res)
             {
@@ -1409,6 +1670,21 @@ int main(int argc, char **argv)
         const json data = format_tokenizer_response(tokens);
         return res.set_content(data.dump(), "application/json"); });
 
+    svr.Post("/detokenize", [&llama](const Request &req, Response &res)
+             {
+        auto lock = llama.lock();
+
+        const json body = json::parse(req.body);
+        std::string content;
+        if (body.count("tokens") != 0)
+        {
+            const std::vector<llama_token> tokens = body["tokens"];
+            content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
+        }
+
+        const json data = format_detokenized_response(content);
+        return res.set_content(data.dump(), "application/json"); });
+
     svr.Post("/embedding", [&llama](const Request &req, Response &res)
              {
         auto lock = llama.lock();
@@ -1437,7 +1713,7 @@ int main(int argc, char **argv)
 
     svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep)
                               {
-        const auto * fmt = "500 Internal Server Error\n%s";
+        const char fmt[] = "500 Internal Server Error\n%s";
         char buf[BUFSIZ];
         try {
             std::rethrow_exception(std::move(ep));
@@ -1472,7 +1748,7 @@ int main(int argc, char **argv)
     svr.set_base_dir(sparams.public_path);
 
     // to make it ctrl+clickable:
-    fprintf(stdout, "\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+    printf("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
 
     LOG_INFO("HTTP server listening", {
                                           {"hostname", sparams.hostname},
diff --git a/examples/simple/CMakeLists.txt b/examples/simple/CMakeLists.txt
index 0ac9cb03a..7da5ff6f3 100644
--- a/examples/simple/CMakeLists.txt
+++ b/examples/simple/CMakeLists.txt
@@ -3,6 +3,3 @@ add_executable(${TARGET} simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
diff --git a/examples/simple/README.md b/examples/simple/README.md
new file mode 100644
index 000000000..5d24b1046
--- /dev/null
+++ b/examples/simple/README.md
@@ -0,0 +1,21 @@
+# llama.cpp/example/simple
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
+
+```bash
+./simple ./models/llama-7b-v2/ggml-model-f16.gguf "Hello my name is"
+
+...
+
+main: n_len = 32, n_ctx = 2048, n_parallel = 1, n_kv_req = 32
+
+ Hello my name is Shawn and I'm a 20 year old male from the United States. I'm a 20 year old
+
+main: decoded 27 tokens in 2.31 s, speed: 11.68 t/s
+
+llama_print_timings:        load time =   579.15 ms
+llama_print_timings:      sample time =     0.72 ms /    28 runs   (    0.03 ms per token, 38888.89 tokens per second)
+llama_print_timings: prompt eval time =   655.63 ms /    10 tokens (   65.56 ms per token,    15.25 tokens per second)
+llama_print_timings:        eval time =  2180.97 ms /    27 runs   (   80.78 ms per token,    12.38 tokens per second)
+llama_print_timings:       total time =  2891.13 ms
+```
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 132f7fbf9..24fb16b78 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -1,9 +1,3 @@
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"
 
@@ -32,99 +26,167 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    // total length of the sequence including the prompt
+    const int n_len = 32;
+
     // init LLM
 
     llama_backend_init(params.numa);
 
-    llama_context_params ctx_params = llama_context_default_params();
+    // initialize the model
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
+    llama_model_params model_params = llama_model_default_params();
+
+    // model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
         fprintf(stderr , "%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = 2048;
+    ctx_params.n_threads = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
 
-    if ((int) tokens_list.size() > max_tokens_list_size) {
-        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_kv_req);
+
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
-    fprintf(stderr, "\n\n");
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
 
     for (auto id : tokens_list) {
-        fprintf(stderr, "%s", llama_token_to_str(ctx, id).c_str());
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
     }
 
     fflush(stderr);
 
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
+
+    llama_batch batch = llama_batch_init(512, 0);
+
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
+
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
+
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
     // main loop
 
-    // The LLM keeps a contextual cache memory of previous token evaluation.
-    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
-    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
-    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
 
-    const int n_gen = std::min(32, max_context_size);
+    const auto t_main_start = ggml_time_us();
 
-    while (llama_get_kv_cache_token_count(ctx) < n_gen) {
-        // evaluate the transformer
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
 
-        if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            // prepare the next batch
+            batch.n_tokens = 0;
+
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = 0;
+            batch.logits[batch.n_tokens] = true;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
-
-        tokens_list.clear();
-
-        // sample the next token
-
-        llama_token new_token_id = 0;
-
-        auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
-
-        // is it an end of stream ?
-        if (new_token_id == llama_token_eos(ctx)) {
-            fprintf(stderr, " [end of text]\n");
-            break;
-        }
-
-        // print the new token :
-        printf("%s", llama_token_to_str(ctx, new_token_id).c_str());
-        fflush(stdout);
-
-        // push this new token for next evaluation
-        tokens_list.push_back(new_token_id);
     }
 
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
     llama_free(ctx);
     llama_free_model(model);
 
     llama_backend_free();
 
-    fprintf(stderr, "\n\n");
-
     return 0;
 }
diff --git a/examples/speculative/CMakeLists.txt b/examples/speculative/CMakeLists.txt
new file mode 100644
index 000000000..6c5c9456e
--- /dev/null
+++ b/examples/speculative/CMakeLists.txt
@@ -0,0 +1,8 @@
+set(TARGET speculative)
+add_executable(${TARGET} speculative.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
new file mode 100644
index 000000000..75a2e5e22
--- /dev/null
+++ b/examples/speculative/speculative.cpp
@@ -0,0 +1,314 @@
+#include "build-info.h"
+
+#include "common.h"
+#include "llama.h"
+#include "grammar-parser.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.model_draft.empty()) {
+        fprintf(stderr, "%s: error: --model-draft is required\n", __func__);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("speculative", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
+
+    // init llama.cpp
+    llama_backend_init(params.numa);
+
+    llama_model * model_tgt = NULL;
+    llama_model * model_dft = NULL;
+
+    llama_context * ctx_tgt = NULL;
+    llama_context * ctx_dft = NULL;
+
+    // load the target model
+    params.logits_all = true;
+    std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
+
+    // load the draft model
+    params.model = params.model_draft;
+    params.n_gpu_layers = params.n_gpu_layers_draft;
+    std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
+
+    const int max_context_size     = llama_n_ctx(ctx_tgt);
+    const int max_tokens_list_size = max_context_size - 4;
+
+    if ((int) inp.size() > max_tokens_list_size) {
+        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
+        return 1;
+    }
+
+    fprintf(stderr, "\n\n");
+
+    for (auto id : inp) {
+        fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str());
+    }
+
+    fflush(stderr);
+
+    const int n_input = inp.size();
+
+    const auto t_enc_start = ggml_time_us();
+
+    // eval the prompt with both models
+    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0));
+
+    const auto t_enc_end = ggml_time_us();
+
+    // the 2 models should have the same vocab
+    const int n_ctx   = llama_n_ctx(ctx_tgt);
+    const int n_vocab = llama_n_vocab(model_tgt);
+    //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
+
+    // how many tokens to draft each time
+    int n_draft = params.n_draft;
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    int n_past_tgt = inp.size();
+    int n_past_dft = inp.size();
+
+    std::vector<llama_token> drafted;
+
+    std::vector<llama_token> last_tokens(n_ctx);
+    std::fill(last_tokens.begin(), last_tokens.end(), 0);
+
+    for (auto & id : inp) {
+        last_tokens.erase(last_tokens.begin());
+        last_tokens.push_back(id);
+    }
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    // used to determine end of generation
+    bool has_eos = false;
+
+    // grammar stuff
+    struct llama_grammar * grammar_dft = NULL;
+    struct llama_grammar * grammar_tgt = NULL;
+
+    grammar_parser::parse_state parsed_grammar;
+
+    // if requested - load the grammar, error checking is omitted for brevity
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
+
+        int i_dft = 0;
+
+        while (true) {
+            // sample from the target model
+            llama_token id = llama_sample_token(ctx_tgt, NULL, grammar_tgt, params, last_tokens, candidates, i_dft);
+
+            // remember which tokens were sampled - used for repetition penalties during sampling
+            last_tokens.erase(last_tokens.begin());
+            last_tokens.push_back(id);
+
+            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
+
+            const std::string token_str = llama_token_to_piece(ctx_tgt, id);
+            printf("%s", token_str.c_str());
+            fflush(stdout);
+
+            if (id == llama_token_eos(ctx_tgt)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
+
+            // check if the draft matches the target
+            if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
+                LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
+                ++n_accept;
+                ++n_past_tgt;
+                ++n_past_dft;
+                ++i_dft;
+
+                continue;
+            }
+
+            // the drafted token was rejected or we are out of drafted tokens
+
+            if (i_dft < (int) drafted.size()) {
+                LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
+                        i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
+            } else {
+                LOG("out of drafted tokens\n");
+            }
+
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
+            ++n_past_dft;
+
+            // heuristic for n_draft
+            {
+                const int  n_draft_cur  = (int) drafted.size();
+                const bool all_accepted = i_dft == n_draft_cur;
+
+                LOG("n_draft      = %d\n", n_draft);
+                LOG("n_draft_cur  = %d\n", n_draft_cur);
+                LOG("i_dft        = %d\n", i_dft);
+                LOG("all_accepted = %d\n", all_accepted);
+
+                if (all_accepted && n_draft == n_draft_cur) {
+                    LOG(" - max drafted tokens accepted - n_draft += 8\n");
+                    n_draft = std::min(30, n_draft + 8);
+                } else if (all_accepted) {
+                    LOG(" - partially drafted tokens accepted - no change\n");
+                } else {
+                    LOG(" - drafted token rejected - n_draft -= 1\n");
+                    n_draft = std::max(2, n_draft - 1);
+                }
+            }
+
+            drafted.clear();
+            drafted.push_back(id);
+
+            break;
+        }
+
+        if (n_predict > params.n_predict || has_eos) {
+            break;
+        }
+
+        if (grammar_tgt) {
+            if (grammar_dft) {
+                llama_grammar_free(grammar_dft);
+            }
+            grammar_dft = llama_grammar_copy(grammar_tgt);
+
+            LOG("copied target grammar to draft grammar\n");
+        }
+
+        // sample n_draft tokens from the draft model using greedy decoding
+        int n_past_cur = n_past_dft;
+        for (int i = 0; i < n_draft; ++i) {
+            float * logits = llama_get_logits(ctx_dft);
+
+            candidates.clear();
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+            }
+
+            llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
+
+            if (grammar_dft != NULL) {
+                llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
+            }
+
+            // computes softmax and sorts the candidates
+            llama_sample_softmax(ctx_dft, &cur_p);
+
+            for (int i = 0; i < 3; ++i) {
+                LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
+            }
+
+            // TODO: better logic?
+            if (cur_p.data[0].p < 2*cur_p.data[1].p) {
+                LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
+                break;
+            }
+
+            // drafted token
+            const llama_token id = cur_p.data[0].id;
+
+            drafted.push_back(id);
+            ++n_drafted;
+
+            // no need to evaluate the last drafted token, since we won't use the result
+            if (i == n_draft - 1) {
+                break;
+            }
+
+            // evaluate the drafted token on the draft model
+            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
+            llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
+            ++n_past_cur;
+
+            if (grammar_dft != NULL) {
+                llama_grammar_accept_token(ctx_dft, grammar_dft, id);
+            }
+        }
+
+        // evaluate the target model on the drafted tokens
+        llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
+        llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
+        ++n_past_tgt;
+
+        // the first token is always proposed by the traget model before the speculation loop
+        drafted.erase(drafted.begin());
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    LOG_TEE("\n\n");
+
+    LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
+
+    // TODO: make sure these numbers are computed correctly
+    LOG_TEE("\n");
+    LOG_TEE("n_draft   = %d\n", n_draft);
+    LOG_TEE("n_predict = %d\n", n_predict);
+    LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_TEE("\ndraft:\n");
+    llama_print_timings(ctx_dft);
+
+    LOG_TEE("\ntarget:\n");
+    llama_print_timings(ctx_tgt);
+
+    llama_free(ctx_tgt);
+    llama_free_model(model_tgt);
+
+    llama_free(ctx_dft);
+    llama_free_model(model_dft);
+
+    if (grammar_dft != NULL) {
+        llama_grammar_free(grammar_dft);
+        llama_grammar_free(grammar_tgt);
+    }
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}
diff --git a/examples/train-text-from-scratch/README.md b/examples/train-text-from-scratch/README.md
index 726ec47c0..1b3454069 100644
--- a/examples/train-text-from-scratch/README.md
+++ b/examples/train-text-from-scratch/README.md
@@ -8,15 +8,20 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 
 # train
 ./bin/train-text-from-scratch \
-        --vocab-model ../models/ggml-vocab.bin \
+        --vocab-model ../models/ggml-vocab-llama.gguf \
         --ctx 64 --embd 256 --head 8 --layer 16 \
-        --checkpoint-in  chk-shakespeare-256x16.bin \
-        --checkpoint-out chk-shakespeare-256x16.bin \
-        --model-out ggml-shakespeare-256x16-f32.bin \
+        --checkpoint-in  chk-shakespeare-256x16-LATEST.gguf \
+        --checkpoint-out chk-shakespeare-256x16-ITERATION.gguf \
+        --model-out ggml-shakespeare-256x16-f32-ITERATION.gguf \
         --train-data "shakespeare.txt" \
-        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
-        --print-details-interval 0 --predict 16 --use-flash
+        -t 6 -b 16 --seed 1 --adam-iter 256 \
+        --no-checkpointing
 
 # predict
-./bin/main -m ggml-shakespeare-256x16-f32.bin
+./bin/main -m ggml-shakespeare-256x16-f32.gguf
 ```
+
+Output files will be saved every N iterations (config with `--save-every N`).
+The pattern "ITERATION" in the output filenames will be replaced with the iteration number and "LATEST" for the latest output.
+
+To train GGUF models just pass them to `--checkpoint-in FN`.
diff --git a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
new file mode 100644
index 000000000..887ed2e21
--- /dev/null
+++ b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
@@ -0,0 +1,499 @@
+#!/usr/bin/env python3
+# train-text-from-scratch checkpoint --> gguf conversion
+
+import argparse
+import os
+import struct
+import sys
+import numpy as np
+from pathlib import Path
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py' / 'gguf'))
+import gguf
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        assert(tuple(ne) == tuple(self.ne))
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationParamsV0:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.type                 = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_threads            = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.past                 = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.delta                = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.print_forward_graph  = struct.unpack('<?', bytes(data[offset:offset + 1]))[0];  offset += 4 # 32bit-aligned
+        self.print_backward_graph = struct.unpack('<?', bytes(data[offset:offset + 1]))[0];  offset += 4 # 32bit-aligned
+        self.adam_n_iter          = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_sched           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_decay           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_alpha           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_beta1           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_beta2           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_eps             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_eps_f           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.adam_eps_g           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_n_iter         = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_max_linesearch = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_eps            = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_ftol           = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_wolfe          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_min_step       = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_max_step       = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_linesearch     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version == 0:
+            params = OptimizationParamsV0()
+            offset = params.load(data, offset)
+            self.past = params.past
+            self.lbfgs_m = params.lbfgs_m
+            self.nx = struct.unpack('N', bytes(data[offset:offset + 8]))[0];  offset += 8
+            self.iter = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+            self.type = params.type
+
+            self.adam_m  = Tensor('f', [self.nx])
+            self.adam_v  = Tensor('f', [self.nx])
+            self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+            self.lbfgs_x    = Tensor('f', [self.nx])
+            self.lbfgs_xp   = Tensor('f', [self.nx])
+            self.lbfgs_g    = Tensor('f', [self.nx])
+            self.lbfgs_gp   = Tensor('f', [self.nx])
+            self.lbfgs_d    = Tensor('f', [self.nx])
+            self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+            self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+            self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+            self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+            self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+            if self.type == 0:
+                # these tensors are stored, but we don't need their data
+                x  = Tensor('f', [self.nx])
+                g  = Tensor('f', [self.nx])
+                g2 = Tensor('f', [self.nx])
+                mh = Tensor('f', [self.nx])
+                vh = Tensor('f', [self.nx])
+
+                offset = x.load(data, offset)
+                offset = g.load(data, offset)
+                offset = g2.load(data, offset)
+                offset = self.adam_m.load(data, offset)
+                offset = self.adam_v.load(data, offset)
+                offset = mh.load(data, offset)
+                offset = vh.load(data, offset)
+                offset = self.adam_pf.load(data, offset)
+
+                self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+            elif self.type == 1:
+                offset = self.lbfgs_x.load(data, offset)
+                offset = self.lbfgs_xp.load(data, offset)
+                offset = self.lbfgs_g.load(data, offset)
+                offset = self.lbfgs_gp.load(data, offset)
+                offset = self.lbfgs_d.load(data, offset)
+                offset = self.lbfgs_pf.load(data, offset)
+                offset = self.lbfgs_lmal.load(data, offset)
+                offset = self.lbfgs_lmys.load(data, offset)
+                offset = self.lbfgs_lms.load(data, offset)
+                offset = self.lbfgs_lmy.load(data, offset)
+
+                self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+            else:
+                raise ValueError('Unknown optimizer type')
+
+
+        elif self.version == 1:
+            self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+            self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+            self.adam_m  = Tensor('f', [self.nx])
+            self.adam_v  = Tensor('f', [self.nx])
+            self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+            self.lbfgs_x    = Tensor('f', [self.nx])
+            self.lbfgs_xp   = Tensor('f', [self.nx])
+            self.lbfgs_g    = Tensor('f', [self.nx])
+            self.lbfgs_gp   = Tensor('f', [self.nx])
+            self.lbfgs_d    = Tensor('f', [self.nx])
+            self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+            self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+            self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+            self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+            self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+            # forgot to save type in version 1:
+            # guess self.type from number of remaining bytes
+            size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                    [self.adam_m, self.adam_v]
+                                    +([self.adam_pf] if (self.past > 0) else [])])
+            size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                    [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                     self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                     self.lbfgs_lmal, self.lbfgs_lmys,
+                                     self.lbfgs_lms, self.lbfgs_lmy]
+                                     +([self.lbfgs_pf] if (self.past > 0) else [])])
+            # due to alignment padding the size might not by exact
+            # but the difference in size for both types is significant,
+            # so we can just use whichever is closest
+            remaining = len(data) - offset
+            if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+                self.type = 0
+            else:
+                self.type = 1
+
+            if self.type == 0:
+                offset = self.adam_m.load(data, offset)
+                offset = self.adam_v.load(data, offset)
+                offset = self.adam_pf.load(data,offset)
+
+                self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+            elif self.type == 1:
+                offset = self.lbfgs_x.load(data, offset)
+                offset = self.lbfgs_xp.load(data, offset)
+                offset = self.lbfgs_g.load(data, offset)
+                offset = self.lbfgs_gp.load(data, offset)
+                offset = self.lbfgs_d.load(data, offset)
+                offset = self.lbfgs_pf.load(data, offset)
+                offset = self.lbfgs_lmal.load(data, offset)
+                offset = self.lbfgs_lmys.load(data, offset)
+                offset = self.lbfgs_lms.load(data, offset)
+                offset = self.lbfgs_lmy.load(data, offset)
+
+                self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+                self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError('Invalid version of checkpoint file')
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class ModelParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+        return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None):
+    return gguf.TENSOR_NAMES[key].format(bid=bid) + ".weight"
+
+class Layer:
+    def __init__(self, params, bid):
+        self.bid = bid
+        self.att_norm = Tensor('f', [params.n_embd])
+        self.wq       = Tensor('f', [params.n_embd, params.n_embd])
+        self.wk       = Tensor('f', [params.n_embd, params.n_embd])
+        self.wv       = Tensor('f', [params.n_embd, params.n_embd])
+        self.wo       = Tensor('f', [params.n_embd, params.n_embd])
+        self.ffn_norm = Tensor('f', [params.n_embd])
+        self.w1       = Tensor('f', [params.n_embd, params.get_n_ff()])
+        self.w2       = Tensor('f', [params.get_n_ff(), params.n_embd])
+        self.w3       = Tensor('f', [params.n_embd, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm.load(data, offset)
+        offset = self.wq.load(data, offset)
+        offset = self.wk.load(data, offset)
+        offset = self.wv.load(data, offset)
+        offset = self.wo.load(data, offset)
+        offset = self.ffn_norm.load(data, offset)
+        offset = self.w1.load(data, offset)
+        offset = self.w2.load(data, offset)
+        offset = self.w3.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid))
+        self.wq.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid))
+        self.wk.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid))
+        self.wv.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid))
+        self.wo.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid))
+        self.ffn_norm.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid))
+        self.w1.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid))
+        self.w2.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid))
+        self.w3.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid))
+
+class Model:
+    def __init__(self):
+        self.params = ModelParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+
+        self.tok_embd = Tensor('f', [self.params.n_embd, self.params.n_vocab])
+        self.norm     = Tensor('f', [self.params.n_embd])
+        self.output   = Tensor('f', [self.params.n_embd, self.params.n_vocab])
+
+        offset = self.tok_embd.load(data, offset)
+        offset = self.norm.load(data, offset)
+        offset = self.output.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+
+        self.tok_embd.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD))
+        self.norm.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM))
+        self.output.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class Checkpoint:
+    def __init__(self):
+        self.model = Model()
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcp':
+            raise ValueError(f"File header magic indicates, that this is no checkpoint file. Expected 'ggcp', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_TRAIN_MODEL)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert train-text-from-scratch checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input train checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help ='Output GGUF filename', required=True)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = Checkpoint()
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    print("    gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("    gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("    gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+
+if __name__ == '__main__':
+    main()
diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
index 79b117df7..be693b3ac 100644
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -1,5 +1,7 @@
 #include "ggml.h"
+#include "ggml-alloc.h"
 #include "common.h"
+#include "train.h"
 #include "llama.h"
 #include <unordered_map>
 #include <vector>
@@ -17,183 +19,22 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const float rms_norm_eps = 1e-5f;
-
-struct random_normal_distribution {
-    std::mt19937 gen;
-    std::normal_distribution<float> rd;
-    float min;
-    float max;
-};
-
-struct random_uniform_distribution {
-    std::mt19937 gen;
-    std::uniform_real_distribution<float> rd;
-};
-
-void init_random_normal_distribution(struct random_normal_distribution * rnd, int seed, float mean, float std, float min, float max) {
-    rnd->gen = std::mt19937(seed);
-    rnd->rd = std::normal_distribution<float>{mean, std};
-    rnd->min = min;
-    rnd->max = max;
-}
-
-void init_random_uniform_distribution(struct random_uniform_distribution * rnd, int seed, float min, float max) {
-    rnd->gen = std::mt19937(seed);
-    rnd->rd = std::uniform_real_distribution<float>{min, max};
-}
-
-int clamp(const int v, const int min, const int max) {
-    return ((v < min) ? (min) : (v > max) ? (max) : v);
-}
-
-float fclamp(const float v, const float min, const float max) {
-    return ((v < min) ? (min) : (v > max) ? (max) : v);
-}
-
-float frand() {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-float frand_normal(struct random_normal_distribution * rnd) {
-    return fclamp(rnd->rd(rnd->gen), rnd->min, rnd->max);
-}
-
-float frand_uniform(struct random_uniform_distribution * rnd) {
-    return rnd->rd(rnd->gen);
-}
-
-void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
-
-    if (plan.work_size > 0) {
-        buf.resize(plan.work_size);
-        plan.work_data = buf.data();
-    }
-
-    ggml_graph_compute(graph, &plan);
-}
-
-struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
-    float scale = 1.0f; // xavier
-    switch (tensor->n_dims) {
-        case 1:
-            scale /= sqrtf(tensor->ne[0]);
-            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
-                *dst = scale * frand_normal(rnd);
-            }
-            break;
-        case 2:
-            scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
-            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-                    *dst = scale * frand_normal(rnd);
-                }
-            }
-            break;
-        case 3:
-            scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
-            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
-                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
-                        *dst = scale * frand_normal(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            scale /= sqrtf(tensor->ne[0]+tensor->ne[1]);
-            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
-                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
-                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
-                            *dst = scale * frand_normal(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
-    return tensor;
-}
-
-struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
-    switch (tensor->n_dims) {
-        case 1:
-            for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
-                *dst = frand_uniform(rnd);
-            }
-            break;
-        case 2:
-            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                    float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-                    *dst = frand_uniform(rnd);
-                }
-            }
-            break;
-        case 3:
-            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
-                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                        float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
-                        *dst = frand_uniform(rnd);
-                    }
-                }
-            }
-            break;
-        case 4:
-            for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
-                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
-                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                            float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3]);
-                            *dst = frand_uniform(rnd);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    };
-    return tensor;
-}
-
-struct llama_vocab {
-    using id    = int32_t;
-    using token = std::string;
-    using ttype = llama_token_type;
-
-    struct token_data {
-        token text;
-        float score;
-        ttype type;
-    };
-
-    std::unordered_map<token, id> token_to_id;
-    std::vector<token_data> id_to_token;
-};
+static const size_t tensor_alignment = 32;
 
 struct my_llama_hparams {
     uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
+    uint32_t n_ctx   = 512;
     uint32_t n_embd  = 4096;
-    uint32_t n_mult  = 4;
     uint32_t n_head  = 32;
     uint32_t n_layer = 32;
     uint32_t n_rot   = 64;
+    uint32_t n_ff    = 11008;
 
-    bool operator!=(const my_llama_hparams& other) const {
-        return memcmp(this, &other, sizeof(my_llama_hparams));
-    }
+    // float f_norm_eps     = 1e-5f; // falcon
+    float f_norm_rms_eps = 1e-5f; // llama
+
+    float rope_freq_base  = 10000.0f;
+    float rope_freq_scale = 1.0f;
 };
 
 struct my_llama_layer {
@@ -215,19 +56,9 @@ struct my_llama_layer {
     struct ggml_tensor * w3;
 };
 
-struct my_llama_kv_cache {
-    struct ggml_context * ctx = NULL;
-
-    struct ggml_tensor * k;
-    struct ggml_tensor * v;
-
-    // llama_ctx_buffer buf;
-
-    int n; // number of tokens currently in the cache
-};
-
 struct my_llama_model {
     struct ggml_context * ctx = NULL;
+    std::vector<uint8_t> data;
 
     my_llama_hparams hparams;
 
@@ -237,86 +68,60 @@ struct my_llama_model {
     struct ggml_tensor * output;
 
     std::vector<my_llama_layer> layers;
-
-    uint32_t train_its = 0;
-    uint32_t train_samples = 0;
-    uint32_t train_tokens = 0;
 };
 
-uint32_t get_n_ff(const struct my_llama_hparams* hparams) {
-    const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult;
-    return n_ff;
-}
+// gguf constants (sync with gguf.py)
+static const char * LLM_KV_TRAINING_TYPE_TRAIN_MODEL     = "train_model";
+static const char * LLM_KV_TRAINING_TYPE                 = "training.type";
 
-void print_params(struct my_llama_hparams * params) {
+static const char * LLM_KV_GENERAL_ARCHITECTURE        = "general.architecture";
+static const char * LLM_KV_GENERAL_FILE_TYPE           = "general.file_type";
+
+static const char * LLM_KV_CONTEXT_LENGTH              = "%s.context_length";
+static const char * LLM_KV_EMBEDDING_LENGTH            = "%s.embedding_length";
+static const char * LLM_KV_BLOCK_COUNT                 = "%s.block_count";
+static const char * LLM_KV_FEED_FORWARD_LENGTH         = "%s.feed_forward_length";
+static const char * LLM_KV_ATTENTION_HEAD_COUNT        = "%s.attention.head_count";
+static const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon";
+static const char * LLM_KV_ROPE_DIMENSION_COUNT        = "%s.rope.dimension_count";
+static const char * LLM_KV_ROPE_FREQ_BASE              = "%s.rope.freq_base"; // TODO load in llama.cpp
+static const char * LLM_KV_ROPE_SCALE_LINEAR           = "%s.rope.scale_linear";
+
+static const char * LLM_KV_TOKENIZER_MODEL             = "tokenizer.ggml.model";
+static const char * LLM_KV_TOKENIZER_LIST              = "tokenizer.ggml.tokens";
+static const char * LLM_KV_TOKENIZER_TOKEN_TYPE        = "tokenizer.ggml.token_type";
+static const char * LLM_KV_TOKENIZER_SCORES            = "tokenizer.ggml.scores";
+static const char * LLM_KV_TOKENIZER_MERGES            = "tokenizer.ggml.merges";
+static const char * LLM_KV_TOKENIZER_BOS_ID            = "tokenizer.ggml.bos_token_id";
+static const char * LLM_KV_TOKENIZER_EOS_ID            = "tokenizer.ggml.eos_token_id";
+static const char * LLM_KV_TOKENIZER_UNK_ID            = "tokenizer.ggml.unknown_token_id";
+static const char * LLM_KV_TOKENIZER_SEP_ID            = "tokenizer.ggml.seperator_token_id";
+static const char * LLM_KV_TOKENIZER_PAD_ID            = "tokenizer.ggml.padding_token_id";
+
+static const char * LLM_TENSOR_TOKEN_EMBD    = "token_embd";
+static const char * LLM_TENSOR_OUTPUT_NORM   = "output_norm";
+static const char * LLM_TENSOR_OUTPUT        = "output";
+static const char * LLM_TENSOR_ATTN_NORM     = "blk.%d.attn_norm";
+static const char * LLM_TENSOR_ATTN_Q        = "blk.%d.attn_q";
+static const char * LLM_TENSOR_ATTN_K        = "blk.%d.attn_k";
+static const char * LLM_TENSOR_ATTN_V        = "blk.%d.attn_v";
+static const char * LLM_TENSOR_ATTN_OUT      = "blk.%d.attn_output";
+static const char * LLM_TENSOR_FFN_NORM      = "blk.%d.ffn_norm";
+static const char * LLM_TENSOR_FFN_GATE      = "blk.%d.ffn_gate";
+static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
+static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
+
+static void print_params(struct my_llama_hparams * params) {
     printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
     printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
     printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
     printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, get_n_ff(params));
+    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
     printf("%s: n_layer: %d\n", __func__, params->n_layer);
     printf("%s: n_rot:   %d\n", __func__, params->n_rot);
 }
 
-void init_model(struct my_llama_model * model) {
-    const auto & hparams = model->hparams;
-
-    const uint32_t n_embd  = hparams.n_embd;
-    const uint32_t n_layer = hparams.n_layer;
-    const uint32_t n_vocab = hparams.n_vocab;
-
-    const uint32_t n_ff = get_n_ff(&hparams);
-
-    struct ggml_context * ctx = model->ctx;
-
-    model->train_its = 0;
-    model->train_samples = 0;
-    model->train_tokens = 0;
-
-    model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-
-    ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
-    ggml_set_name(model->norm,           "norm.weight");
-    ggml_set_name(model->output,         "output.weight");
-
-    model->layers.resize(n_layer);
-    for (uint32_t i = 0; i < n_layer; ++i) {
-        auto & layer = model->layers[i];
-
-        std::string layers_i = "layers." + std::to_string(i);
-
-        layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-        layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-
-        layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-
-        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
-        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-
-        ggml_set_name(layer.attention_norm, (layers_i + ".attention_norm.weight").c_str());
-
-        ggml_set_name(layer.wq, (layers_i + ".attention.wq.weight").c_str());
-        ggml_set_name(layer.wk, (layers_i + ".attention.wk.weight").c_str());
-        ggml_set_name(layer.wv, (layers_i + ".attention.wv.weight").c_str());
-        ggml_set_name(layer.wo, (layers_i + ".attention.wo.weight").c_str());
-
-        ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str());
-
-        ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str());
-        ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
-        ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
-    }
-}
-
-void set_param_model(struct my_llama_model * model) {
+static void set_param_model(struct my_llama_model * model) {
     const auto& hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
@@ -342,1102 +147,172 @@ void set_param_model(struct my_llama_model * model) {
     }
 }
 
-void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
+static void alloc_model(struct ggml_allocr * alloc, struct my_llama_model * model) {
+    ggml_allocr_alloc(alloc, model->tok_embeddings);
+    ggml_allocr_alloc(alloc, model->norm);
+    ggml_allocr_alloc(alloc, model->output);
+    for (uint32_t i = 0; i < model->layers.size(); ++i) {
+        auto & layer = model->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm);
+        ggml_allocr_alloc(alloc, layer.wq);
+        ggml_allocr_alloc(alloc, layer.wk);
+        ggml_allocr_alloc(alloc, layer.wv);
+        ggml_allocr_alloc(alloc, layer.wo);
+        ggml_allocr_alloc(alloc, layer.ffn_norm);
+        ggml_allocr_alloc(alloc, layer.w1);
+        ggml_allocr_alloc(alloc, layer.w2);
+        ggml_allocr_alloc(alloc, layer.w3);
+    }
+    ggml_allocr_alloc(alloc, model->tok_embeddings->grad);
+    ggml_allocr_alloc(alloc, model->norm->grad);
+    ggml_allocr_alloc(alloc, model->output->grad);
+    for (uint32_t i = 0; i < model->layers.size(); ++i) {
+        auto & layer = model->layers[i];
+        ggml_allocr_alloc(alloc, layer.attention_norm->grad);
+        ggml_allocr_alloc(alloc, layer.wq->grad);
+        ggml_allocr_alloc(alloc, layer.wk->grad);
+        ggml_allocr_alloc(alloc, layer.wv->grad);
+        ggml_allocr_alloc(alloc, layer.wo->grad);
+        ggml_allocr_alloc(alloc, layer.ffn_norm->grad);
+        ggml_allocr_alloc(alloc, layer.w1->grad);
+        ggml_allocr_alloc(alloc, layer.w2->grad);
+        ggml_allocr_alloc(alloc, layer.w3->grad);
+    }
+}
+
+static void init_model(struct my_llama_model * model) {
+    const auto & hparams = model->hparams;
+
+    const uint32_t n_embd  = hparams.n_embd;
+    const uint32_t n_layer = hparams.n_layer;
+    const uint32_t n_vocab = hparams.n_vocab;
+    const uint32_t n_ff    = hparams.n_ff;
+
+
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
+        return tn_buf.data();
+    };
+
+    // context for model tensors without their data
+    struct ggml_init_params ctx_model_params;
+    ctx_model_params.mem_size   = ggml_tensor_overhead()*2*(6 + n_layer*18);
+    ctx_model_params.mem_buffer = NULL;
+    ctx_model_params.no_alloc   = true;
+
+    struct ggml_context * ctx = ggml_init(ctx_model_params);
+    model->ctx = ctx;
+
+    model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+    model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+    model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
+
+    ggml_set_name(model->tok_embeddings, tn(LLM_TENSOR_TOKEN_EMBD));
+    ggml_set_name(model->norm,           tn(LLM_TENSOR_OUTPUT_NORM));
+    ggml_set_name(model->output,         tn(LLM_TENSOR_OUTPUT));
+
+    model->layers.resize(n_layer);
+    for (uint32_t i = 0; i < n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+        layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+
+        layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
+        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+
+        ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
+
+        ggml_set_name(layer.wq,             tni(LLM_TENSOR_ATTN_Q, i));
+        ggml_set_name(layer.wk,             tni(LLM_TENSOR_ATTN_K, i));
+        ggml_set_name(layer.wv,             tni(LLM_TENSOR_ATTN_V, i));
+        ggml_set_name(layer.wo,             tni(LLM_TENSOR_ATTN_OUT, i));
+
+        ggml_set_name(layer.ffn_norm,       tni(LLM_TENSOR_FFN_NORM, i));
+
+        ggml_set_name(layer.w1,             tni(LLM_TENSOR_FFN_GATE, i));
+        ggml_set_name(layer.w2,             tni(LLM_TENSOR_FFN_DOWN, i));
+        ggml_set_name(layer.w3,             tni(LLM_TENSOR_FFN_UP, i));
+    }
+
+    set_param_model(model);
+
+    // measure data size
+    struct ggml_allocr * alloc = NULL;
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc_model(alloc, model);
+
+    // allocate data
+    model->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
+    ggml_allocr_free(alloc);
+    alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
+    alloc_model(alloc, model);
+    ggml_allocr_free(alloc);
+}
+
+static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
     const auto & hparams = model->hparams;
 
     const uint32_t n_layer = hparams.n_layer;
 
-    struct random_normal_distribution rnd;
-    init_random_normal_distribution(&rnd, seed, mean, std, min, max);
+    struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
 
-    randomize_tensor_normal(model->tok_embeddings, &rnd);
-    randomize_tensor_normal(model->norm,           &rnd);
-    randomize_tensor_normal(model->output,         &rnd);
+    randomize_tensor_normal(model->tok_embeddings, rnd);
+    randomize_tensor_normal(model->norm,           rnd);
+    randomize_tensor_normal(model->output,         rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = model->layers[i];
-        randomize_tensor_normal(layer.attention_norm, &rnd);
+        randomize_tensor_normal(layer.attention_norm, rnd);
 
-        randomize_tensor_normal(layer.wq, &rnd);
-        randomize_tensor_normal(layer.wk, &rnd);
-        randomize_tensor_normal(layer.wv, &rnd);
-        randomize_tensor_normal(layer.wo, &rnd);
+        randomize_tensor_normal(layer.wq, rnd);
+        randomize_tensor_normal(layer.wk, rnd);
+        randomize_tensor_normal(layer.wv, rnd);
+        randomize_tensor_normal(layer.wo, rnd);
 
-        randomize_tensor_normal(layer.ffn_norm, &rnd);
+        randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, &rnd);
-        randomize_tensor_normal(layer.w2, &rnd);
-        randomize_tensor_normal(layer.w3, &rnd);
-    }
-}
-
-bool init_kv_cache(struct my_llama_kv_cache* cache, struct my_llama_model * model, int n_batch) {
-    const auto & hparams = model->hparams;
-
-    const uint32_t n_ctx   = hparams.n_ctx;
-    const uint32_t n_embd  = hparams.n_embd;
-    const uint32_t n_layer = hparams.n_layer;
-
-    const int64_t n_mem      = n_layer*n_ctx*n_batch;
-    const int64_t n_elements = n_embd*n_mem;
-
-    // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
-
-    // struct ggml_init_params params;
-    // params.mem_size   = cache.buf.size;
-    // params.mem_buffer = cache.buf.addr;
-    // params.no_alloc   = false;
-    if (!cache->ctx) {
-        struct ggml_init_params params;
-        params.mem_size   = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024;
-        params.mem_buffer = NULL;
-        params.no_alloc   = false;
-
-        cache->ctx = ggml_init(params);
-
-        if (!cache->ctx) {
-            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
-            return false;
-        }
+        randomize_tensor_normal(layer.w1, rnd);
+        randomize_tensor_normal(layer.w2, rnd);
+        randomize_tensor_normal(layer.w3, rnd);
     }
 
-    cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-    cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements);
-
-    return true;
+    free_random_normal_distribution(rnd);
 }
 
-struct ggml_tensor * forward(
-        struct my_llama_model    * model,
-        struct my_llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past) {
-
-    const int N = n_tokens;
-
-    struct my_llama_kv_cache& kv_self = *cache;
-    const auto & hparams = model->hparams;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_head  = hparams.n_head;
-    const int n_rot   = hparams.n_rot;
-
-    struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(tokens->data, tokens_input->data, N*ggml_element_size(tokens));
-
-    struct ggml_tensor * kc = kv_self.k;
-    struct ggml_tensor * vc = kv_self.v;
-
-    // inpL shape [n_embd,N,1,1]
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * inpSA = inpL;
-
-        struct ggml_tensor * cur;
-
-        // lctx.use_buf(ctx0, 0);
-
-        // norm
-        {
-            // cur shape [n_embd,N,1,1]
-            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-
-            // cur = attention_norm*cur
-            cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
-                        cur);
-        }
-
-        // self-attention
-        {
-            // compute Q and K and RoPE them
-            // wq   shape [n_embd, n_embd, 1, 1]
-            // wk   shape [n_embd, n_embd, 1, 1]
-            // Qcur shape [n_embd/n_head, n_head, N, 1]
-            // Kcur shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
-
-            // store key and value to memory
-            {
-                // compute the transposed [N, n_embd] V matrix
-                // wv   shape [n_embd, n_embd, 1, 1]
-                // Vcur shape [n_embd, N, 1, 1]
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wv, cur), n_embd, N)));
-
-                // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
-                // kv_self.v shape [n_embd * n_ctx * n_layer, 1]
-                // k         shape [n_embd * N, 1]   == kv_self.k[:,n_past:n_past+N,il,0]
-                // v         shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0]
-
-                /* {
-                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                    struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
-                            (   n_ctx)*ggml_element_size(kv_self.v),
-                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
-
-                    // important: storing RoPE-ed version of K in the KV cache!
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
-                } //*/
-
-                kc = ggml_set_1d_inplace(ctx0, kc, ggml_reshape_1d(ctx0, Kcur, n_embd*N), (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                vc = ggml_set_2d_inplace(ctx0, vc, Vcur, (   n_ctx)*ggml_element_size(kv_self.v),
-                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
-            }
-
-            // Qcur shape [n_embd/n_head, n_head, N, 1]
-            // Q shape    [n_embd/n_head, N, n_head, 1]
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        Qcur,
-                        0, 2, 1, 3);
-
-            // kv_self.k shape [n_embd * n_ctx * n_layer, 1]
-            // K shape [n_embd/n_head, n_past + N, n_head, 1]
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, kc, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kc)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        0, 2, 1, 3);
-
-            // K * Q
-            // KQ shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // KQ_masked shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            // KQ_soft_max shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
-
-            // split cached V into n_head heads
-            //// V shape [n_past + N, n_embd/n_head, n_head, 1]
-            // V shape [n_past + N, n_embd/n_head, n_head, 1] == kv_self.v[:,:(n_past+N),il,1]
-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, vc,
-                        n_past + N, n_embd/n_head, n_head,
-                        n_ctx*ggml_element_size(vc),
-                        n_ctx*ggml_element_size(vc)*n_embd/n_head,
-                        il*n_ctx*ggml_element_size(vc)*n_embd);
-
-            // KQV shape [n_embd/n_head, N, n_head, 1]
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            // KQV_merged shape [n_embd/n_head, n_head, N, 1]
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-            // KQV_merged shape
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            // cur shape [n_embd,N,1,1]
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N);
-            // cur = ggml_cpy(ctx0,
-            //         KQV_merged,
-            //         ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-
-            // projection (no bias)
-            // cur shape [n_embd,N,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].wo,
-                    cur);
-        }
-
-        // lctx.use_buf(ctx0, 1);
-
-        // inpFF shape [n_embd,N,1,1]
-        struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
-
-        // feed-forward network
-        {
-            // norm
-            {
-                // cur shape [n_embd,N,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
-
-                // cur = ffn_norm*cur
-                // cur shape [n_embd,N,1,1]
-                cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
-                        cur);
-            }
-
-            // tmp shape [n_ff,N,1,1]
-            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
-                    model->layers[il].w3,
-                    cur);
-
-            // cur shape [n_ff,N,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w1,
-                    cur);
-
-            // SILU activation
-            // cur shape [n_ff,N,1,1]
-            cur = ggml_silu(ctx0, cur);
-
-            // cur shape [n_ff,N,1,1]
-            cur = ggml_mul(ctx0, cur, tmp);
-
-            // cur shape [n_embd,N,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w2,
-                    cur);
-        }
-
-        // cur shape [n_embd,N,1,1]
-        cur = ggml_add(ctx0, cur, inpFF);
-
-        // input for next layer
-        // inpL shape [n_embd,N,1,1]
-        inpL = cur;
-    }
-
-    // norm
-    {
-
-        // inpL shape [n_embd,N,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-
-        // inpL = norm*inpL
-        // inpL shape [n_embd,N,1,1]
-        inpL = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model->norm, inpL),
-                    inpL);
-
-        //embeddings = inpL;
-    }
-
-    // lm_head
-    // inpL shape [n_vocab,N,1,1]
-    inpL = ggml_mul_mat(ctx0, model->output, inpL);
-
-    // run the computation
-    ggml_build_forward_expand(gf, inpL);
-
-    return inpL;
-}
-
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
-    GGML_ASSERT(tensor->n_dims == 1);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-}
-
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
-    GGML_ASSERT(tensor->n_dims == 2);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-}
-
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
-    GGML_ASSERT(tensor->n_dims == 3);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-}
-
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
-    GGML_ASSERT(tensor->n_dims == 4);
-    GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == ne3);
-}
-
-struct ggml_tensor * forward_batch(
-        struct my_llama_model    * model,
-        struct my_llama_kv_cache * cache,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_past,
-        const  int              n_batch) {
-
-    const int N = n_tokens;
-
-    struct my_llama_kv_cache& kv_self = *cache;
-    const auto & hparams = model->hparams;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_vocab = hparams.n_vocab;
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_head  = hparams.n_head;
-    const int n_rot   = hparams.n_rot;
-    const int n_ff    = get_n_ff(&hparams);
-
-    struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
-    memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
-
-    struct ggml_tensor * kc = kv_self.k;
-    struct ggml_tensor * vc = kv_self.v;
-
-    // inpL shape [n_embd,N*n_batch,1]
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
-    assert_shape_2d(inpL, n_embd, N*n_batch);
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * inpSA = inpL;
-
-        struct ggml_tensor * cur;
-
-        // lctx.use_buf(ctx0, 0);
-
-        // norm
-        {
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-
-            // cur = attention_norm*cur
-            cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
-                        cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // self-attention
-        {
-            // compute Q and K and RoPE them
-            // wq   shape [n_embd, n_embd, 1, 1]
-            // wk   shape [n_embd, n_embd, 1, 1]
-            // Qcur shape [n_embd/n_head, n_head, N, n_batch]
-            // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
-            assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
-
-            // store key and value to memory
-            {
-                // compute the transposed [N, n_embd] V matrix
-                // wv   shape [n_embd, n_embd, 1, 1]
-                // Vcur shape [N, n_embd, n_batch, 1]
-                struct ggml_tensor * Vcur = ggml_cont(ctx0,
-                    ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_mul_mat(ctx0,
-                                model->layers[il].wv,
-                                cur),
-                        n_embd, N, n_batch),
-                        1, 0, 2, 3));
-                assert_shape_3d(Vcur, N, n_embd, n_batch);
-
-                // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
-                // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
-                // k         shape [n_embd * N, n_batch]   == kv_self.k[:,n_past:n_past+N,:,il]
-                // v         shape [N, n_embd, n_batch, 1] == kv_self.v[:,n_past:n_past+N,:,il]
-
-                /* {
-                    struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                    struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
-                            (   n_ctx)*ggml_element_size(kv_self.v),
-                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
-
-                    // important: storing RoPE-ed version of K in the KV cache!
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
-                    ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
-                } //*/
-
-                kc = ggml_set_2d_inplace(ctx0, kc,
-                        ggml_reshape_2d(ctx0, Kcur, n_embd*N, n_batch),
-                        ggml_element_size(kc)*n_embd*n_ctx,
-                        (ggml_element_size(kc)*n_embd)*(il*n_batch*n_ctx + n_past));
-                vc = ggml_set_2d_inplace(ctx0, vc,
-                        ggml_reshape_2d(ctx0, Vcur, N*n_embd, n_batch),
-                        ggml_element_size(vc)*n_ctx*n_embd,
-                        ggml_element_size(vc)*(n_past + il*n_embd*n_batch*n_ctx));
-
-                assert_shape_1d(kc, n_embd * n_ctx * n_batch * n_layer);
-                assert_shape_1d(vc, n_embd * n_ctx * n_batch * n_layer);
-            }
-
-            // Qcur shape [n_embd/n_head, n_head, N, n_batch]
-            // Q shape    [n_embd/n_head, N, n_head, n_batch]
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        Qcur,
-                        0, 2, 1, 3);
-            assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
-
-            // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
-            // K shape [n_embd/n_head, n_past + N, n_head, n_batch]
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_reshape_4d(ctx0,
-                            ggml_view_3d(ctx0,
-                                kc,
-                                n_embd,
-                                (n_past + N),
-                                n_batch,
-                                n_embd*ggml_element_size(kc),
-                                n_ctx*n_embd*ggml_element_size(kc),
-                                il*n_batch*n_ctx*n_embd*ggml_element_size(kc)),
-                            n_embd/n_head, n_head, n_past + N, n_batch),
-                        0, 2, 1, 3);
-            assert_shape_4d(K, n_embd/n_head, n_past + N, n_head, n_batch);
-
-            // K * Q
-            // KQ shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-            assert_shape_4d(KQ, n_past + N, N, n_head, n_batch);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale_inplace(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
-            assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // KQ_masked shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
-            assert_shape_4d(KQ_masked, n_past + N, N, n_head, n_batch);
-
-            // KQ = soft_max(KQ_masked)
-            // KQ_soft_max shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
-            assert_shape_4d(KQ_soft_max, n_past + N, N, n_head, n_batch);
-
-            // split cached V into n_head heads
-            // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer]
-            // V shape [n_past + N, n_embd/n_head, n_head, n_batch] == kv_self.v[:(n_past+N),:,:,il]
-            struct ggml_tensor * V =
-                ggml_view_4d(ctx0, vc,
-                        n_past + N, n_embd/n_head, n_head, n_batch,
-                        ggml_element_size(vc)*n_ctx,
-                        ggml_element_size(vc)*n_ctx*n_embd/n_head,
-                        ggml_element_size(vc)*n_ctx*n_embd,
-                        il*n_batch*n_ctx*n_embd*ggml_element_size(vc));
-            assert_shape_4d(V, n_past + N, n_embd/n_head, n_head, n_batch);
-
-            // KQV shape [n_embd/n_head, N, n_head, n_batch]
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-            assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            // KQV_merged shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-            assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
-            // KQV_merged shape
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-            // cur = ggml_cpy(ctx0,
-            //         KQV_merged,
-            //         ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-
-            // projection (no bias)
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].wo,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // lctx.use_buf(ctx0, 1);
-
-        // inpFF shape [n_embd,N*n_batch,1,1]
-        struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
-        assert_shape_2d(inpFF, n_embd, N*n_batch);
-
-        // feed-forward network
-        {
-            // norm
-            {
-                // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-
-                // cur = ffn_norm*cur
-                // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
-                        cur);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-            }
-
-            // tmp shape [n_ff,N*n_batch,1,1]
-            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
-                    model->layers[il].w3,
-                    cur);
-            assert_shape_2d(tmp, n_ff, N*n_batch);
-
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w1,
-                    cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // SILU activation
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_silu(ctx0, cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_mul(ctx0, cur, tmp);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w2,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // cur shape [n_embd,N*n_batch,1,1]
-        cur = ggml_add_inplace(ctx0, cur, inpFF);
-        assert_shape_2d(cur, n_embd, N*n_batch);
-
-        // input for next layer
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = cur;
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-    }
-
-    // norm
-    {
-
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-
-        // inpL = norm*inpL
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model->norm, inpL),
-                    inpL);
-
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-
-        //embeddings = inpL;
-    }
-
-    // lm_head
-    // inpL shape [n_vocab,N*n_batch,1,1]
-    inpL = ggml_mul_mat(ctx0, model->output, inpL);
-    assert_shape_2d(inpL, n_vocab, N*n_batch);
-
-    {
-        // inpL shape [n_vocab,N,n_batch,1]
-        inpL = ggml_reshape_3d(ctx0,
-                        inpL,
-                        n_vocab, N, n_batch);
-        assert_shape_3d(inpL, n_vocab, N, n_batch);
-    }
-
-    // run the computation
-    ggml_build_forward_expand(gf, inpL);
-
-    return inpL;
-}
-
-struct ggml_tensor * forward_batch_wo_cache(
+static struct ggml_tensor * llama_build_train_graphs(
         struct my_llama_model * model,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_batch) {
-
-    const int n_past = 0;
-    const int N = n_tokens;
-
-    const auto & hparams = model->hparams;
-    //const int n_ctx   = hparams.n_ctx;
-    const int n_vocab = hparams.n_vocab;
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_head  = hparams.n_head;
-    const int n_rot   = hparams.n_rot;
-    const int n_ff    = get_n_ff(&hparams);
-
-    struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
-    memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
-
-    // inpL shape [n_embd,N*n_batch,1]
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
-    assert_shape_2d(inpL, n_embd, N*n_batch);
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * inpSA = inpL;
-
-        struct ggml_tensor * cur;
-
-        // lctx.use_buf(ctx0, 0);
-
-        // norm
-        {
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-
-            // cur = attention_norm*cur
-            cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
-                        cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // self-attention
-        {
-            // compute Q and K and RoPE them
-            // wq   shape [n_embd, n_embd, 1, 1]
-            // wk   shape [n_embd, n_embd, 1, 1]
-            // Qcur shape [n_embd/n_head, n_head, N, n_batch]
-            // Kcur shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
-            assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
-
-            // Vcur shape [N, n_batch, n_embd/n_head, n_head]
-            struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head);
-            assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head);
-
-            // Qcur shape [n_embd/n_head, n_head, N, n_batch]
-            // Q shape    [n_embd/n_head, N, n_head, n_batch]
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        Qcur,
-                        0, 2, 1, 3);
-            assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
-
-            // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer]
-            // K shape [n_embd/n_head, N, n_head, n_batch]
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        Kcur,
-                        0, 2, 1, 3);
-            assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch);
-
-            // K * Q
-            // KQ shape [N, N, n_head, n_batch]
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-            assert_shape_4d(KQ, N, N, n_head, n_batch);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // KQ_scaled shape [N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale_inplace(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
-            assert_shape_4d(KQ_scaled, N, N, n_head, n_batch);
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // KQ_masked shape [N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
-            assert_shape_4d(KQ_masked, N, N, n_head, n_batch);
-
-            // KQ = soft_max(KQ_masked)
-            // KQ_soft_max shape [N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
-            assert_shape_4d(KQ_soft_max, N, N, n_head, n_batch);
-
-            // Vcur shape [N, n_batch, n_embd/n_head, n_head]
-            // V shape    [N, n_embd/n_head, n_head, n_batch]
-            struct ggml_tensor * V =
-                ggml_permute(ctx0,
-                    Vcur,
-                    0, 3, 1, 2);
-            assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch);
-
-            // KQV shape [n_embd/n_head, N, n_head, n_batch]
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-            assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            // KQV_merged shape [n_embd/n_head, n_head, N, n_batch]
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-            assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
-            // KQV_merged shape
-
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-
-            // projection (no bias)
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].wo,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // lctx.use_buf(ctx0, 1);
-
-        // inpFF shape [n_embd,N*n_batch,1,1]
-        struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
-        assert_shape_2d(inpFF, n_embd, N*n_batch);
-
-        // feed-forward network
-        {
-            // norm
-            {
-                // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-
-                // cur = ffn_norm*cur
-                // cur shape [n_embd,N*n_batch,1,1]
-                cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
-                        cur);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-            }
-
-            // tmp shape [n_ff,N*n_batch,1,1]
-            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
-                    model->layers[il].w3,
-                    cur);
-            assert_shape_2d(tmp, n_ff, N*n_batch);
-
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w1,
-                    cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // SILU activation
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_silu(ctx0, cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // cur shape [n_ff,N*n_batch,1,1]
-            cur = ggml_mul(ctx0, cur, tmp);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // cur shape [n_embd,N*n_batch,1,1]
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w2,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // cur shape [n_embd,N*n_batch,1,1]
-        cur = ggml_add_inplace(ctx0, cur, inpFF);
-        assert_shape_2d(cur, n_embd, N*n_batch);
-
-        // input for next layer
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = cur;
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-    }
-
-    // norm
-    {
-
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-
-        // inpL = norm*inpL
-        // inpL shape [n_embd,N*n_batch,1,1]
-        inpL = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model->norm, inpL),
-                    inpL);
-
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-
-        //embeddings = inpL;
-    }
-
-    // lm_head
-    // inpL shape [n_vocab,N*n_batch,1,1]
-    inpL = ggml_mul_mat(ctx0, model->output, inpL);
-    assert_shape_2d(inpL, n_vocab, N*n_batch);
-
-    {
-        // inpL shape [n_vocab,N,n_batch,1]
-        inpL = ggml_reshape_3d(ctx0,
-                        inpL,
-                        n_vocab, N, n_batch);
-        assert_shape_3d(inpL, n_vocab, N, n_batch);
-    }
-
-    // run the computation
-    ggml_build_forward_expand(gf, inpL);
-
-    return inpL;
-}
-
-struct ggml_tensor * forward_batch_wo_cache_flash_attn(
-        struct my_llama_model * model,
-        struct ggml_context   * ctx0,
-        struct ggml_cgraph    * gf,
-        struct ggml_tensor    * tokens_input,
-        const  int              n_tokens,
-        const  int              n_batch) {
-
-    const int n_past = 0;
-    const int N = n_tokens;
-
-    const auto & hparams = model->hparams;
-    //const int n_ctx   = hparams.n_ctx;
-    const int n_vocab = hparams.n_vocab;
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_head  = hparams.n_head;
-    const int n_rot   = hparams.n_rot;
-    const int n_ff    = get_n_ff(&hparams);
-
-    struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch);
-    memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch);
-
-    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens);
-    assert_shape_2d(inpL, n_embd, N*n_batch);
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * inpSA = inpL;
-
-        struct ggml_tensor * cur;
-
-        // norm
-        {
-            cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-
-            // cur = attention_norm*cur
-            cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].attention_norm, cur),
-                        cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        // self-attention
-        {
-            // compute Q and K and RoPE them
-            // wq   shape [n_embd, n_embd, 1, 1]
-            // wk   shape [n_embd, n_embd, 1, 1]
-            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0);
-            assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch);
-            assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch);
-
-            struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head);
-            assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head);
-
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        Qcur,
-                        0, 2, 1, 3);
-            assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch);
-
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        Kcur,
-                        0, 2, 1, 3);
-            assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch);
-
-            struct ggml_tensor * V =
-                ggml_permute(ctx0,
-                    Vcur,
-                    0, 3, 1, 2);
-            assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch);
-
-            bool masked = true;
-            struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, masked);
-            assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch);
-
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-            assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch);
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-
-            // projection (no bias)
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].wo,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA);
-        assert_shape_2d(inpFF, n_embd, N*n_batch);
-
-        // feed-forward network
-        {
-            // norm
-            {
-                cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-
-                // cur = ffn_norm*cur
-                cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model->layers[il].ffn_norm, cur),
-                        cur);
-                assert_shape_2d(cur, n_embd, N*n_batch);
-            }
-
-            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
-                    model->layers[il].w3,
-                    cur);
-            assert_shape_2d(tmp, n_ff, N*n_batch);
-
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w1,
-                    cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            // SILU activation
-            cur = ggml_silu(ctx0, cur);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            cur = ggml_mul(ctx0, cur, tmp);
-            assert_shape_2d(cur, n_ff, N*n_batch);
-
-            cur = ggml_mul_mat(ctx0,
-                    model->layers[il].w2,
-                    cur);
-            assert_shape_2d(cur, n_embd, N*n_batch);
-        }
-
-        cur = ggml_add_inplace(ctx0, cur, inpFF);
-        assert_shape_2d(cur, n_embd, N*n_batch);
-
-        // input for next layer
-        inpL = cur;
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-    }
-
-    // norm
-    {
-
-        inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps);
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-
-        // inpL = norm*inpL
-        inpL = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model->norm, inpL),
-                    inpL);
-
-        assert_shape_2d(inpL, n_embd, N*n_batch);
-    }
-
-    // lm_head
-    inpL = ggml_mul_mat(ctx0, model->output, inpL);
-    assert_shape_2d(inpL, n_vocab, N*n_batch);
-
-    {
-        inpL = ggml_reshape_3d(ctx0,
-                        inpL,
-                        n_vocab, N, n_batch);
-        assert_shape_3d(inpL, n_vocab, N, n_batch);
-    }
-
-    // run the computation
-    ggml_build_forward_expand(gf, inpL);
-
-    return inpL;
-}
-
-// expand the graph nodes without creating leafs.
-struct ggml_tensor * expand(struct ggml_cgraph * g, struct ggml_tensor * t) {
-    // check if already visited
-    for (int i = 0; i < g->n_nodes; i++) {
-        if (g->nodes[i] == t) {
-            return t;
-        }
-    }
-
-    for (int i = 0; i < g->n_leafs; i++) {
-        if (g->leafs[i] == t) {
-            return t;
-        }
-    }
-
-    for (int i = 0; i < GGML_MAX_SRC; ++i) {
-        if (t->src[i]) {
-            expand(g, t->src[i]);
-        }
-    }
-
-    GGML_ASSERT(g->n_nodes < GGML_MAX_NODES);
-
-    if (strlen(t->name) == 0) {
-        snprintf(t->name, sizeof(t->name), "node_%d", g->n_nodes);
-    }
-
-    g->nodes[g->n_nodes] = t;
-    g->grads[g->n_nodes] = t->grad;
-    g->n_nodes++;
-    return t;
-}
-
-void graph_set_leafs_grads(struct ggml_cgraph * g) {
-    // moves leaf nodes to g->leafs.
-    // i.e. g->n_nodes might change.
-    int n_nodes = 0;
-    for (int i = 0; i < g->n_nodes; ++i) {
-        struct ggml_tensor * node = g->nodes[i];
-        const bool is_leaf = node->op == GGML_OP_NONE && node->grad == NULL;
-        if (is_leaf) {
-            GGML_ASSERT(g->n_leafs < GGML_MAX_NODES);
-
-            if (strlen(node->name) == 0) {
-                snprintf(node->name, sizeof(node->name), "leaf_%d", g->n_leafs);
-            }
-
-            g->leafs[g->n_leafs] = node;
-            g->n_leafs++;
-        } else {
-            GGML_ASSERT(n_nodes < GGML_MAX_NODES);
-
-            if (strlen(node->name) == 0) {
-                snprintf(node->name, sizeof(node->name), "node_%d", n_nodes);
-            }
-
-            g->nodes[n_nodes] = node;
-            g->grads[n_nodes] = node->grad;
-            n_nodes++;
-        }
-    }
-    for (int i=n_nodes; i < g->n_nodes; ++i) {
-        g->nodes[n_nodes] = NULL;
-        g->grads[n_nodes] = NULL;
-    }
-    g->n_nodes = n_nodes;
-}
-
-struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
-        struct my_llama_model * model,
-        struct ggml_context   * ctx0,
+        struct ggml_allocr    * alloc,
+        struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
+        struct ggml_cgraph    * gb_tmp,
         struct ggml_tensor  * * logits,
         struct ggml_tensor    * tokens_input,
         struct ggml_tensor    * targets,
-        void                  * compute_buf_0,
-        void                  * compute_buf_1,
-        size_t                  size_buf_0,
-        size_t                  size_buf_1,
         const  int              n_tokens,
-        const  int              n_batch) {
-
-    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+        const  int              n_batch,
+        const  bool             enable_flash_attn,
+        const  bool             enable_checkpointing) {
 
+    ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
     const int N = n_tokens;
-
-    gf->n_nodes = 0;
-    gf->n_leafs = 0;
-    gf->perf_runs = 0;
-    gf->perf_cycles = 0;
-    gf->perf_time_us = 0;
-
     const auto & hparams = model->hparams;
     const int n_ctx      = hparams.n_ctx;
     const int n_vocab    = hparams.n_vocab;
@@ -1445,1374 +320,515 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
     const int n_layer    = hparams.n_layer;
     const int n_head     = hparams.n_head;
     const int n_rot      = hparams.n_rot;
-    const int n_ff       = get_n_ff(&hparams);
-    const int rope_mode  = 0;
+    const int n_ff       = hparams.n_ff;
+    const float f_norm_rms_eps  = hparams.f_norm_rms_eps;
+    const float rope_freq_base  = hparams.rope_freq_base;
+    const float rope_freq_scale = hparams.rope_freq_scale;
 
-    int last_buf = -1;
-    size_t buf_offs[2] = { 0, 0 };
-    size_t buf_size[2] = { size_buf_0,
-                           size_buf_1 };
-    void * buf_data[2] = { compute_buf_0,
-                           compute_buf_1 };
-    auto use_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data] (int buf) {
-        size_t last_offs = 0;
-        last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, });
-        if (last_buf >= 0) {
-            buf_offs[last_buf] = last_offs;
-        }
-        if (buf >= 0) {
-            size_t offs = buf_offs[buf];
-            size_t size = buf_size[buf];
-            void * data = buf_data[buf];
-            ggml_set_scratch(ctx0, { offs, size, data, });
-        }
-        last_buf = buf;
-    };
-
-    bool track_max_mem = false;
-    size_t buf_maxs[2] = { 0, 0 };
-
-    auto clr_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data, &buf_maxs, track_max_mem] (int buf) {
-        if (buf < 0) return;
-        if (track_max_mem) {
-            size_t last_offs = 0;
-            last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, });
-            if (last_buf >= 0) {
-                buf_offs[last_buf] = last_offs;
-                buf_maxs[last_buf] = std::max(buf_maxs[last_buf], buf_offs[last_buf]);
-            }
-        }
-        buf_offs[buf] = 0;
-        if (track_max_mem && last_buf >= 0) {
-            size_t offs = buf_offs[last_buf];
-            size_t size = buf_size[last_buf];
-            void * data = buf_data[last_buf];
-            ggml_set_scratch(ctx0, { offs, size, data, });
+    auto set_name = [](struct ggml_tensor * t, const char * n) {
+        ggml_set_name(t, n);
+        if (t->grad) {
+            ggml_format_name(t->grad, "%s->grad", n);
         }
     };
 
-
-    auto view__q = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
-        int64_t ne0 = n_embd/n_head;
-        int64_t ne1 = N;
-        int64_t ne2 = n_head;
-        int64_t ne3 = n_batch;
-        size_t  nb0 = ggml_element_size(t);
-        size_t  nb1 = nb0*ne0;
-        size_t  nb2 = nb1*ne1;
-        size_t  nb3 = nb2*ne2;
-        size_t offset = 0;
-        return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
-    };
-
-    auto view__k = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
-        int64_t ne0 = n_embd/n_head;
-        int64_t ne1 = N;
-        int64_t ne2 = n_head;
-        int64_t ne3 = n_batch;
-        size_t  nb0 = ggml_element_size(t);
-        size_t  nb1 = nb0*ne0;
-        size_t  nb2 = nb1*ne1;
-        size_t  nb3 = nb2*ne2;
-        size_t offset = nb3*ne3;
-        return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
-    };
-
-    auto view__v = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * {
-        int64_t ne0 = N;
-        int64_t ne1 = n_embd/n_head;
-        int64_t ne2 = n_head;
-        int64_t ne3 = n_batch;
-        size_t  nb0 = ggml_element_size(t);
-        size_t  nb1 = nb0*ne0;
-        size_t  nb2 = nb1*ne1;
-        size_t  nb3 = nb2*ne2;
-        size_t offset = 2*nb3*ne3;
-        return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset);
-    };
-
-    auto add_or_set = [ctx0] (struct ggml_tensor * a, struct ggml_tensor * b) -> struct ggml_tensor * {
-        if (a == NULL) {
-            return b;
-        } else {
-            return ggml_add_inplace(ctx0, a, b);
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < N; ++i) {
+            data[i] = n_past + i;
         }
-    };
-
-    use_buf(-1);
-
-    model->tok_embeddings->grad    = NULL;
-    model->norm->grad              = NULL;
-    model->output->grad            = NULL;
-
-    for (int il = 0; il < n_layer; ++il) {
-        struct my_llama_layer & layer = model->layers[il];
-        layer.attention_norm->grad = NULL;
-        layer.wq->grad             = NULL;
-        layer.wk->grad             = NULL;
-        layer.wv->grad             = NULL;
-        layer.wo->grad             = NULL;
-        layer.ffn_norm->grad       = NULL;
-        layer.w1->grad             = NULL;
-        layer.w2->grad             = NULL;
-        layer.w3->grad             = NULL;
     }
 
-    clr_buf(0);
-    clr_buf(1);
+    // rope has so much parameters that we make a custom function for it
+    auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
+                (struct ggml_tensor * t) -> struct ggml_tensor * {
+        // not capturing these, to silcence warnings
+        const int rope_mode = 0;
 
-    use_buf(-1);
+        return ggml_rope_custom(ctx,
+            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            rope_freq_base, rope_freq_scale);
+    };
 
-    struct ggml_tensor * t00 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); assert_shape_1d(t00, N*n_batch);
-    memcpy(t00->data, tokens_input->data, ggml_element_size(t00)*N*n_batch);
+    set_name(tokens_input, "tokens_input");
+    set_name(targets,      "targets");
 
-    use_buf(-1);
-
-    struct ggml_tensor * t01 = expand(gf, ggml_get_rows(ctx0, model->tok_embeddings, t00)); assert_shape_2d(t01, n_embd, N*n_batch);
-
-    // need to remember these for the backward pass
-    std::vector<struct ggml_tensor *> t02L; t02L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t03L; t03L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t04L; t04L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t05L; t05L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t06L; t06L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t07L; t07L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t08L; t08L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t09L; t09L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t10L; t10L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t11L; t11L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t12L; t12L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t13L; t13L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t14L; t14L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t15L; t15L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t16L; t16L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t17L; t17L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t18L; t18L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t19L; t19L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t20L; t20L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t21L; t21L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t22L; t22L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t23L; t23L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t24L; t24L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t25L; t25L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t26L; t26L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t27L; t27L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t28L; t28L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t29L; t29L.resize(n_layer, NULL);
-    std::vector<struct ggml_tensor *> t30L; t30L.resize(n_layer, NULL);
+    GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
+    struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch);  set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch);
+    struct ggml_tensor * t01 = ggml_get_rows(ctx, model->tok_embeddings, t00); set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch);
 
     struct ggml_tensor * cur = t01;
 
+    std::vector<struct ggml_tensor *> checkpoints;
+    checkpoints.push_back(tokens_input);
+    checkpoints.push_back(targets);
+    checkpoints.push_back(t00);
+    checkpoints.push_back(t01);
+
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
     for (int il = 0; il < n_layer; ++il) {
-        clr_buf(0);
         struct my_llama_layer & layer = model->layers[il];
-        // tensors with values necessary for backward pass are in persistent buf(-1)
-        // other tensors with buf(0) and buf(1) are only temporary needed, and their memory reused after layer is completed.
-        use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm     (ctx0, cur, rms_norm_eps));                      assert_shape_2d(t02, n_embd, N*n_batch);
-        use_buf( 0); struct ggml_tensor * t03 = expand(gf, ggml_repeat       (ctx0, layer.attention_norm, t02));              assert_shape_2d(t03, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t04 = expand(gf, ggml_mul          (ctx0, t02, t03));                               assert_shape_2d(t04, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t05 = expand(gf, ggml_mul_mat      (ctx0, layer.wq, t04));                          assert_shape_2d(t05, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t06 = expand(gf, ggml_reshape_4d   (ctx0, t05, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t07 = expand(gf, ggml_rope_inplace (ctx0, t06, n_past, n_rot, rope_mode, 0));       assert_shape_4d(t07, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t08 = expand(gf, ggml_mul_mat      (ctx0, layer.wk, t04));                          assert_shape_2d(t08, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t09 = expand(gf, ggml_reshape_4d   (ctx0, t08, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t10 = expand(gf, ggml_rope_inplace (ctx0, t09, n_past, n_rot, rope_mode, 0));       assert_shape_4d(t10, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t11 = expand(gf, ggml_mul_mat      (ctx0, t04, layer.wv));                          assert_shape_2d(t11, N*n_batch, n_embd);
-        use_buf(-1); struct ggml_tensor * t12 = expand(gf, ggml_reshape_4d   (ctx0, t11, N, n_batch, n_embd/n_head, n_head)); assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head);
-        use_buf(-1); struct ggml_tensor * t13 = expand(gf, ggml_permute      (ctx0, t07, 0, 2, 1, 3));                        assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch);
-        use_buf(-1); struct ggml_tensor * t14 = expand(gf, ggml_permute      (ctx0, t10, 0, 2, 1, 3));                        assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch);
-        use_buf(-1); struct ggml_tensor * t15 = expand(gf, ggml_permute      (ctx0, t12, 0, 3, 1, 2));                        assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
-        use_buf(-1); struct ggml_tensor * t16 = expand(gf, ggml_flash_attn   (ctx0, t13, t14, t15, true));                    assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
-        use_buf( 0); struct ggml_tensor * t17 = expand(gf, ggml_permute      (ctx0, t16, 0, 2, 1, 3));                        assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t18 = expand(gf, ggml_cont         (ctx0, t17));                                    assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch);
-        use_buf(-1); struct ggml_tensor * t19 = expand(gf, ggml_reshape_2d   (ctx0, t18, n_embd, N*n_batch));                 assert_shape_2d(t19, n_embd, N*n_batch);
-        use_buf( 0); struct ggml_tensor * t20 = expand(gf, ggml_mul_mat      (ctx0, layer.wo, t19));                          assert_shape_2d(t20, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t21 = expand(gf, ggml_add          (ctx0, t20, cur));                               assert_shape_2d(t21, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm     (ctx0, t21, rms_norm_eps));                      assert_shape_2d(t22, n_embd, N*n_batch);
-        use_buf( 0); struct ggml_tensor * t23 = expand(gf, ggml_repeat       (ctx0, layer.ffn_norm, t22));                    assert_shape_2d(t23, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t24 = expand(gf, ggml_mul          (ctx0, t23, t22));                               assert_shape_2d(t24, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t25 = expand(gf, ggml_mul_mat      (ctx0, layer.w3, t24));                          assert_shape_2d(t25, n_ff, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t26 = expand(gf, ggml_mul_mat      (ctx0, layer.w1, t24));                          assert_shape_2d(t26, n_ff, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t27 = expand(gf, ggml_silu         (ctx0, t26));                                    assert_shape_2d(t27, n_ff, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t28 = expand(gf, ggml_mul          (ctx0, t27, t25));                               assert_shape_2d(t28, n_ff, N*n_batch);
-        use_buf( 0); struct ggml_tensor * t29 = expand(gf, ggml_mul_mat      (ctx0, layer.w2, t28));                          assert_shape_2d(t29, n_embd, N*n_batch);
-        use_buf(-1); struct ggml_tensor * t30 = expand(gf, ggml_add          (ctx0, t21, t29));                               assert_shape_2d(t30, n_embd, N*n_batch);
-        t02L[il] = t02;
-        t03L[il] = t03;
-        t04L[il] = t04;
-        t05L[il] = t05;
-        t06L[il] = t06;
-        t07L[il] = t07;
-        t08L[il] = t08;
-        t09L[il] = t09;
-        t10L[il] = t10;
-        t11L[il] = t11;
-        t12L[il] = t12;
-        t13L[il] = t13;
-        t14L[il] = t14;
-        t15L[il] = t15;
-        t16L[il] = t16;
-        t17L[il] = t17;
-        t18L[il] = t18;
-        t19L[il] = t19;
-        t20L[il] = t20;
-        t21L[il] = t21;
-        t22L[il] = t22;
-        t23L[il] = t23;
-        t24L[il] = t24;
-        t25L[il] = t25;
-        t26L[il] = t26;
-        t27L[il] = t27;
-        t28L[il] = t28;
-        t29L[il] = t29;
-        t30L[il] = t30;
-
-        cur      = t30;
-    }
-    clr_buf(0);
-    use_buf(0);
-    struct ggml_tensor * t31   = expand(gf, ggml_rms_norm  (ctx0, cur, rms_norm_eps));         assert_shape_2d(t31, n_embd, N*n_batch);
-    struct ggml_tensor * t32   = expand(gf, ggml_repeat    (ctx0, model->norm, t31));          assert_shape_2d(t32, n_embd, N*n_batch);
-    struct ggml_tensor * t33   = expand(gf, ggml_mul       (ctx0, t32, t31));                  assert_shape_2d(t33, n_embd, N*n_batch);
-    use_buf(-1);
-    struct ggml_tensor * t34   = expand(gf, ggml_mul_mat   (ctx0, model->output, t33));        assert_shape_2d(t34, n_vocab, N*n_batch);
-    struct ggml_tensor * t35   = expand(gf, ggml_reshape_3d(ctx0, t34, n_vocab, N, n_batch));  assert_shape_3d(t35, n_vocab, N, n_batch);
-    struct ggml_tensor * t36   = expand(gf, ggml_cross_entropy_loss(ctx0, t35, targets));      assert_shape_1d(t36, 1);
-
-    {
-        /*
-        tok_embeddings                                                        | grad_tok_embeddings = ggml_get_rows_back(grad_t01, t00)
-        L0_att_norm                                                           | grad_L0_att_norm    = ggml_repeat_back(grad_t03L0, L0_att_norm.shape)
-        L0_wq                                                                 | grad_L0_wq          = ggml_out_prod(t04L0, grad_t05L0)
-        L0_wk                                                                 | grad_L0_wk          = ggml_out_prod(t04L0, grad_t08L0)
-        L0_wv                                                                 | grad_L0_wv          = ggml_out_prod(t04L0, ggml_transpose(grad_t11L0))
-        L0_wo                                                                 | grad_L0_wo          = ggml_out_prod(t19L0, grad_t20L0)
-        L0_ffn_norm                                                           | grad_L0_ffn_norm    = ggml_repeat_back(grad_t23L0, L0_ffn_norm.shape)
-        L0_w1                                                                 | grad_L0_w1          = ggml_out_prod(t24L0, grad_t26L0)
-        L0_w2                                                                 | grad_L0_w2          = ggml_out_prod(t28L0, grad_t29L0)
-        L0_w3                                                                 | grad_L0_w3          = ggml_out_prod(t24L0, grad_t25L0)
-        L1_att_norm                                                           | grad_L1_att_norm    = ggml_repeat_back(grad_t03L1, L1_att_norm.shape)
-        L1_wq                                                                 | grad_L1_wq          = ggml_out_prod(t04L1, grad_t05L1)
-        L1_wk                                                                 | grad_L1_wk          = ggml_out_prod(t04L1, grad_t08L1)
-        L1_wv                                                                 | grad_L1_wv          = ggml_out_prod(t04L1, ggml_transpose(grad_t11L1))
-        L1_wo                                                                 | grad_L1_wo          = ggml_out_prod(t19L1, grad_t20L1)
-        L1_ffn_norm                                                           | grad_L1_ffn_norm    = ggml_repeat_back(grad_t23L1, L1_ffn_norm.shape)
-        L1_w1                                                                 | grad_L1_w1          = ggml_out_prod(t24L1, grad_t26L1)
-        L1_w2                                                                 | grad_L1_w2          = ggml_out_prod(t28L1, grad_t29L1)
-        L1_w3                                                                 | grad_L1_w3          = ggml_out_prod(t24L1, grad_t25L1)
-        norm                                                                  | grad_norm           = ggml_repeat_back(grad_t32, norm.shape)
-        output                                                                | grad_output         = ggml_out_prod(t33, grad_t34)
-                                                                              |
-        t01 = ggml_get_rows(tok_embeddings, t00)                              | grad_t01   = grad_t21L0 + ggml_rms_norm_back(t01, grad_t02L0)
-        for layer:                                                            |
-        t02L0*= ggml_rms_norm     (t01)                                       | grad_t02L0 = ggml_mul(grad_t04L0, t03L0)
-        t03L0 = ggml_repeat       (L0_att_norm, t02L0_shape)                  | grad_t03L0 = ggml_mul(grad_t04L0, t02L0)
-        t04L0*= ggml_mul          (t02L0, t03L0)                              | grad_t04L0 = ggml_out_prod(L0_wv, grad_t11L0) + ggml_out_prod(L0_wk, ggml_transpose(grad_t08L0)) + ggml_out_prod(L0_wq, ggml_transpose(grad_t05L0))
-        t05L0 = ggml_mul_mat      (L0_wq, t04L0)                              | grad_t05L0 = ggml_reshape(grad_t06L0, t05L0_shape)
-        t06L0 = ggml_reshape_4d   (t05L0, n_embd/n_head, n_head, N, n_batch)  | grad_t06L0 = ggml_rope_back(grad_t07L0)
-        t07L0 = ggml_rope_inplace (t06L0)                                     | grad_t07L0 = ggml_permute_back(grad_t13L0, 0, 2, 1, 3) = ggml_permute(grad_t13L0, 0, 2, 1, 3)
-        t08L0 = ggml_mul_mat      (L0_wk, t04L0)                              | grad_t08L0 = ggml_reshape(grad_t09L0, t08L0_shape)
-        t09L0 = ggml_reshape_4d   (t08L0, n_embd/n_head, n_head, N, n_batch)  | grad_t09L0 = ggml_rope_back(grad_t10L0)
-        t10L0 = ggml_rope_inplace (t09L0)                                     | grad_t10L0 = ggml_permute_back(grad_t14L0, 0, 2, 1, 3) = ggml_permute(grad_t14L0, 0, 2, 1, 3)
-        t11L0 = ggml_mul_mat      (t04L0, L0_wv)                              | grad_t11L0 = ggml_reshape(grad_t12L0, t11L0_shape)
-        t12L0 = ggml_reshape_4d   (t11L0, N, n_batch, n_embd/n_head, n_head)  | grad_t12L0 = ggml_permute_back(grad_t15L0, 0, 3, 1, 2) = ggml_permute(grad_t15L0, 0, 2, 3, 1)
-        t13L0*= ggml_permute      (t07L0, 0, 2, 1, 3)                         | grad_t13L0 = view__q(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
-        t14L0*= ggml_permute      (t10L0, 0, 2, 1, 3)                         | grad_t14L0 = view__k(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
-        t15L0*= ggml_permute      (t12L0, 0, 3, 1, 2)                         | grad_t15L0 = view__v(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0))
-        t16L0 = ggml_flash_attn   (t13L0, t14L0, t15L0)                       | grad_t16L0 = ggml_permute_back(grad_t17L0, 0, 2, 1, 3) = ggml_permute(grad_t17L0, 0, 2, 1, 3)
-        t17L0 = ggml_permute      (t16L0, 0, 2, 1, 3)                         | grad_t17L0 = grad_t18L0
-        t18L0 = ggml_cont         (t17L0)                                     | grad_t18L0 = ggml_reshape(grad_t19L0, t18L0_shape)
-        t19L0*= ggml_reshape_2d   (t18L0, n_embd, N*n_batch)                  | grad_t19L0 = ggml_out_prod(L0_wo, ggml_transpose(grad_t20L0))
-        t20L0 = ggml_mul_mat      (L0_wo, t19L0)                              | grad_t20L0 = grad_t21L0
-        t21L0*= ggml_add          (t20L0, t01)                                | grad_t21L0 = grad_t30L0 + ggml_rms_norm_back(t21L0, grad_t22L0)
-        t22L0*= ggml_rms_norm     (t21L0)                                     | grad_t22L0 = ggml_mul(grad_t24L0, t23L0)
-        t23L0 = ggml_repeat       (L0_ffn_norm, t22L0_shape)                  | grad_t23L0 = ggml_mul(grad_t24L0, t22L0)
-        t24L0*= ggml_mul          (t23L0, t22L0)                              | grad_t24L0 = ggml_out_prod(L0_w1, ggml_transpose(grad_t26L0)) + ggml_out_prod(L0_w3, ggml_transpose(grad_t25L0))
-        t25L0*= ggml_mul_mat      (L0_w3, t24L0)                              | grad_t25L0 = ggml_mul(grad_t28L0, t27L0)
-        t26L0*= ggml_mul_mat      (L0_w1, t24L0)                              | grad_t26L0 = ggml_silu_back(t26L0, grad_t27L0)
-        t27L0*= ggml_silu         (t26L0)                                     | grad_t27L0 = ggml_mul(grad_t28L0, t25L0)
-        t28L0*= ggml_mul          (t27L0, t25L0)                              | grad_t28L0 = ggml_out_prod(L0_w2, ggml_transpose(grad_t29L0))
-        t29L0 = ggml_mul_mat      (L0_w2, t28L0)                              | grad_t29L0 = grad_t30L0
-        t30L0*= ggml_add          (t21L0, t29L0)                              | grad_t30L0 = ggml_rms_norm_back(t30L0, grad_t02L1) + grad_t21L1
-                                                                              ^
-        t02L1*= ggml_rms_norm     (t30L0)                                     | grad_t02L1 = ggml_mul(grad_t04L1, t03L1)
-        t03L1 = ggml_repeat       (L1_att_norm, t02L1_shape)                  | grad_t03L1 = ggml_mul(grad_t04L1, t02L1)
-        t04L1*= ggml_mul          (t02L1, t03L1)                              | grad_t04L1 = ggml_out_prod(L1_wv, grad_t11L1) + ggml_out_prod(L1_wk, ggml_transpose(grad_t08L1)) + ggml_out_prod(L1_wq, ggml_transpose(grad_t05L1))
-        t05L1 = ggml_mul_mat      (L1_wq, t04L1)                              | grad_t05L1 = ggml_reshape(grad_t06L1, t05L1_shape)
-        t06L1 = ggml_reshape_4d   (t05L1, n_embd/n_head, n_head, N, n_batch)  | grad_t06L1 = ggml_rope_back(grad_t07L1)
-        t07L1 = ggml_rope_inplace (t06L1)                                     | grad_t07L1 = ggml_permute_back(grad_t13L1, 0, 2, 1, 3) = ggml_permute(grad_t13L1, 0, 2, 1, 3)
-        t08L1 = ggml_mul_mat      (L1_wk, t04L1)                              | grad_t08L1 = ggml_reshape(grad_t09L1, t08L1_shape)
-        t09L1 = ggml_reshape_4d   (t08L1, n_embd/n_head, n_head, N, n_batch)  | grad_t09L1 = ggml_rope_back(grad_t10L1)
-        t10L1 = ggml_rope_inplace (t09L1)                                     | grad_t10L1 = ggml_permute_back(grad_t14L1, 0, 2, 1, 3) = ggml_permute(grad_t14L1, 0, 2, 1, 3)
-        t11L1 = ggml_mul_mat      (t04L1, L1_wv)                              | grad_t11L1 = ggml_reshape(grad_t12L1, t11L1_shape)
-        t12L1 = ggml_reshape_4d   (t11L1, N, n_batch, n_embd/n_head, n_head)  | grad_t12L1 = ggml_permute_back(grad_t15L1, 0, 3, 1, 2) = ggml_permute(grad_t15L1, 0, 2, 3, 1)
-        t13L1*= ggml_permute      (t07L1, 0, 2, 1, 3)                         | grad_t13L1 = view__q(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
-        t14L1*= ggml_permute      (t10L1, 0, 2, 1, 3)                         | grad_t14L1 = view__k(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
-        t15L1*= ggml_permute      (t12L1, 0, 3, 1, 2)                         | grad_t15L1 = view__v(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1))
-        t16L1 = ggml_flash_attn   (t13L1, t14L1, t15L1)                       | grad_t16L1 = ggml_permute_back(grad_t17L1, 0, 2, 1, 3) = ggml_permute(grad_t17L1, 0, 2, 1, 3)
-        t17L1 = ggml_permute      (t16L1, 0, 2, 1, 3)                         | grad_t17L1 = grad_t18L1
-        t18L1 = ggml_cont         (t17L1)                                     | grad_t18L1 = ggml_reshape(grad_t19L1, t18L1_shape)
-        t19L1*= ggml_reshape_2d   (t18L1, n_embd, N*n_batch)                  | grad_t19L1 = ggml_out_prod(L1_wo, ggml_transpose(grad_t20L1))
-        t20L1 = ggml_mul_mat      (L1_wo, t19L1)                              | grad_t20L1 = grad_t21L1
-        t21L1*= ggml_add          (t20L1, t30L0)                              | grad_t21L1 = grad_t30L1 + ggml_rms_norm_back(t21L1, grad_t22L1)
-        t22L1*= ggml_rms_norm     (t21L1)                                     | grad_t22L1 = ggml_mul(grad_t24L1, t23L1)
-        t23L1 = ggml_repeat       (L1_ffn_norm, t22L1_shape)                  | grad_t23L1 = ggml_mul(grad_t24L1, t22L1)
-        t24L1*= ggml_mul          (t23L1, t22L1)                              | grad_t24L1 = ggml_out_prod(L1_w1, ggml_transpose(grad_t26L1)) + ggml_out_prod(L1_w3, ggml_transpose(grad_t25L1))
-        t25L1*= ggml_mul_mat      (L1_w3, t24L1)                              | grad_t25L1 = ggml_mul(grad_t28L1, t27L1)
-        t26L1*= ggml_mul_mat      (L1_w1, t24L1)                              | grad_t26L1 = ggml_silu_back(t26L1, grad_t27L1)
-        t27L1*= ggml_silu         (t26L1)                                     | grad_t27L1 = ggml_mul(grad_t28L1, t25L1)
-        t28L1*= ggml_mul          (t27L1, t25L1)                              | grad_t28L1 = ggml_out_prod(L1_w2, ggml_transpose(grad_t29L1))
-        t29L1 = ggml_mul_mat      (L1_w2, t28L1)                              | grad_t29L1 = grad_t30L1
-        t30L1*= ggml_add          (t21L1, t29L1)                              | grad_t30L1 = ggml_rms_norm_back(t30L1, grad_t31)
-                                                                              ^
-        t31   = ggml_rms_norm     (t30L1)                                     | grad_t31   = ggml_mul(grad_t33, t32)
-        t32   = ggml_repeat       (norm, t31.shape)                           | grad_t32   = ggml_mul(grad_t33, t31)
-        t33   = ggml_mul          (t32, t31)                                  | grad_t33   = ggml_out_prod(output, ggml_transpose(grad_t34))
-        t34   = ggml_mul_mat      (output, t33)                               | grad_t34   = ggml_reshape(grad_t35, t34.shape)
-        t35   = ggml_reshape_3d   (t34, n_vocab, N, n_batch)                  | grad_t35   = ggml_cross_entropy_loss_back(t35, targets, grad_t36)
-        t36   = ggml_cross_entropy_loss(t35, targets)                         | grad_t36   = 1 (optimizer)
-        tensors marked with * need to be stored until grad computation
-        tensors during grad computation are all temporary
-        */
-    }
-
-    *gb = *gf;
-
-    // t36->grad gets set to one by optimizer, so we need the tensor.
-    // initialize it with 1.0f to make sure.
-    use_buf(-1);
-    t36->grad = expand(gb, ggml_new_f32(ctx0, 1.0f));
-
-    use_buf(0);
-    t35->grad = expand(gb, ggml_cross_entropy_loss_back(ctx0, t35, targets, t36->grad));              assert_shape_3d(t35->grad, n_vocab, N, n_batch);
-    t34->grad = expand(gb, ggml_reshape_2d (ctx0, t35->grad, n_vocab, N*n_batch));                    assert_shape_2d(t34->grad, n_vocab, N*n_batch);
-    t33->grad = expand(gb, ggml_out_prod   (ctx0, model->output, ggml_transpose(ctx0, t34->grad)));   assert_shape_2d(t33->grad, n_embd, N*n_batch);
-    t32->grad = expand(gb, ggml_mul        (ctx0, t33->grad, t31));                                   assert_shape_2d(t32->grad, n_embd, N*n_batch);
-
-    use_buf(-1);
-
-    model->norm->grad   = expand(gb, add_or_set(model->norm->grad,   ggml_repeat_back(ctx0, t32->grad, model->norm))); assert_shape_1d(model->norm->grad, n_embd);
-    model->output->grad = expand(gb, add_or_set(model->output->grad, ggml_out_prod(ctx0, t33, t34->grad)));            assert_shape_2d(model->output->grad, n_embd, n_vocab);
-
-    clr_buf(1);
-    use_buf(1);
-    t31->grad = expand(gb, ggml_mul(ctx0, t33->grad, t32));  assert_shape_2d(t31->grad, n_embd, N*n_batch);
-
-    struct ggml_tensor * back_layer_inp = t31;
-    struct ggml_tensor * grad_layer_inp = NULL;
-
-    for (int k = 0; k < n_layer; ++k) {
-        int il = n_layer-1-k;
-        struct my_llama_layer & layer = model->layers[il];
-
-        struct ggml_tensor * t02 = t02L[il];
-        struct ggml_tensor * t03 = t03L[il];
-        struct ggml_tensor * t04 = t04L[il];
-        struct ggml_tensor * t05 = t05L[il];
-        struct ggml_tensor * t06 = t06L[il];
-        struct ggml_tensor * t07 = t07L[il];
-        struct ggml_tensor * t08 = t08L[il];
-        struct ggml_tensor * t09 = t09L[il];
-        struct ggml_tensor * t10 = t10L[il];
-        struct ggml_tensor * t11 = t11L[il];
-        struct ggml_tensor * t12 = t12L[il];
-        struct ggml_tensor * t13 = t13L[il];
-        struct ggml_tensor * t14 = t14L[il];
-        struct ggml_tensor * t15 = t15L[il];
-        struct ggml_tensor * t16 = t16L[il];
-        struct ggml_tensor * t17 = t17L[il];
-        struct ggml_tensor * t18 = t18L[il];
-        struct ggml_tensor * t19 = t19L[il];
-        struct ggml_tensor * t20 = t20L[il];
-        struct ggml_tensor * t21 = t21L[il];
-        struct ggml_tensor * t22 = t22L[il];
-        struct ggml_tensor * t23 = t23L[il];
-        struct ggml_tensor * t24 = t24L[il];
-        struct ggml_tensor * t25 = t25L[il];
-        struct ggml_tensor * t26 = t26L[il];
-        struct ggml_tensor * t27 = t27L[il];
-        struct ggml_tensor * t28 = t28L[il];
-        struct ggml_tensor * t29 = t29L[il];
-        struct ggml_tensor * t30 = t30L[il];
-
-        clr_buf(0);
-        use_buf(0);
-        t30->grad = expand(gb, ggml_rms_norm_back(ctx0, t30, back_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch);
-        if (grad_layer_inp) {
-            t30->grad = expand(gb, ggml_add(ctx0, t30->grad, grad_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch);
+        struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, f_norm_rms_eps);                    set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
+        struct ggml_tensor * t03 = ggml_repeat       (ctx, layer.attention_norm, t02);              set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
+        struct ggml_tensor * t04 = ggml_mul          (ctx, t03, t02);                               set_name(t04, "t04");     assert_shape_2d(t04, n_embd, N*n_batch);
+        struct ggml_tensor * t05 = ggml_mul_mat      (ctx, layer.wq, t04);                          set_name(t05, "t05");     assert_shape_2d(t05, n_embd, N*n_batch);
+        struct ggml_tensor * t06 = ggml_reshape_4d   (ctx, t05, n_embd/n_head, n_head, N, n_batch); set_name(t06, "t06");     assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t07 = rope              (t06);                                         set_name(t07, "t07");     assert_shape_4d(t07, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t08 = ggml_mul_mat      (ctx, layer.wk, t04);                          set_name(t08, "t08");     assert_shape_2d(t08, n_embd, N*n_batch);
+        struct ggml_tensor * t09 = ggml_reshape_4d   (ctx, t08, n_embd/n_head, n_head, N, n_batch); set_name(t09, "t09");     assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t10 = rope              (t09);                                         set_name(t10, "t10");     assert_shape_4d(t10, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t11 = ggml_mul_mat      (ctx, t04, layer.wv);                          set_name(t11, "t11");     assert_shape_2d(t11, N*n_batch, n_embd);
+        struct ggml_tensor * t12 = ggml_reshape_4d   (ctx, t11, N, n_batch, n_embd/n_head, n_head); set_name(t12, "t12");     assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head);
+        struct ggml_tensor * t13 = ggml_permute      (ctx, t07, 0, 2, 1, 3);                        set_name(t13, "t13");     assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch);
+        struct ggml_tensor * t14 = ggml_permute      (ctx, t10, 0, 2, 1, 3);                        set_name(t14, "t14");     assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch);
+        struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                        set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
+        struct ggml_tensor * t16;
+        if (enable_flash_attn) {
+            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
+        } else {
+            struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                 set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);          set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past);            set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch);
+            struct ggml_tensor * t16_3 = ggml_soft_max_inplace     (ctx, t16_2);                    set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch);
+            t16 = ggml_mul_mat(ctx, t15, t16_3);                                                    set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
         }
-        clr_buf(1);
-        t29->grad = t30->grad;                                                                                        assert_shape_2d(t29->grad, n_embd, N*n_batch);
-        t28->grad = expand(gb, ggml_out_prod(ctx0, layer.w2, ggml_transpose(ctx0, t29->grad)));                       assert_shape_2d(t28->grad, n_ff, N*n_batch);
-        t27->grad = expand(gb, ggml_mul(ctx0, t28->grad, t25));                                                       assert_shape_2d(t27->grad, n_ff, N*n_batch);
-        t26->grad = expand(gb, ggml_silu_back(ctx0, t26, t27->grad));                                                 assert_shape_2d(t26->grad, n_ff, N*n_batch);
-        t25->grad = expand(gb, ggml_mul(ctx0, t28->grad, t27));                                                       assert_shape_2d(t25->grad, n_ff, N*n_batch);
-        t24->grad = expand(gb, ggml_add_inplace(ctx0,
-                        ggml_out_prod(ctx0, layer.w1, ggml_transpose(ctx0, t26->grad)),
-                        ggml_out_prod(ctx0, layer.w3, ggml_transpose(ctx0, t25->grad))));                             assert_shape_2d(t24->grad, n_embd, N*n_batch);
-        t23->grad = expand(gb, ggml_mul(ctx0, t24->grad, t22));                                                       assert_shape_2d(t23->grad, n_embd, N*n_batch);
-        t22->grad = expand(gb, ggml_mul(ctx0, t24->grad, ggml_repeat(ctx0, layer.ffn_norm, t24->grad)));              assert_shape_2d(t22->grad, n_embd, N*n_batch);
-        use_buf(1);
-        t21->grad = expand(gb, ggml_add(ctx0, t30->grad, ggml_rms_norm_back(ctx0, t21, t22->grad)));                  assert_shape_2d(t21->grad, n_embd, N*n_batch);
-        grad_layer_inp = t21;
-        use_buf(0);
-        t20->grad = t21->grad;                                                                                        assert_shape_2d(t20->grad, n_embd, N*n_batch);
-        t19->grad = expand(gb, ggml_out_prod(ctx0, layer.wo, ggml_transpose(ctx0, t20->grad)));                       assert_shape_2d(t19->grad, n_embd, N*n_batch);
-        t18->grad = expand(gb, ggml_reshape_4d(ctx0, t19->grad, n_embd/n_head, n_head, N, n_batch));                  assert_shape_4d(t18->grad, n_embd/n_head, n_head, N, n_batch);
-        t17->grad = t18->grad;                                                                                        assert_shape_4d(t17->grad, n_embd/n_head, n_head, N, n_batch);
-        t16->grad = expand(gb, ggml_permute(ctx0, t17->grad, 0, 2, 1, 3));                                            assert_shape_4d(t16->grad, n_embd/n_head, N, n_head, n_batch);
-        struct ggml_tensor * flash_attn = expand(gb, ggml_flash_attn_back(ctx0, t13, t14, t15, t16->grad, true));     assert_shape_4d(flash_attn, n_embd/n_head, N*3, n_head, n_batch);
-        t15->grad = expand(gb, view__v(flash_attn));                                                                  assert_shape_4d(t15->grad, N, n_embd/n_head, n_head, n_batch);
-        t14->grad = expand(gb, view__k(flash_attn));                                                                  assert_shape_4d(t14->grad, n_embd/n_head, N, n_head, n_batch);
-        t13->grad = expand(gb, view__q(flash_attn));                                                                  assert_shape_4d(t13->grad, n_embd/n_head, N, n_head, n_batch);
-        t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1));                                            assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
-        t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd));                 assert_shape_2d(t11->grad, N*n_batch, n_embd);
-        t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3));                                            assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
-        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false));        assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
-        t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch));                                  assert_shape_2d(t08->grad, n_embd, N*n_batch);
-        t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3));                                            assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
-        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false));        assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
-        t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch));                                  assert_shape_2d(t05->grad, n_embd, N*n_batch);
-        t04->grad = expand(gb, ggml_add_inplace(ctx0,
-                        ggml_add_inplace(ctx0,
-                            ggml_out_prod(ctx0, layer.wv, t11->grad),
-                            ggml_out_prod(ctx0, layer.wk, ggml_transpose(ctx0, t08->grad))),
-                        ggml_out_prod(ctx0, layer.wq, ggml_transpose(ctx0, t05->grad))));                             assert_shape_2d(t04->grad, n_embd, N*n_batch);
-        t03->grad = expand(gb, ggml_mul(ctx0, t04->grad, t02));                                                       assert_shape_2d(t04->grad, n_embd, N*n_batch);
-        use_buf(1);
-        t02->grad = expand(gb, ggml_mul(ctx0, t04->grad, ggml_repeat(ctx0, layer.attention_norm, t02)));              assert_shape_2d(t02->grad, n_embd, N*n_batch);
-        back_layer_inp = t02;
-        // use_buf(0);
-
-        use_buf(-1);
-        layer.attention_norm->grad = expand(gb, add_or_set(layer.attention_norm->grad, ggml_repeat_back(ctx0, t03->grad, layer.attention_norm)));   assert_shape_1d(layer.attention_norm->grad, n_embd);
-        layer.wq->grad             = expand(gb, add_or_set(layer.wq->grad,             ggml_out_prod(ctx0, t04, t05->grad)));                       assert_shape_2d(layer.wq->grad,             n_embd, n_embd);
-        layer.wk->grad             = expand(gb, add_or_set(layer.wk->grad,             ggml_out_prod(ctx0, t04, t08->grad)));                       assert_shape_2d(layer.wk->grad,             n_embd, n_embd);
-        layer.wv->grad             = expand(gb, add_or_set(layer.wv->grad,             ggml_out_prod(ctx0, t04, ggml_transpose(ctx0, t11->grad)))); assert_shape_2d(layer.wv->grad,             n_embd, n_embd);
-        layer.wo->grad             = expand(gb, add_or_set(layer.wo->grad,             ggml_out_prod(ctx0, t19, t20->grad)));                       assert_shape_2d(layer.wo->grad,             n_embd, n_embd);
-        layer.ffn_norm->grad       = expand(gb, add_or_set(layer.ffn_norm->grad,       ggml_repeat_back(ctx0, t23->grad, layer.ffn_norm)));         assert_shape_1d(layer.ffn_norm->grad,       n_embd);
-        layer.w1->grad             = expand(gb, add_or_set(layer.w1->grad,             ggml_out_prod(ctx0, t24, t26->grad)));                       assert_shape_2d(layer.w1->grad,             n_embd, n_ff);
-        layer.w2->grad             = expand(gb, add_or_set(layer.w2->grad,             ggml_out_prod(ctx0, t28, t29->grad)));                       assert_shape_2d(layer.w2->grad,             n_ff, n_embd);
-        layer.w3->grad             = expand(gb, add_or_set(layer.w3->grad,             ggml_out_prod(ctx0, t24, t25->grad)));                       assert_shape_2d(layer.w3->grad,             n_embd, n_ff);
-        // use_buf(0);
+        struct ggml_tensor * t17 = ggml_permute      (ctx, t16, 0, 2, 1, 3);                        set_name(t17, "t17");     assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t18 = ggml_cont         (ctx, t17);                                    set_name(t18, "t18");     assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch);
+        struct ggml_tensor * t19 = ggml_reshape_2d   (ctx, t18, n_embd, N*n_batch);                 set_name(t19, "t19");     assert_shape_2d(t19, n_embd, N*n_batch);
+        struct ggml_tensor * t20 = ggml_mul_mat      (ctx, layer.wo, t19);                          set_name(t20, "t20");     assert_shape_2d(t20, n_embd, N*n_batch);
+        struct ggml_tensor * t21 = ggml_add          (ctx, t20, cur);                               set_name(t21, "t21");     assert_shape_2d(t21, n_embd, N*n_batch);
+        struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, f_norm_rms_eps);                    set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
+        struct ggml_tensor * t23 = ggml_repeat       (ctx, layer.ffn_norm, t22);                    set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
+        struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                               set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.w3, t24);                          set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.w1, t24);                          set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                    set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
+        struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                               set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.w2, t28);                          set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                               set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
+        cur = t30;
+        checkpoints.push_back(cur);
+    }
+    struct ggml_tensor * t31   = ggml_rms_norm          (ctx, cur, f_norm_rms_eps);                 set_name(t31, "t31");     assert_shape_2d(t31, n_embd, N*n_batch);
+    struct ggml_tensor * t32   = ggml_repeat            (ctx, model->norm, t31);                    set_name(t32, "t32");     assert_shape_2d(t32, n_embd, N*n_batch);
+    struct ggml_tensor * t33   = ggml_mul               (ctx, t32, t31);                            set_name(t33, "t33");     assert_shape_2d(t33, n_embd, N*n_batch);
+    struct ggml_tensor * t34   = ggml_mul_mat           (ctx, model->output, t33);                  set_name(t34, "t34");     assert_shape_2d(t34, n_vocab, N*n_batch);
+    struct ggml_tensor * t35   = ggml_reshape_3d        (ctx, t34, n_vocab, N, n_batch);            set_name(t35, "t35");     assert_shape_3d(t35, n_vocab, N, n_batch);
+    struct ggml_tensor * t36   = ggml_cross_entropy_loss(ctx, t35, targets);                        set_name(t36, "t36");     assert_shape_1d(t36, 1);
+
+    checkpoints.push_back(t31);
+    checkpoints.push_back(t32);
+    checkpoints.push_back(t33);
+    checkpoints.push_back(t34);
+    checkpoints.push_back(t35);
+    checkpoints.push_back(t36);
+
+    ggml_build_forward_expand(gf, t36);
+
+    if (enable_checkpointing) {
+        ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
+    } else {
+        *gb = *gf;
+        ggml_build_backward_expand(ctx, gf, gb, true);
+    }
+
+    if (alloc) {
+        // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
+        int n_leafs_before = gb->n_leafs;
+        int n_nodes_before = gb->n_nodes;
+        struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
+        // output tensors
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+        // input gradient
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+        // KQ_pos
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+        GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
+
+        ggml_allocr_alloc(alloc, t36->grad);
+
+        // allocating checkpoints in one block to reduce memory fragmentation
+        // note: they will be freed in reverse order
+        for (int i = 0; i < (int) checkpoints.size(); ++i) {
+            if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
+                ggml_allocr_alloc(alloc, checkpoints[i]);
+            }
+        }
+
+        //int n_leafs_after = gb->n_leafs;
+        //int n_nodes_after = gb->n_nodes;
+
+        ggml_allocr_alloc_graph(alloc, gb);
+
+        // remove the additional nodes and leafs
+        for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
+            gb->leafs[i] = NULL;
+        }
+        for (int i = n_nodes_before; i < gb->n_nodes; ++i) {
+            gb->nodes[i] = NULL;
+        }
+        gb->n_leafs = n_leafs_before;
+        gb->n_nodes = n_nodes_before;
     }
-    clr_buf(0);
-    use_buf(0);
-    t01->grad = expand(gb, ggml_add_inplace(ctx0, grad_layer_inp->grad, ggml_rms_norm_back(ctx0, t01, back_layer_inp->grad)));  assert_shape_2d(t01->grad, n_embd, N*n_batch);
-    use_buf(-1);
-    model->tok_embeddings->grad = expand(gb, ggml_get_rows_back(ctx0, t01->grad, t00, model->tok_embeddings));                  assert_shape_2d(model->tok_embeddings->grad, n_embd, n_vocab);
-    // clr_buf(1);
-    // clr_buf(0);
 
     *logits = t35;
-
-    if (track_max_mem) {
-        printf("%s: max size compute buf0: %zu\n", __func__, buf_maxs[0]);
-        printf("%s: max size compute buf1: %zu\n", __func__, buf_maxs[1]);
-    }
-
-    // now that all grads are created, set the graph leafs and grads
-    graph_set_leafs_grads(gf);
-    graph_set_leafs_grads(gb);
-
     return t36;
 }
 
-void set_f32_3d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, int64_t i2, float value) {
-    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2]);
-    *ptr = value;
-}
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+do { \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        die_fmt("key not found in model: %s", skey.c_str()); \
+    } \
+} while (0)
 
-void set_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, float value) {
-    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-    *ptr = value;
-}
+static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model) {
+    // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read
+    std::string arch;
 
-void set_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1, int32_t value) {
-    int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-    *ptr = value;
-}
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [&arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch.c_str());
+        return keybuf.data();
+    };
 
-float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
-    float * ptr = (float *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-    return *ptr;
-}
+    std::vector<char> tn_buf;
+    tn_buf.resize(GGML_MAX_NAME);
+    auto tn = [&tn_buf](const char * key) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key);
+        return tn_buf.data();
+    };
+    auto tni = [&tn_buf](const char * key, int bid) -> const char * {
+        snprintf(tn_buf.data(), tn_buf.size(), key, bid);
+        std::string s = tn_buf.data();
+        snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str());
+        return tn_buf.data();
+    };
 
-int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
-    int32_t * ptr = (int32_t *) ((char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1]);
-    return *ptr;
-}
+    GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE);
+    GGML_ASSERT(arch == "llama");
 
-void print_row(struct ggml_tensor * probs, int i) {
-    for (int k = 0; k < probs->ne[0]; ++k) {
-        float p = get_f32_2d(probs, k, i);
-        printf(" %.2f", p);
-    }
-    printf("\n");
-}
+    uint32_t ftype_u;
+    GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE);
+    GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32);
 
-void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
-    for (int i = 0; i < probs->ne[1]; ++i) {
-        for (int k = 0; k < probs->ne[0]; ++k) {
-            float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
-        }
-        printf("\n");
-    }
-}
+    // n_ctx was not saved in earlier checkpoint file versions, so we make it optional here
+    GGUF_GET_KEY(fctx, model->hparams.n_ctx,   gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH));
 
+    GGUF_GET_KEY(fctx, model->hparams.n_embd,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH));
+    GGUF_GET_KEY(fctx, model->hparams.n_ff,    gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH));
+    GGUF_GET_KEY(fctx, model->hparams.n_head,  gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
+    GGUF_GET_KEY(fctx, model->hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT));
 
-void print_token(struct llama_context * ctx, llama_token token) {
-    printf("%s", llama_token_to_str(ctx, token).c_str());
-}
+    model->hparams.n_rot = model->hparams.n_embd / model->hparams.n_head;
+    GGUF_GET_KEY(fctx, model->hparams.n_rot,   gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
 
-void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) {
-    for (int i=0; i<tokens->ne[0]; ++i) {
-        int token = ggml_get_i32_1d(tokens, i);
-        print_token(ctx, token);
-    }
-}
-
-void print_tokens_batch(struct llama_context* ctx, struct ggml_tensor * tokens) {
-    for (int i1=0; i1<tokens->ne[1]; ++i1) {
-        //int num_newline = 0;
-        for (int i0=0; i0<tokens->ne[0]; ++i0) {
-            int token = get_i32_2d(tokens, i0, i1);
-            print_token(ctx, token);
-            // bool isnl = (token == llama_token_nl());
-            // if (isnl) {
-            //     ++num_newline;
-            // }
-            // if (isnl) {
-            //     if (num_newline < 2) {
-            //         print_token(ctx, token);
-            //     } else {
-            //         printf("\\n");
-            //     }
-            // } else {
-            //     print_token(ctx, token);
-            // }
-        }
-        printf("\n--\n");
-    }
-}
-
-void get_example_targets(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
-    int n_tokens = tokens_input->ne[0];
-    int n_vocab  = target_logits->ne[0];
-
-    size_t sample = train_samples[example_id % n_train_samples];
-    GGML_ASSERT(sample+n_tokens-1 < n_train_data);
-
-    ggml_set_f32(target_logits, -1.0f/n_vocab);
-    ggml_set_f32(target_probs, 0.0f);
-    ggml_set_i32_1d(tokens_input, 0, llama_token_bos(lctx));
-    for (int i=1; i<n_tokens+1; ++i) {
-        int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
-        set_f32_2d(target_logits, token, i-1, +1.0f);
-        set_f32_2d(target_probs,  token, i-1, +1.0f);
-        if (i<n_tokens) {
-            ggml_set_i32_1d(tokens_input, i, token);
-        }
-    }
-}
-
-void get_example_targets_batch(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) {
-    GGML_ASSERT(tokens_input->n_dims  == 2);
-    GGML_ASSERT(target_logits->n_dims == 3);
-    GGML_ASSERT(target_probs->n_dims  == 3);
-    int n_vocab  = target_logits->ne[0];
-    int n_tokens = tokens_input->ne[0];
-    int n_batch  = tokens_input->ne[1];
-    GGML_ASSERT(n_tokens == target_logits->ne[1]);
-    GGML_ASSERT(n_batch  == target_logits->ne[2]);
-    GGML_ASSERT(n_vocab  == target_probs->ne[0]);
-    GGML_ASSERT(n_tokens == target_probs->ne[1]);
-    GGML_ASSERT(n_batch  == target_probs->ne[2]);
-
-    ggml_set_f32(target_logits, -1.0f/n_vocab);
-    ggml_set_f32(target_probs, 0.0f);
-    for (int k=0; k<n_batch; ++k) {
-        // printf("%s: batch %d\n", __func__, k);
-        size_t sample = train_samples[(example_id*n_batch + k) % n_train_samples];
-        GGML_ASSERT(sample+n_tokens-1 < n_train_data);
-
-        set_i32_2d(tokens_input, 0, k, llama_token_bos(lctx));
-        for (int i=1; i<n_tokens+1; ++i) {
-            int token = clamp(train_data[sample+i-1], 0, n_vocab-1);
-            // print_token(lctx, token);
-            set_f32_3d(target_logits, token, i-1, k, +1.0f);
-            set_f32_3d(target_probs,  token, i-1, k, +1.0f);
-            if (i<n_tokens) {
-                set_i32_2d(tokens_input, i, k, token);
-            }
-        }
-        // printf("\n=\n");
-        // for (int i=0; i<n_tokens; ++i) {
-        //     int token = get_i32_2d(tokens_input, i, k);
-        //     print_token(lctx, token);
-        // }
-        // printf("\n-\n");
-    }
-}
-
-
-void lshift_examples(struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs, int n_shift) {
-    int n_tokens = tokens_input->ne[0];
-    int n_vocab = target_logits->ne[0];
-    for (int i=0; i<n_tokens-n_shift; ++i) {
-        ggml_set_i32_1d(tokens_input, i, ggml_get_i32_1d(tokens_input, i + n_shift));
-        for (int k=0; k<n_vocab; ++k) {
-            ggml_set_f32_1d(target_logits, i*n_vocab + k, ggml_get_f32_1d(target_logits, (i + n_shift)*n_vocab + k));
-            ggml_set_f32_1d(target_probs, i*n_vocab + k,  ggml_get_f32_1d(target_probs,  (i + n_shift)*n_vocab + k));
-        }
-    }
-}
-
-struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * target) {
-    return ggml_sum(ctx, ggml_sqr(ctx, ggml_sub(ctx, target, a)));
-}
-
-struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * probs) {
-    return ggml_cross_entropy_loss(ctx, a, probs);
-}
-
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
-struct llama_file {
-    // use FILE * so we don't have to re-open the file to mmap
-    FILE * fp;
-    size_t size;
-
-    llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
-        if (fp == NULL) {
-            size = 0;
-        } else {
-            seek(0, SEEK_END);
-            size = tell();
-            seek(0, SEEK_SET);
-        }
+    float rope_freq_scale = 1.0f;
+    GGUF_GET_KEY(fctx, model->hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+    GGUF_GET_KEY(fctx, model->hparams.rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
+    GGUF_GET_KEY(fctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    if (rope_freq_scale != 1.0f) {
+        model->hparams.rope_freq_scale = 1.0f / rope_freq_scale;
     }
 
-    size_t tell() const {
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        GGML_ASSERT(ret != -1); // this really shouldn't fail
-        return (size_t) ret;
-    }
+    init_model(model);
 
-    void seek(size_t offset, int whence) {
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        GGML_ASSERT(ret == 0); // same
-    }
-
-    void read_raw(void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        std::size_t ret = std::fread(ptr, size, 1, fp);
-        if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
-        }
-        if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
-        }
-    }
-
-    std::uint32_t read_u32() {
-        std::uint32_t ret;
-        read_raw(&ret, sizeof(ret));
-        return ret;
-    }
-
-    std::string read_string(std::uint32_t len) {
-        std::vector<char> chars(len);
-        read_raw(chars.data(), len);
-        return std::string(chars.data(), len);
-    }
-
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
-        if (ret != 1) {
-            throw std::runtime_error(format("write error: %s", strerror(errno)));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
-    ~llama_file() {
-        if (fp) {
-            std::fclose(fp);
-        }
-    }
-};
-
-int tokenize_file(struct llama_context * lctx, const char * filename, std::vector<llama_token>& out) {
-    struct llama_file f(filename, "rb");
-
-    std::vector<char> buf;
-    buf.resize(f.size+1);
-
-    f.read_raw(buf.data(), f.size);
-    buf[f.size] = '\0';
-
-    int n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false);
-    if (n_tokens < 0) {
-        out.resize(-n_tokens);
-        llama_tokenize(lctx, buf.data(), out.data(), out.size(), false);
-    }
-
-    bool verify = false;
-    if (verify) {
-        const char * in  = buf.data();
-        const char * end = buf.data() + buf.size();
-        for (int i = 0; i < (int) out.size(); ++i) {
-            std::string s = llama_token_to_str(lctx, out[i]);
-            int len = s.length();
-            if (in >= end) {
-                printf("%s: unexpected end of original text.\n", __func__);
-                break;
-            }
-            const bool matches = (strncmp(in, s.c_str(), len) == 0);
-            if (matches) {
-                in += len;
-            } else {
-                printf("%s: mismatch: expected '%s', but got '%s'\n", __func__, std::string(in, len).c_str(), s.c_str());
-            }
-        }
-    }
-
-    return n_tokens;
-}
-
-void shuffle_ints(int * begin, int * end) {
-    if (end <= begin) return;
-    int max=begin[0];
-    for (int i=1; i<end-begin; ++i) {
-        if (begin[i] > max) {
-            max = begin[i];
-        }
-    }
-    std::vector<float> vals;
-    vals.resize(max+1);
-    for (int i=0; i<max+1; ++i) {
-       vals[i] = frand();
-    }
-    std::sort(begin, end, [&vals](int a, int b){
-       return vals.at(a) < vals.at(b);
-    });
-}
-
-struct my_llama_sampler_params {
-    float temp            = 0.0f;  // <= 0.0 disabled
-    int   top_k           = 20;    // <= 0 to use vocab size
-    float top_p           = 0.95f; // 1.0 = disabled
-    float tfs_z           = 1.00f; // 1.0 = disabled
-    float typical_p       = 1.00f; // 1.0 = disabled
-    int   repeat_last_n   = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float repeat_penalty  = 1.0f;  // 1.0 = disabled
-    float alpha_presence  = 0.0f;  // 0.0 = disabled
-    float alpha_frequency = 0.0f;  // 0.0 = disabled
-    int   mirostat        = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float mirostat_tau    = 5.00f; // target entropy
-    float mirostat_eta    = 0.10f; // learning rate
-    bool  penalize_nl     = true;  // consider newlines as a repeatable token
-};
-
-struct my_llama_sampler {
-    struct llama_context * ctx = NULL;
-    my_llama_sampler_params params;
-
-    int n_vocab = 0;
-    int n_ctx = 0;
-
-    float mirostat_mu;
-
-    std::vector<llama_token_data> candidates;
-    llama_token_data_array candidates_p;
-
-};
-
-void init_sampler(struct my_llama_sampler * sampler, struct llama_context * ctx) {
-    sampler->ctx = ctx;
-    sampler->n_vocab = llama_n_vocab(sampler->ctx);
-    sampler->n_ctx   = llama_n_ctx(sampler->ctx);
-    sampler->mirostat_mu = 2.0f * sampler->params.mirostat_tau;
-}
-
-llama_token sample(struct my_llama_sampler * sampler, float * logits, const llama_token * last_tokens, int n_last_tokens) {
-    GGML_ASSERT(sampler->ctx != NULL);
-
-    struct llama_context * ctx = sampler->ctx;
-
-    sampler->candidates.resize(sampler->n_vocab);
-    for (llama_token token_id = 0; token_id < sampler->n_vocab; ++token_id) {
-        sampler->candidates[token_id].id = token_id;
-        sampler->candidates[token_id].logit = logits[token_id];
-        sampler->candidates[token_id].p = 0.0;
-    }
-
-    llama_token_data_array * candidates_p = & sampler->candidates_p;
-
-    candidates_p->data = sampler->candidates.data();
-    candidates_p->size = sampler->candidates.size();
-    candidates_p->sorted = false;
-
-    const auto params = sampler->params;
-
-    // Apply penalties
-    const float nl_logit = logits[llama_token_nl(ctx)];
-
-    const int n_last = std::min(std::min(n_last_tokens, params.repeat_last_n), sampler->n_ctx);
-
-    llama_sample_repetition_penalty(
-        ctx,
-        candidates_p,
-        last_tokens + n_last_tokens - n_last,
-        n_last,
-        params.repeat_penalty);
-    llama_sample_frequency_and_presence_penalties(
-        ctx,
-        candidates_p,
-        last_tokens + n_last_tokens - n_last,
-        n_last,
-        params.alpha_frequency,
-        params.alpha_presence);
-
-    if (!params.penalize_nl) {
-        logits[llama_token_nl(ctx)] = nl_logit;
-    }
-
-    llama_token token = 0;
-    if (params.temp <= 0) {
-        // Greedy sampling
-        token = llama_sample_token_greedy(ctx, candidates_p);
-    } else {
-        if (params.mirostat == 1) {
-            int mirostat_m = 100;
-            llama_sample_temperature(ctx, candidates_p, params.temp);
-            token = llama_sample_token_mirostat(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, mirostat_m, &sampler->mirostat_mu);
-        } else if (params.mirostat == 2) {
-            llama_sample_temperature(ctx, candidates_p, params.temp);
-            token = llama_sample_token_mirostat_v2(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, &sampler->mirostat_mu);
-        } else {
-            // Temperature sampling
-            llama_sample_top_k        (ctx, candidates_p, params.top_k, 1);
-            llama_sample_tail_free    (ctx, candidates_p, params.tfs_z, 1);
-            llama_sample_typical      (ctx, candidates_p, params.typical_p, 1);
-
-            llama_sample_top_p        (ctx, candidates_p, params.top_p, 1);
-            llama_sample_temperature  (ctx, candidates_p, params.temp);
-            token = llama_sample_token(ctx, candidates_p);
-        }
-    }
-    return token;
-}
-
-void set_logits_masked(struct ggml_tensor * logits, std::vector<bool>& mask, float value) {
-    GGML_ASSERT(logits->ne[0] == (int64_t) mask.size());
-    for (int i2 = 0; i2 < logits->ne[2]; ++i2) {
-        for (int i1 = 0; i1 < logits->ne[1]; ++i1) {
-            for (int i0 = 0; i0 < logits->ne[0]; ++i0) {
-                if (!mask[i0]) continue;
-                float * ptr = (float *) ((char *) logits->data + i2*logits->nb[2] + i1*logits->nb[1] + i0*logits->nb[0]);
-                *ptr = value;
-            }
-        }
-    }
-}
-
-void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
-    if (tensor == NULL) {
-        file->write_u32(0);
-        file->write_u32(0);
-        file->write_u32(GGML_TYPE_F32);
-        file->seek((0-file->tell()) & 31, SEEK_CUR);
-        return;
-    }
-    const char * name = ggml_get_name(tensor);
-    uint32_t name_len = strlen(name);
-    uint32_t nd = tensor->n_dims;
-    uint32_t ne[4] = { (uint32_t)tensor->ne[0],
-                       (uint32_t)tensor->ne[1],
-                       (uint32_t)tensor->ne[2],
-                       (uint32_t)tensor->ne[3] };
-    file->write_u32(nd);
-    file->write_u32(name_len);
-    file->write_u32(tensor->type);
-    file->write_raw(ne, sizeof(ne[0]) * nd);
-    file->write_raw(name, name_len);
-    file->seek((0-file->tell()) & 31, SEEK_CUR);
-    file->write_raw(tensor->data, ggml_nbytes(tensor));
-}
-
-void read_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
-    int32_t nd = file->read_u32();
-    GGML_ASSERT(nd == tensor->n_dims);
-
-    uint32_t name_len       = file->read_u32();
-    enum     ggml_type type = (enum ggml_type) file->read_u32();
-    GGML_ASSERT(type == tensor->type);
-
-    uint32_t ne[4];
-    file->read_raw(ne, sizeof(ne[0]) * nd);
-    for (int i=0; i<nd; ++i) {
-        GGML_ASSERT(ne[i] == tensor->ne[i]);
-    }
-
-    std::string name = file->read_string(name_len);
-    GGML_ASSERT(strncmp(ggml_get_name(tensor), name.c_str(), sizeof(tensor->name)-1) == 0);
-
-    file->seek((0-file->tell()) & 31, SEEK_CUR);
-    file->read_raw(tensor->data, ggml_nbytes(tensor));
-}
-
-void write_opt_context(struct llama_file * file, struct ggml_opt_context * opt) {
-    const uint32_t version = 0;
-    GGML_ASSERT(opt->nx   >= 0);
-    GGML_ASSERT(opt->iter >= 0);
-    file->write_u32(version);
-    file->write_raw(&opt->params, sizeof(opt->params));
-    file->write_raw(&opt->nx,     sizeof(opt->nx));
-    file->write_raw(&opt->iter,   sizeof(opt->iter));
-    file->write_u32((uint32_t)  opt->just_initialized);
-    switch (opt->params.type) {
-        case GGML_OPT_ADAM:
-            {
-                GGML_ASSERT(opt->adam.x != NULL);
-                write_tensor(file, opt->adam.x);
-                write_tensor(file, opt->adam.g1);
-                write_tensor(file, opt->adam.g2);
-                write_tensor(file, opt->adam.m);
-                write_tensor(file, opt->adam.v);
-                write_tensor(file, opt->adam.mh);
-                write_tensor(file, opt->adam.vh);
-                write_tensor(file, opt->adam.pf);
-                file->write_raw(&opt->adam.fx_best,          sizeof(opt->adam.fx_best));
-                file->write_raw(&opt->adam.fx_prev,          sizeof(opt->adam.fx_prev));
-                file->write_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement));
-            } break;
-        case GGML_OPT_LBFGS:
-            {
-                GGML_ASSERT(opt->adam.x != NULL);
-                write_tensor(file, opt->lbfgs.x);
-                write_tensor(file, opt->lbfgs.xp);
-                write_tensor(file, opt->lbfgs.g);
-                write_tensor(file, opt->lbfgs.gp);
-                write_tensor(file, opt->lbfgs.d);
-                write_tensor(file, opt->lbfgs.pf);
-                write_tensor(file, opt->lbfgs.lmal);
-                write_tensor(file, opt->lbfgs.lmys);
-                write_tensor(file, opt->lbfgs.lms);
-                write_tensor(file, opt->lbfgs.lmy);
-                file->write_raw(&opt->lbfgs.fx_best,          sizeof(opt->lbfgs.fx_best));
-                file->write_raw(&opt->lbfgs.step,             sizeof(opt->lbfgs.step));
-                file->write_raw(&opt->lbfgs.j,                sizeof(opt->lbfgs.j));
-                file->write_raw(&opt->lbfgs.k,                sizeof(opt->lbfgs.k));
-                file->write_raw(&opt->lbfgs.end,              sizeof(opt->lbfgs.end));
-                file->write_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement));
-            } break;
-    }
-}
-
-void read_opt_context(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) {
-    uint32_t version = file->read_u32();
-    GGML_ASSERT(version == 0);
-
-    file->read_raw(&opt->params, sizeof(opt->params));
-    file->read_raw(&opt->nx,     sizeof(opt->nx));
-    ggml_opt_init(ctx, opt, opt->params, opt->nx);
-
-    file->read_raw(&opt->iter,   sizeof(opt->iter));
-    opt->just_initialized = (bool) file->read_u32();
-
-    switch (opt->params.type) {
-        case GGML_OPT_ADAM:
-            {
-                read_tensor(file, opt->adam.x);
-                read_tensor(file, opt->adam.g1);
-                read_tensor(file, opt->adam.g2);
-                read_tensor(file, opt->adam.m);
-                read_tensor(file, opt->adam.v);
-                read_tensor(file, opt->adam.mh);
-                read_tensor(file, opt->adam.vh);
-                if (opt->adam.pf) { read_tensor(file, opt->adam.pf); }
-                file->read_raw(&opt->adam.fx_best,          sizeof(opt->adam.fx_best));
-                file->read_raw(&opt->adam.fx_prev,          sizeof(opt->adam.fx_prev));
-                file->read_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement));
-            } break;
-        case GGML_OPT_LBFGS:
-            {
-                GGML_ASSERT(opt->adam.x != NULL);
-                read_tensor(file, opt->lbfgs.x);
-                read_tensor(file, opt->lbfgs.xp);
-                read_tensor(file, opt->lbfgs.g);
-                read_tensor(file, opt->lbfgs.gp);
-                read_tensor(file, opt->lbfgs.d);
-                if (opt->lbfgs.pf) { read_tensor(file, opt->lbfgs.pf); }
-                read_tensor(file, opt->lbfgs.lmal);
-                read_tensor(file, opt->lbfgs.lmys);
-                read_tensor(file, opt->lbfgs.lms);
-                read_tensor(file, opt->lbfgs.lmy);
-                file->read_raw(&opt->lbfgs.fx_best,          sizeof(opt->lbfgs.fx_best));
-                file->read_raw(&opt->lbfgs.step,             sizeof(opt->lbfgs.step));
-                file->read_raw(&opt->lbfgs.j,                sizeof(opt->lbfgs.j));
-                file->read_raw(&opt->lbfgs.k,                sizeof(opt->lbfgs.k));
-                file->read_raw(&opt->lbfgs.end,              sizeof(opt->lbfgs.end));
-                file->read_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement));
-            } break;
-    }
-}
-
-void save_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename) {
-    struct llama_file file(filename, "wb");
-    if (file.fp == NULL) {
-        return;
-    }
-
-    const uint32_t magic   = 'ggcp';
-    const uint32_t version = 0;
-
-    file.write_u32(magic);
-    file.write_u32(version);
-    file.write_u32(model->train_its);
-    file.write_u32(model->train_samples);
-    file.write_u32(model->train_tokens);
-    file.write_u32(model->hparams.n_vocab);
-    file.write_u32(model->hparams.n_embd);
-    file.write_u32(model->hparams.n_mult);
-    file.write_u32(model->hparams.n_head);
-    file.write_u32(model->hparams.n_layer);
-    file.write_u32(model->hparams.n_rot);
-
-    write_tensor(&file, model->tok_embeddings);
-    write_tensor(&file, model->norm);
-    write_tensor(&file, model->output);
+    copy_tensor_by_name(model->tok_embeddings, f_ggml_ctx, tn(LLM_TENSOR_TOKEN_EMBD));
+    copy_tensor_by_name(model->norm,           f_ggml_ctx, tn(LLM_TENSOR_OUTPUT_NORM));
+    copy_tensor_by_name(model->output,         f_ggml_ctx, tn(LLM_TENSOR_OUTPUT));
 
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
         auto & layer = model->layers[i];
 
-        write_tensor(&file, layer.attention_norm);
-        write_tensor(&file, layer.wq);
-        write_tensor(&file, layer.wk);
-        write_tensor(&file, layer.wv);
-        write_tensor(&file, layer.wo);
-        write_tensor(&file, layer.ffn_norm);
-        write_tensor(&file, layer.w1);
-        write_tensor(&file, layer.w2);
-        write_tensor(&file, layer.w3);
+        copy_tensor_by_name(layer.attention_norm, f_ggml_ctx, tni(LLM_TENSOR_ATTN_NORM, i));
+        copy_tensor_by_name(layer.wq,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_Q, i));
+        copy_tensor_by_name(layer.wk,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_K, i));
+        copy_tensor_by_name(layer.wv,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
+        copy_tensor_by_name(layer.wo,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
+        copy_tensor_by_name(layer.ffn_norm,       f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
+        copy_tensor_by_name(layer.w1,             f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
+        copy_tensor_by_name(layer.w2,             f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
+        copy_tensor_by_name(layer.w3,             f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
     }
-
-    write_opt_context(&file, opt);
 }
 
-bool load_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename, bool init) {
-    struct llama_file file(filename, "rb");
+static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model) {
+    const char * arch = "llama";
+    enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
 
-    uint32_t magic;
-    uint32_t version;
+    std::vector<char> keybuf;
+    keybuf.resize(512);
+    auto kv = [arch, &keybuf](const char * key) -> const char * {
+        snprintf(keybuf.data(), keybuf.size(), key, arch);
+        return keybuf.data();
+    };
 
-    uint32_t train_its = 0;
-    uint32_t train_samples = 0;
-    uint32_t train_tokens = 0;
+    // set arch
+    gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
+    gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
 
-    if (file.fp) {
-        printf("%s: Loading model from '%s'.\n", __func__, filename);
-        magic                  = file.read_u32();
-        GGML_ASSERT(magic     == 'ggcp');
-        version                = file.read_u32();
-        GGML_ASSERT(version   == 0);
-        train_its              = file.read_u32();
-        train_samples          = file.read_u32();
-        train_tokens           = file.read_u32();
-        model->hparams.n_vocab = file.read_u32();
-        model->hparams.n_embd  = file.read_u32();
-        model->hparams.n_mult  = file.read_u32();
-        model->hparams.n_head  = file.read_u32();
-        model->hparams.n_layer = file.read_u32();
-        model->hparams.n_rot   = file.read_u32();
-        print_params(&model->hparams);
-    }
+    // set hparams
+    gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH),              model->hparams.n_ctx                  );
+    gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH),            model->hparams.n_embd                 );
+    gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH),         model->hparams.n_ff                   );
+    gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT),        model->hparams.n_head                 );
+    gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT),                 model->hparams.n_layer                );
+    gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT),        model->hparams.n_rot                  );
 
-    if (init) {
-        init_model(model);
-    }
+    gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps         );
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE),              model->hparams.rope_freq_base         ); // TODO load in llama.cpp
+    gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR),           1.0f / model->hparams.rope_freq_scale );
 
-    if (file.fp) {
-        model->train_its = train_its;
-        model->train_samples = train_samples;
-        model->train_tokens = train_tokens;
-    }
+    // set vocab by copying from vocab_model gguf file
+    {
+        struct gguf_init_params params = {
+            /*.no_alloc = */ false,
+            /*.ctx      = */ NULL,
+        };
+        struct gguf_context * vctx = gguf_init_from_file(fn_vocab_model, params);
 
-    printf("%s: Training iterations: %u.\n", __func__, model->train_its);
-    printf("%s: Training samples:    %u.\n", __func__, model->train_samples);
-    printf("%s: Training tokens:     %u.\n", __func__, model->train_tokens);
+        const int token_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_LIST));
+        if (token_idx == -1) {
+            die("cannot find tokenizer vocab in model file");
+        }
+        const uint32_t n_vocab = gguf_get_arr_n(vctx, token_idx);
 
-    if (file.fp) {
-        read_tensor(&file, model->tok_embeddings);
-        read_tensor(&file, model->norm);
-        read_tensor(&file, model->output);
-
-        for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
-            auto & layer = model->layers[i];
-
-            read_tensor(&file, layer.attention_norm);
-            read_tensor(&file, layer.wq);
-            read_tensor(&file, layer.wk);
-            read_tensor(&file, layer.wv);
-            read_tensor(&file, layer.wo);
-            read_tensor(&file, layer.ffn_norm);
-            read_tensor(&file, layer.w1);
-            read_tensor(&file, layer.w2);
-            read_tensor(&file, layer.w3);
+        const int score_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_SCORES));
+        if (score_idx == -1) {
+            die("cannot find tokenizer scores in model file");
         }
 
-        read_opt_context(&file, model->ctx, opt);
+        const float * scores = (const float * ) gguf_get_arr_data(vctx, score_idx);
+
+        const int toktype_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE));
+        if (toktype_idx == -1) {
+            die("cannot find token type list in GGUF file");
+        }
+
+        const int * toktypes = (const int * ) gguf_get_arr_data(vctx, toktype_idx);
+
+        std::string tokenizer_name;
+        GGUF_GET_KEY(vctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL));
+
+        gguf_set_val_str(fctx, kv(LLM_KV_TOKENIZER_MODEL), tokenizer_name.c_str());
+        gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_SCORES), GGUF_TYPE_FLOAT32, scores, n_vocab);
+        gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE), GGUF_TYPE_INT32, toktypes, n_vocab);
+
+        int32_t special_bos_id = 1;
+        int32_t special_eos_id = 2;
+        int32_t special_unk_id = 0;
+        int32_t special_sep_id = -1;
+        int32_t special_pad_id = -1;
+        if (tokenizer_name == "llama") {
+            // default special tokens
+            special_bos_id = 1;
+            special_eos_id = 2;
+            special_unk_id = 0;
+            special_sep_id = -1;
+            special_pad_id = -1;
+        } else if (tokenizer_name == "gpt2") {
+            // read and copy bpe merges
+            const int merges_keyidx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_MERGES));
+            if (merges_keyidx == -1) {
+                die("cannot find tokenizer merges in model file");
+            }
+
+            const int n_merges = gguf_get_arr_n(vctx, merges_keyidx);
+
+            std::vector<const char*> merges;
+            merges.resize(n_merges);
+            for (int i = 0; i < n_merges; i++) {
+                merges[i] = gguf_get_arr_str(vctx, merges_keyidx, i);
+            }
+            gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_MERGES), merges.data(), n_merges);
+
+            // default special tokens
+            special_bos_id = 11;
+            special_eos_id = 11;
+            special_unk_id = -1;
+            special_sep_id = -1;
+            special_pad_id = -1;
+        } else {
+            fprintf(stderr, "%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str());
+            fprintf(stderr, "%s: using default tokenizer: 'llama'", __func__);
+        }
+
+        std::vector<const char*> tokens;
+        tokens.resize(n_vocab);
+        for (uint32_t i = 0; i < n_vocab; i++) {
+            tokens[i] = gguf_get_arr_str(vctx, token_idx, i);
+        }
+        gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_LIST), tokens.data(), n_vocab);
+
+        GGUF_GET_KEY(vctx, special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
+        GGUF_GET_KEY(vctx, special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
+        GGUF_GET_KEY(vctx, special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
+        GGUF_GET_KEY(vctx, special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
+        GGUF_GET_KEY(vctx, special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
+
+        gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_BOS_ID), special_bos_id);
+        gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_EOS_ID), special_eos_id);
+        gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_UNK_ID), special_unk_id);
+        gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_SEP_ID), special_sep_id);
+        gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_PAD_ID), special_pad_id);
+
+        gguf_free(vctx);
     }
 
-    return (file.fp != NULL);
+    // add tensors
+    gguf_add_tensor(fctx, model->tok_embeddings);
+    gguf_add_tensor(fctx, model->norm);
+    gguf_add_tensor(fctx, model->output);
+    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
+        auto & layer = model->layers[i];
+
+
+        gguf_add_tensor(fctx, layer.attention_norm);
+        gguf_add_tensor(fctx, layer.wq);
+        gguf_add_tensor(fctx, layer.wk);
+        gguf_add_tensor(fctx, layer.wv);
+        gguf_add_tensor(fctx, layer.wo);
+        gguf_add_tensor(fctx, layer.ffn_norm);
+        gguf_add_tensor(fctx, layer.w1);
+        gguf_add_tensor(fctx, layer.w2);
+        gguf_add_tensor(fctx, layer.w3);
+    }
 }
 
-void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, const char * filename) {
-    struct llama_file file(filename, "wb");
-    if (file.fp == NULL) {
-        return;
-    }
+static void save_llama_model_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct gguf_context * fctx = gguf_init_empty();
 
-#pragma message("TODO: implement file saving using gguf")
-    (void) vocab;
-    (void) model;
-//    // write_magic
-//    file.write_u32(LLAMA_FILE_MAGIC);   // magic
-//    file.write_u32(LLAMA_FILE_VERSION); // version
-//    // write_hparams
-//    file.write_u32(model->hparams.n_vocab);
-//    file.write_u32(model->hparams.n_embd);
-//    file.write_u32(model->hparams.n_mult);
-//    file.write_u32(model->hparams.n_head);
-//    file.write_u32(model->hparams.n_layer);
-//    file.write_u32(model->hparams.n_rot);
-//    file.write_u32(LLAMA_FTYPE_ALL_F32);
-//    // write_vocab
-//    uint32_t n_vocab = model->hparams.n_vocab;
-//    for (uint32_t i = 0; i < n_vocab; i++) {
-//        const auto & token_data = vocab->id_to_token.at(i);
-//        file.write_u32((uint32_t) token_data.tok.size());
-//        file.write_raw(token_data.tok.data(), token_data.tok.size());
-//        file.write_raw(&token_data.score, sizeof(token_data.score));
-//    }
-//    // write tensors
-//    write_tensor(&file, model->tok_embeddings);
-//    write_tensor(&file, model->norm);
-//    write_tensor(&file, model->output);
-//    for (uint32_t i = 0; i < model->hparams.n_layer; ++i) {
-//        auto & layer = model->layers[i];
-//
-//        write_tensor(&file, layer.attention_norm);
-//        write_tensor(&file, layer.wq);
-//        write_tensor(&file, layer.wk);
-//        write_tensor(&file, layer.wv);
-//        write_tensor(&file, layer.wo);
-//        write_tensor(&file, layer.ffn_norm);
-//        write_tensor(&file, layer.w1);
-//        write_tensor(&file, layer.w2);
-//        write_tensor(&file, layer.w3);
-//    }
+    save_llama_model_gguf(fctx, fn_vocab_model, model);
+
+    // write file
+    const bool only_meta = false;
+    gguf_write_to_file(fctx, filename, only_meta);
+    gguf_free(fctx);
 }
 
-float cosine_decay(const int decay_steps, const float alpha, int step) {
-    if (step > decay_steps) {
-        step = decay_steps;
+static void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct train_state * train) {
+    load_llama_model_gguf(fctx, f_ggml_ctx, model);
+    if (load_train_state_gguf(fctx, f_ggml_ctx, train)) {
+        std::string train_type = LLM_KV_TRAINING_TYPE_TRAIN_MODEL;
+        GGUF_GET_KEY(fctx, train_type, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE);
+        GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_TRAIN_MODEL);
+    } else {
+        printf("%s: loaded llama model as checkpoint\n", __func__);
     }
-    const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps));
-    const float decay = (1 - alpha)*cosine_decay + alpha;
-    return decay;
 }
 
-float cosine_decay_restart(int decay_steps, const float alpha, int step, float restart_step_mult) {
-    while (step > decay_steps) {
-        step -= decay_steps;
-        decay_steps = (int) restart_step_mult * decay_steps;
+static void save_checkpoint_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model, struct train_state * train) {
+    gguf_set_val_str(fctx, LLM_KV_TRAINING_TYPE, LLM_KV_TRAINING_TYPE_TRAIN_MODEL);
+    save_llama_model_gguf(fctx, fn_vocab_model, model);
+    save_train_state_gguf(fctx, train);
+}
+
+static bool load_checkpoint_file(const char * filename, struct my_llama_model * model, struct train_state * train) {
+    struct ggml_context * f_ggml_ctx;
+    struct gguf_init_params params;
+    params.no_alloc = false;
+    params.ctx = &f_ggml_ctx;
+    struct gguf_context * fctx = gguf_init_from_file(filename, params);
+    if (fctx == NULL) {
+        return false;
     }
-    return cosine_decay(decay_steps, alpha, step);
+
+    load_checkpoint_gguf(fctx, f_ggml_ctx, model, train);
+
+    return true;
+}
+
+static void save_checkpoint_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model, struct train_state * train) {
+    printf("%s: saving to %s\n", __func__, filename);
+    struct gguf_context * fctx = gguf_init_empty();
+
+    save_checkpoint_gguf(fctx, fn_vocab_model, model, train);
+
+    // write file
+    const bool only_meta = false;
+    gguf_write_to_file(fctx, filename, only_meta);
+    gguf_free(fctx);
 }
 
 struct train_params {
+    struct train_params_common common;
+
     const char * fn_vocab_model;
-    const char * fn_train_data;
-    const char * fn_checkpoint_in;
-    const char * fn_checkpoint_out;
     const char * fn_model_out;
 
-    uint32_t seed;
+    bool only_write_model;
 
     int n_ctx;
     int n_embd;
-    int n_mult;
     int n_head;
     int n_layer;
-    int n_rotmax;
+    int n_ff;
 
-    int n_threads;
-    int n_batch;
-    int n_examples;
-    int n_predict;
-
-    int print_info_interval;
-    int print_details_interval;
-
-    bool samples_start_after_nl;
-    bool use_adam;
-    bool use_flash;
-    bool use_scratch;
-
-    // only adam
-    int   warmup;
-    int   cos_decay_steps;
-    float cos_decay_restart;
-    float cos_decay_alpha;
-
-    int   lbfgs_n_iter;
-    int   adam_n_iter;
-    float adam_alpha;
-    float adam_decay;
-
-    int mem_model_gb;
-    int mem_compute_gb;
-    int mem_compute0_gb;
-    int mem_compute1_gb;
+    float f_norm_rms_eps;
+    float rope_freq_base;
+    float rope_freq_scale;
 };
 
-struct train_params get_default_train_params() {
+static struct train_params get_default_train_params() {
     struct train_params params;
+    params.common = get_default_train_params_common();
     params.fn_vocab_model    = "ggml-vic7b-uncensored-q4_0.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
     params.fn_model_out      = "ggml-checkpoint-f32.bin";
 
-    params.seed       =   -1;
+    params.only_write_model = false;
 
     params.n_ctx      =  128;
     params.n_embd     =  256;
-    params.n_mult     =  256;
     params.n_head     =    8;
     params.n_layer    =   16;
-    params.n_rotmax   =   64;
+    params.n_ff       =  768;
 
-    params.n_threads  =    6;
-    params.n_batch    =    8;
-    params.n_examples =    8;
-    params.n_predict  = 1024;
-
-    params.print_info_interval    = 1;
-    params.print_details_interval = 2;
-
-    params.samples_start_after_nl = false;
-    params.use_adam               = true;
-    params.use_flash              = true;
-    params.use_scratch            = true;
-
-    // only adam
-    params.warmup            =  100;
-    params.cos_decay_steps   = 1000;
-    params.cos_decay_restart = 1.1f;
-    params.cos_decay_alpha   = 0.0f;
-
-    params.lbfgs_n_iter      = 16;
-    params.adam_n_iter       = 16;
-    params.adam_alpha        = 1e-3f;
-    params.adam_decay        = 1e-3f;
-
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
-    params.mem_compute0_gb = 8;
-    params.mem_compute1_gb = 2;
+    params.f_norm_rms_eps  = 1e-5f;
+    params.rope_freq_base  = 10000.0f;
+    params.rope_freq_scale = 1.0f;
 
     return params;
 }
 
-void train_print_usage(int /*argc*/, char ** argv, const struct train_params * params) {
+static void train_print_usage(int argc, char ** argv, const struct train_params * params) {
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                 show this help message and exit\n");
+
     fprintf(stderr, "  --vocab-model FNAME        model path from which to load vocab (default '%s')\n", params->fn_vocab_model);
-    fprintf(stderr, "  --train-data FNAME         path from which to load training data (default '%s')\n", params->fn_train_data);
-    fprintf(stderr, "  --checkpoint-in FNAME      path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
-    fprintf(stderr, "  --checkpoint-out FNAME     path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
     fprintf(stderr, "  --model-out FNAME          path to save ggml model (default '%s')\n", params->fn_model_out);
-    fprintf(stderr, "  -s SEED, --seed SEED       RNG seed (default: -1, use random seed for -1)\n");
-    fprintf(stderr, "  -c N, --ctx N              Context size used during training (default %d)\n", params->n_ctx);
+    fprintf(stderr, "  --only-write-model         only save llama model, don't do any training. use this if you only want to convert a checkpoint to a model.\n");
     fprintf(stderr, "  --embd N                   Embedding size used for new models (default %d)\n", params->n_embd);
-    fprintf(stderr, "  --mult N                   Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult);
+    fprintf(stderr, "  --ff N                     Feedforward size used for new models. (default %d)\n", params->n_ff);
     fprintf(stderr, "  --head N                   Number of heads for new models (default %d)\n", params->n_head);
     fprintf(stderr, "  --layer N                  Number of layers for new models (default %d)\n", params->n_layer);
-    fprintf(stderr, "  --rotmax N                 Maximal number Rope dimensions for new models (default %d)\n", params->n_rotmax);
-    fprintf(stderr, "  -t N, --threads N          Number of threads (default %d)\n", params->n_threads);
-    fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
-    fprintf(stderr, "  -n N, --examples N         Number of examples to train (default %d)\n", params->n_examples);
-    fprintf(stderr, "  --predict N                Number of tokens to generate after training (default %d)\n", params->n_predict);
-    fprintf(stderr, "  --print-info-interval N    Print infos during training each N examples (default %d)\n", params->print_info_interval);
-    fprintf(stderr, "  --print-details-interval N Print details during training each N examples (default %d)\n", params->print_details_interval);
-    fprintf(stderr, "  --samples-after-nl         Training samples start after newlines. (default %s)\n", params->samples_start_after_nl ? "on" : "off");
-    fprintf(stderr, "  --use-lbfgs                Use LBFGS optimizer instead of default Adam\n");
-    fprintf(stderr, "  --use-adam                 Use Adam optimizer (default)\n");
-    fprintf(stderr, "  --no-flash                 Don't use flash attention.\n");
-    fprintf(stderr, "  --use-flash                Use flash attention (default)\n");
-    fprintf(stderr, "  --no-scratch               Don't use scratch buffers\n");
-    fprintf(stderr, "  --use-scratch              Use scratch buffers (default)\n");
-    fprintf(stderr, "  --warmup N                 Number of warmup steps (default %d)\n", params->warmup);
-    fprintf(stderr, "  --cos-decay-steps N        Number of cosine decay steps (default %d)\n", params->cos_decay_steps);
-    fprintf(stderr, "  --cos-decay-restart N      Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart);
-    fprintf(stderr, "  --cos-decay-alpha N        Cosine decay alpha (default %f)\n", params->cos_decay_alpha);
-    fprintf(stderr, "  --lbfgs-iter N             Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter);
-    fprintf(stderr, "  --adam-iter N              Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter);
-    fprintf(stderr, "  --adam-alpha N             Adam learning rate alpha (default %f)\n", params->adam_alpha);
-    fprintf(stderr, "  --adam-decay N             AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
-    fprintf(stderr, "  --mem-model N              Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb);
-    fprintf(stderr, "  --mem-compute N            Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb);
-    fprintf(stderr, "  --mem-compute0 N           Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute0_gb);
-    fprintf(stderr, "  --mem-compute1 N           Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute1_gb);
-    fprintf(stderr, "\n");
+    fprintf(stderr, "  --norm-rms-eps F           RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps);
+    fprintf(stderr, "  --rope-freq-base F         Frequency base for ROPE (default %f)\n", params->rope_freq_base);
+    fprintf(stderr, "  --rope-freq-scale F        Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
+
+    print_common_train_usage(argc, argv, &params->common);
 }
 
-bool train_params_parse(int argc, char ** argv, struct train_params * params) {
+static bool train_params_parse(int argc, char ** argv, struct train_params * params) {
     bool invalid_param = false;
     std::string arg;
     struct train_params default_params = get_default_train_params();
@@ -2824,60 +840,39 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
             std::replace(arg.begin(), arg.end(), '_', '-');
         }
 
-        if (arg == "--vocab-model") {
+        if (consume_common_train_arg(argc, argv, &i, &params->common, &invalid_param)) {
+            if (invalid_param) {
+                break;
+            } else if (params->common.print_usage) {
+                train_print_usage(argc, argv, &default_params);
+                exit(0);
+            }
+        } else if (arg == "--vocab-model") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params->fn_vocab_model = argv[i];
-        } else if (arg == "--train-data") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->fn_train_data = argv[i];
-        } else if (arg == "--checkpoint-in") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->fn_checkpoint_in = argv[i];
-        } else if (arg == "--checkpoint-out") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->fn_checkpoint_out = argv[i];
         } else if (arg == "--model-out") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params->fn_model_out = argv[i];
-        } else if (arg == "-s" || arg == "--seed") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->seed = std::stoi(argv[i]);
-        } else if (arg == "-c" || arg == "--ctx") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--only-write-model") {
+            params->only_write_model = true;
         } else if (arg == "--embd") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params->n_embd = std::stoi(argv[i]);
-        } else if (arg == "--mult") {
+        } else if (arg == "--ff") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_mult = std::stoi(argv[i]);
+            params->n_ff = std::stoi(argv[i]);
         } else if (arg == "--head") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -2890,137 +885,24 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
                 break;
             }
             params->n_layer = std::stoi(argv[i]);
-        } else if (arg == "--rotmax") {
+        } else if (arg == "--norm-rms-eps") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rotmax = std::stoi(argv[i]);
-        } else if (arg == "-t" || arg == "--threads") {
+            params->f_norm_rms_eps = std::stof(argv[i]);
+        } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_threads = std::stoi(argv[i]);
-        } else if (arg == "-b" || arg == "--batch") {
+            params->rope_freq_base = std::stof(argv[i]);
+        } else if (arg == "--rope-freq-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_batch = std::stoi(argv[i]);
-        } else if (arg == "-n" || arg == "--examples") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->n_examples = std::stoi(argv[i]);
-        } else if (arg == "--predict") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->n_predict = std::stoi(argv[i]);
-        } else if (arg == "--print-info-interval") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->print_info_interval = std::stoi(argv[i]);
-        } else if (arg == "--print-details-interval") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->print_details_interval = std::stoi(argv[i]);
-        } else if (arg == "--samples-after-nl") {
-            params->samples_start_after_nl = true;
-        } else if (arg == "--use-lbfgs") {
-            params->use_adam = false;
-        } else if (arg == "--use-adam") {
-            params->use_adam = true;
-        } else if (arg == "--no-flash") {
-            params->use_flash = false;
-        } else if (arg == "--use-flash") {
-            params->use_flash = true;
-        } else if (arg == "--no-scratch") {
-            params->use_scratch = false;
-        } else if (arg == "--use-scratch") {
-            params->use_scratch = true;
-        } else if (arg == "--warmup") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->warmup = std::stoi(argv[i]);
-        } else if (arg == "--cos-decay-steps") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->cos_decay_steps = std::stof(argv[i]);
-        } else if (arg == "--cos-decay-restart") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->cos_decay_restart = std::stof(argv[i]);
-        } else if (arg == "--cos-decay-alpha") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->cos_decay_alpha = std::stof(argv[i]);
-        } else if (arg == "--lbfgs-iter") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->lbfgs_n_iter = std::stoi(argv[i]);
-        } else if (arg == "--adam-iter") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->adam_n_iter = std::stoi(argv[i]);
-        } else if (arg == "--adam-alpha") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->adam_alpha = std::stof(argv[i]);
-        } else if (arg == "--adam-decay") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->adam_decay = std::stof(argv[i]);
-        } else if (arg == "--mem-model") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->mem_model_gb = std::stoi(argv[i]);
-        } else if (arg == "--mem-compute") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->mem_compute_gb = std::stoi(argv[i]);
-        } else if (arg == "--mem-compute0") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->mem_compute0_gb = std::stoi(argv[i]);
-        } else if (arg == "--mem-compute1") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params->mem_compute1_gb = std::stoi(argv[i]);
-        } else if (arg == "-h" || arg == "--help") {
-            train_print_usage(argc, argv, &default_params);
-            exit(0);
+            params->rope_freq_scale = std::stof(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -3032,10 +914,56 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
         train_print_usage(argc, argv, &default_params);
         exit(1);
     }
+    finish_processing_train_args(&params->common);
 
     return true;
 }
 
+struct save_train_files_data {
+    const char            * fn_checkpoint_out;
+    const char            * fn_model_out;
+    const char            * fn_vocab_model;
+    const char            * pattern_fn_it;
+    const char            * fn_latest;
+    struct my_llama_model * model;
+};
+
+static void save_train_files(void * vdata, struct train_state * train) {
+    struct save_train_files_data * data   = (struct save_train_files_data *) vdata;
+    int64_t iter = train->opt->iter;
+
+    if (strlen(data->fn_checkpoint_out) > 0) {
+        save_checkpoint_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->fn_vocab_model, data->model, train);
+        save_checkpoint_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->fn_vocab_model, data->model, train);
+
+    }
+    if (strlen(data->fn_model_out) > 0) {
+        save_llama_model_file(get_train_filename(data->fn_model_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->fn_vocab_model, data->model);
+        save_llama_model_file(get_train_filename(data->fn_model_out, data->pattern_fn_it, data->fn_latest, -1  ).c_str(), data->fn_vocab_model, data->model);
+    }
+}
+
+static int64_t get_parameter_count(struct my_llama_model* model) {
+    int64_t nx = 0;
+    nx += ggml_nelements(model->tok_embeddings);
+    nx += ggml_nelements(model->norm);
+    nx += ggml_nelements(model->output);
+
+    for (uint32_t i = 0; i < model->layers.size(); ++i) {
+        auto & layer = model->layers[i];
+        nx += ggml_nelements(layer.attention_norm);
+        nx += ggml_nelements(layer.wq);
+        nx += ggml_nelements(layer.wk);
+        nx += ggml_nelements(layer.wv);
+        nx += ggml_nelements(layer.wo);
+        nx += ggml_nelements(layer.ffn_norm);
+        nx += ggml_nelements(layer.w1);
+        nx += ggml_nelements(layer.w2);
+        nx += ggml_nelements(layer.w3);
+    }
+    return nx;
+}
+
 int main(int argc, char ** argv) {
     struct train_params params = get_default_train_params();
 
@@ -3043,353 +971,340 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
+    if (params.common.seed == LLAMA_DEFAULT_SEED) {
+        params.common.seed = time(NULL);
     }
-    printf("%s: seed: %u\n", __func__, params.seed);
-    srand(params.seed);
+    printf("%s: seed: %u\n", __func__, params.common.seed);
+    srand(params.common.seed);
 
-    struct llama_context_params llama_params = llama_context_default_params();
-    llama_params.vocab_only = true;
+    struct llama_model_params mparams = llama_model_default_params();
+    mparams.vocab_only = true;
 
-    struct llama_model * lmodel = llama_load_model_from_file(params.fn_vocab_model, llama_params);
-    struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params);
+    struct llama_context_params cparams = llama_context_default_params();
 
-    struct llama_vocab vocab;
-    {
-        const int n_vocab = llama_n_vocab(lctx);
-        vocab.id_to_token.resize(n_vocab);
-        for (int i=0; i<n_vocab; ++i) {
-            vocab.id_to_token[i].text  = llama_token_get_text(lctx, i);
-            vocab.id_to_token[i].score = llama_token_get_score(lctx, i);
-            vocab.id_to_token[i].type  = llama_token_get_type(lctx, i);
-            vocab.token_to_id.emplace(vocab.id_to_token[i].text, i);
-        }
-    }
-
-    printf("%s: tokenize training data\n", __func__);
-    std::vector<llama_token> train_tokens;
-    if (tokenize_file(lctx, params.fn_train_data, train_tokens) < 0) {
-        fprintf(stderr, "%s: failed to tokenize file '%s'\n", __func__, params.fn_train_data);
-    }
-    printf("%s: number of training tokens: %d\n", __func__, (int) train_tokens.size());
+    struct llama_model * lmodel = llama_load_model_from_file(params.fn_vocab_model, mparams);
+    struct llama_context * lctx = llama_new_context_with_model(lmodel, cparams);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
+    model.hparams.n_vocab = llama_n_vocab(lmodel);
+    model.hparams.n_ctx   = params.common.n_ctx;
     model.hparams.n_embd  = params.n_embd;
-    model.hparams.n_mult  = params.n_mult;
     model.hparams.n_head  = params.n_head;
     model.hparams.n_layer = params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_ff    = params.n_ff;
+    // llama.cpp requires n_rot to be exactly n_embd / n_head
+    model.hparams.n_rot   = model.hparams.n_embd / model.hparams.n_head;
+    model.hparams.f_norm_rms_eps  = params.f_norm_rms_eps;
+    model.hparams.rope_freq_base  = params.rope_freq_base;
+    model.hparams.rope_freq_scale = params.rope_freq_scale;
+
+    struct train_state      * train = init_train_state();
+    struct ggml_opt_context * opt   = train->opt;
+
+    // set opt params from command line
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    opt->params.print_forward_graph     = false;
+    opt->params.print_backward_graph    = false;
+    opt->params.n_threads               = params.common.n_threads;
+    opt->params.past                    = params.common.opt_past;
+    opt->params.delta                   = params.common.opt_delta;
+    opt->params.max_no_improvement      = params.common.opt_max_no_improvement;
+    opt->params.n_gradient_accumulation = params.common.n_gradient_accumulation;
+    opt->params.adam.n_iter             = params.common.adam_n_iter;
+    opt->params.adam.sched              = 1.0f;
+    opt->params.adam.alpha              = params.common.adam_alpha;
+    opt->params.adam.decay              = params.common.adam_decay;
+    opt->params.adam.decay_min_ndim     = params.common.adam_decay_min_ndim;
+    opt->params.adam.beta1              = params.common.adam_beta1;
+    opt->params.adam.beta2              = params.common.adam_beta2;
+    opt->params.adam.gclip              = params.common.adam_gclip;
+    opt->params.adam.eps_f              = params.common.adam_eps_f;
+
+    printf("%s: init model\n", __func__);
+    bool existed = load_checkpoint_file(params.common.fn_checkpoint_in, &model, train);
+    if (existed) {
+        // overwrite last n_ctx with user provided n_ctx
+        if (params.common.custom_n_ctx) {
+            model.hparams.n_ctx = params.common.n_ctx;
+        }
+
+        const bool opt_past_changed = opt->params.past != params.common.opt_past;
+
+        if (opt_past_changed) {
+            die("Optimizer parameter '--opt-past N' differs from checkpoint file. To use different value train from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting");
+            // need to discard previous optimizer past function value statistics and opt_init with new shapes
+            // TODO
+        }
+    } else {
+        init_model(&model);
+        randomize_model(&model, params.common.seed, 0.0f, 1.0f, -1.0f, +1.0f);
+        if (!params.only_write_model) {
+            ggml_opt_init(opt->ctx, opt, opt->params, get_parameter_count(&model));
+        }
+    }
+    opt->iter = train->train_its;
 
     print_params(&model.hparams);
+    printf("%s: total train_iterations %llu\n", __func__, (long long unsigned) train->train_its);
+    printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
+    printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
+    printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
+    printf("%s: model_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(model.ctx) + model.data.size()), (float) (ggml_used_mem(model.ctx) + model.data.size()) / (1024.0f*1024.0f));
 
-    std::vector<size_t> token_noccurs;
-    std::vector<bool>   token_notavail;
-    token_noccurs.resize(model.hparams.n_vocab, 0);
-    token_notavail.resize(model.hparams.n_vocab, true);
-    for (int i = 0; i < (int) train_tokens.size(); ++i) {
-        ++token_noccurs[train_tokens[i]];
-        token_notavail[train_tokens[i]] = false;
+    if (params.only_write_model) {
+        save_train_files_data save_data;
+        save_data.fn_checkpoint_out = "";
+        save_data.fn_model_out      = params.fn_model_out;
+        save_data.fn_vocab_model    = params.fn_vocab_model;
+        save_data.pattern_fn_it     = params.common.pattern_fn_it;
+        save_data.fn_latest         = params.common.fn_latest;
+        save_data.model             = &model;
+
+        save_train_files(&save_data, train);
+
+        free_train_state(train);
+        ggml_free(model.ctx);
+        llama_free(lctx);
+        llama_free_model(lmodel);
+        return 0;
     }
 
-    std::vector<float> token_freq;
-    token_freq.resize(model.hparams.n_vocab, 0);
-    int n_unique_tokens = 0;
-    for (int i = 0; i < (int) token_noccurs.size(); ++i) {
-        token_freq[i] = (float) token_noccurs[i] / (float) train_tokens.size();
-        n_unique_tokens += (token_noccurs[i] > 0) ? 1 : 0;
-    }
-    printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens);
-
-    struct my_llama_kv_cache kv_self;
-
-
-    struct ggml_init_params lcparams;
-    lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
-    lcparams.mem_buffer = NULL;
-    lcparams.no_alloc   = false;
-
-    model.ctx = ggml_init(lcparams);
-    kv_self.ctx = model.ctx;
-
-    my_llama_sampler sampler;
-
+    printf("%s: opt_size  = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f));
+    printf("%s: opt iter %d\n", __func__, opt->iter);
 
     int n_tokens = model.hparams.n_ctx;
     int n_vocab  = model.hparams.n_vocab;
-    int n_batch  = params.n_batch;
+    int n_batch  = params.common.n_batch;
 
-    struct ggml_opt_context * opt = (struct ggml_opt_context *) alloca(sizeof(struct ggml_opt_context));
-    memset(opt, 0, sizeof(struct ggml_opt_context));
+    std::vector<uint8_t> mem_input_data;
+    std::vector<uint8_t> mem_compute_data;
 
-    struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
-    struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
-    opt_params_adam.print_forward_graph = false;
-    opt_params_adam.print_backward_graph = false;
-    opt_params_adam.n_threads   = params.n_threads;
-    opt_params_adam.adam.n_iter = params.adam_n_iter;
-    opt_params_adam.adam.sched  = 1.0f;
-    opt_params_adam.adam.alpha  = params.adam_alpha;
-    opt_params_adam.adam.decay  = params.adam_decay;
+    ggml_allocr * alloc = NULL;
 
-    opt_params_lbfgs.print_forward_graph = false;
-    opt_params_lbfgs.print_backward_graph = false;
-    opt_params_lbfgs.n_threads    = params.n_threads;
-    opt_params_lbfgs.lbfgs.n_iter = params.lbfgs_n_iter;
+    // context for input tensors without their data
+    struct ggml_init_params ctx_input_params = {
+        ggml_tensor_overhead() * 2, // mem_size
+        NULL,                       // mem_buffer
+        true,                       // no_alloc
+    };
+    struct ggml_context * ctx_input = ggml_init(ctx_input_params);
 
-    opt->ctx = model.ctx;
-    opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs;
+    // the input tensors
+    struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
+    struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
-    printf("%s: init model\n", __func__);
-    bool existed = load_checkpoint(&model, opt, params.fn_checkpoint_in, true);
-    set_param_model(&model);
+    // measure required memory for input tensors
+    alloc = ggml_allocr_new_measure(tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+    ggml_allocr_free(alloc);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
-    opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs;
+    // allocate input tensors
+    mem_input_data.resize(max_input_size);
+    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    ggml_allocr_free(alloc);
 
-    opt->iter = model.train_its;
-    printf("%s: opt iter %d\n", __func__, opt->iter);
+    // context for compute tensors without their data
+    size_t estimated_compute_size_wo_data = (
+        ggml_tensor_overhead()*GGML_MAX_NODES*2
+      + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*(
+            params.common.use_checkpointing ? 3 : 2
+        )
+    );
+    struct ggml_init_params ctx_compute_params = {
+        estimated_compute_size_wo_data, // mem_size
+        NULL,                           // mem_buffer
+        true,                           // no_alloc
+    };
+    struct ggml_context * ctx_compute = NULL;
 
-    bool from_scratch = !existed;
-    if (from_scratch) {
-        randomize_model(&model, params.seed, 0.0f, 1.0f, -1.0f, +1.0f);
-    }
+    struct ggml_tensor * loss   = NULL;
+    struct ggml_tensor * logits = NULL;
 
-    init_kv_cache(&kv_self, &model, 1);
-    // init_kv_cache(&kv_self, &model, n_batch);
-    init_sampler(&sampler, lctx);
+    struct ggml_cgraph * gf     = NULL;
+    struct ggml_cgraph * gb     = NULL;
+    struct ggml_cgraph * gb_tmp = NULL;
 
-    printf("used_mem model+cache: %zu bytes\n", ggml_used_mem(model.ctx));
-    // ggml_print_tensor_objects(model.ctx);
-
-    // TODO: use std::vector<uint8_t> intead of "new"
-    size_t    compute_size = 1024ll*1024ll*1024ll*((size_t) params.mem_compute_gb);
-    uint8_t * compute_addr = new uint8_t[compute_size];
-
-    size_t size_buf_0 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute0_gb);
-    size_t size_buf_1 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute1_gb);
-    uint8_t * compute_buf_0 = new uint8_t[size_buf_0];
-    uint8_t * compute_buf_1 = new uint8_t[size_buf_1];
-
-    GGML_ASSERT(n_tokens < (int) train_tokens.size());
-    std::vector<int> train_samples;
-    train_samples.push_back(0);
-    for (int i = 1; i < (int) train_tokens.size() - n_tokens; ++i) {
-        if (!params.samples_start_after_nl || (train_tokens[i-1] == llama_token_nl(lctx))) {
-            train_samples.push_back(i);
+    // measure required memory for compute tensors
+    size_t best_compute_size = SIZE_MAX;
+    enum ggml_cgraph_eval_order best_order = GGML_CGRAPH_EVAL_ORDER_COUNT;
+    // find best evaluation order
+    for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
+        ctx_compute = ggml_init(ctx_compute_params);
+        alloc = ggml_allocr_new_measure(tensor_alignment);
+        gf = ggml_new_graph(ctx_compute);
+        gf->order = (enum ggml_cgraph_eval_order) order;
+        gb = ggml_new_graph(ctx_compute);
+        gb_tmp = params.common.use_checkpointing
+            ? ggml_new_graph(ctx_compute)
+            : NULL;
+        loss = llama_build_train_graphs(
+            &model, alloc, ctx_compute,
+            gf, gb, gb_tmp,
+            &logits, tokens_input, target_probs,
+            n_tokens, n_batch,
+            params.common.use_flash,
+            params.common.use_checkpointing
+        );
+        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        if (max_compute_size < best_compute_size) {
+            best_compute_size = max_compute_size;
+            best_order = gf->order;
         }
+        ggml_allocr_free(alloc);
+        ggml_free(ctx_compute);
     }
-    shuffle_ints(train_samples.data(), train_samples.data() + train_samples.size());
-    for (int i = 0; i < (int) train_samples.size(); ++i) {
-        GGML_ASSERT(train_samples[i]+n_tokens-1 < (int) train_tokens.size());
+    size_t max_compute_size = best_compute_size;
+    printf("%s: compute_size = %zu bytes (%.1f MB)\n", __func__, max_compute_size, (float) max_compute_size / (1024.0f*1024.0f));
+    printf("%s: evaluation order = %s\n", __func__,
+        (best_order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? "LEFT_TO_RIGHT" :
+        (best_order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? "RIGHT_TO_LEFT" :
+        "invalid");
+
+    // allocate compute tensors
+    mem_compute_data.resize(max_compute_size);
+    ctx_compute = ggml_init(ctx_compute_params);
+    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    gf = ggml_new_graph(ctx_compute);
+    gf->order = best_order;
+    gb = ggml_new_graph(ctx_compute);
+    gb_tmp = params.common.use_checkpointing
+        ? ggml_new_graph(ctx_compute)
+        : NULL;
+    loss = llama_build_train_graphs(
+        &model, alloc, ctx_compute,
+        gf, gb, gb_tmp,
+        &logits, tokens_input, target_probs,
+        n_tokens, n_batch,
+        params.common.use_flash,
+        params.common.use_checkpointing
+    );
+    ggml_allocr_free(alloc);
+
+    std::vector<llama_token> train_tokens;
+    std::vector<size_t> train_samples_begin;
+    std::vector<size_t> train_samples_size;
+    printf("%s: tokenize training data\n", __func__);
+    tokenize_file(lctx,
+            params.common.fn_train_data,
+            params.common.sample_start,
+            params.common.include_sample_start,
+            params.common.overlapping_samples,
+            n_tokens,
+            train_tokens,
+            train_samples_begin,
+            train_samples_size);
+    GGML_ASSERT(train_samples_begin.size() == train_samples_size.size());
+
+    printf("%s: number of training tokens: %zu\n", __func__, train_tokens.size());
+
+    size_t shuffle_samples_hash = compute_samples_hash(params.common.fn_train_data, train_samples_begin.data(), train_samples_size.data(), train_samples_size.size());
+    const bool changed_train_data = (shuffle_samples_hash != train->shuffle_samples_hash) || (train->shuffle_sample_count != train_samples_size.size());
+    if (changed_train_data) {
+        printf("%s: train data seems to have changed. restarting shuffled epoch.\n", __func__);
     }
-
-    std::vector<uint8_t> work_buffer;
-
+    if (params.common.force_reshuffle) {
+        printf("%s: forced reshuffling of data. restarting with newly shuffled epoch.\n", __func__);
+    }
+    if ((train->shuffle_rng_state_current == "") || changed_train_data || params.common.force_reshuffle) {
+        train->shuffle_rng_state_current = mt19937_seed_to_state(params.common.seed);
+        train->shuffle_sample_count = train_samples_size.size();
+        train->shuffle_next_sample = 0;
+        train->shuffle_samples_hash = shuffle_samples_hash;
+    }
+    std::vector<size_t> train_shuffled_samples_offs;
+    std::vector<size_t> train_shuffled_samples_begin;
+    std::vector<size_t> train_shuffled_samples_size;
+    train_shuffled_samples_offs.resize(train_samples_begin.size());
+    train_shuffled_samples_begin.resize(train_samples_begin.size());
+    train_shuffled_samples_size.resize(train_samples_size.size());
+    train->shuffle_rng_state_next = shuffle_samples(
+        train->shuffle_rng_state_current,
+        train_shuffled_samples_offs.data(),
+        train_shuffled_samples_begin.data(),
+        train_shuffled_samples_size.data(),
+        train_samples_begin.data(),
+        train_samples_size.data(),
+        train_samples_size.size());
     printf("%s: begin training\n", __func__);
 
-    for (int ex = 0; ex < params.n_examples; ++ex) {
-        if (ex*n_batch >= (int) train_samples.size()) {
-            shuffle_ints(train_samples.data(), train_samples.data() + train_samples.size());
-            for (int i = 0; i < (int) train_samples.size(); ++i) {
-                GGML_ASSERT(train_samples[i]+n_tokens-1 < (int) train_tokens.size());
-            }
-        }
+    save_train_files_data save_data;
+    save_data.fn_checkpoint_out = params.common.fn_checkpoint_out;
+    save_data.fn_model_out      = params.fn_model_out;
+    save_data.fn_vocab_model    = params.fn_vocab_model;
+    save_data.pattern_fn_it     = params.common.pattern_fn_it;
+    save_data.fn_latest         = params.common.fn_latest;
+    save_data.model             = &model;
 
-        struct ggml_init_params cparams = {
-            /*.mem_size   =*/ compute_size,
-            /*.mem_buffer =*/ compute_addr,
-            /*.no_alloc   =*/ false,
-        };
-        struct ggml_context * ctx0 = ggml_init(cparams);
+    struct train_opt_callback_data opt_cb_data;
+    opt_cb_data.params                 = &params.common;
+    opt_cb_data.train                  = train;
+    opt_cb_data.save_cb                = &save_train_files;
+    opt_cb_data.save_data              = &save_data;
+    opt_cb_data.lctx                   = lctx;
+    opt_cb_data.last_save_iter         = opt->iter;
+    opt_cb_data.tokens_data            = train_tokens.data();
+    opt_cb_data.tokens_size            = train_tokens.size();
+    opt_cb_data.samples_begin          = train_samples_begin.data();
+    opt_cb_data.samples_size           = train_samples_size.data();
+    opt_cb_data.shuffled_samples_offs  = train_shuffled_samples_offs.data();
+    opt_cb_data.shuffled_samples_begin = train_shuffled_samples_begin.data();
+    opt_cb_data.shuffled_samples_size  = train_shuffled_samples_size.data();
+    opt_cb_data.samples_count          = train_samples_size.size();
+    opt_cb_data.tokens_input           = tokens_input;
+    opt_cb_data.target_probs           = target_probs;
+    opt_cb_data.first_iter             = opt->iter;
+    opt_cb_data.first_epoch            = train->train_epochs;
+    opt_cb_data.iter_at_last_epoch     = -1;
+    opt_cb_data.last_time              = ggml_time_ms();
+    opt_cb_data.millis_per_iter        = 0.0;
 
-        struct ggml_tensor * after_opt_best_samples = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch);
-        //struct ggml_tensor * after_opt_probs        = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
-        struct ggml_tensor * tokens_input           = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch);
-        struct ggml_tensor * target_logits          = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
-        struct ggml_tensor * target_probs           = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
+    // measure required memory for work buffer
+    size_t max_work_size = ggml_graph_plan(gb, params.common.n_threads).work_size + GGML_OBJECT_SIZE;
+    printf("%s: work_size = %zu bytes (%.1f MB)\n", __func__, max_work_size, (float) max_work_size / (1024.0f*1024.0f));
 
-        int n_past = 0;
+    // context for work buffer
+    struct ggml_init_params ctx_work_params = {
+        max_work_size, // mem_size
+        NULL,          // mem_buffer
+        false,         // no_alloc
+    };
+    struct ggml_context * ctx_work = ggml_init(ctx_work_params);
 
-        struct ggml_tensor * gfbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0));
-        struct ggml_tensor * gbbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0));
+    int64_t t0 = ggml_time_ms();
 
-        memset(gfbuf->data, 0, ggml_nbytes(gfbuf));
-        memset(gbbuf->data, 0, ggml_nbytes(gbbuf));
+    ggml_opt_resume_g(ctx_work, opt, loss, gf, gb, &train_opt_callback, (void *) &opt_cb_data);
 
-        struct ggml_cgraph * gf = (struct ggml_cgraph *) gfbuf->data;
-        struct ggml_cgraph * gb = (struct ggml_cgraph *) gbbuf->data;
+    ggml_free(ctx_work);
+    ggml_free(ctx_compute);
+    ggml_free(ctx_input);
 
+    int64_t t1 = ggml_time_ms();
+    printf("%s: total training time: ", __func__);
+    print_duration((double) (t1 - t0));
+    printf("\n");
 
-        get_example_targets_batch(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), ex,  tokens_input, target_logits, target_probs);
+    int new_iters = opt->iter - opt_cb_data.last_save_iter;
+    if (new_iters > 0) {
+        train->train_its     += new_iters;
+        train->train_tokens  += new_iters * opt->params.n_gradient_accumulation * n_batch * n_tokens;
 
-        GGML_ASSERT(n_past == 0);
-
-        struct ggml_tensor * loss   = NULL;
-        struct ggml_tensor * logits = NULL;
-
-        if (params.use_scratch) {
-            loss = forward_batch_wo_cache_flash_attn_train(
-                    &model, ctx0,
-                    gf, gb,
-                    &logits, tokens_input, target_probs,
-                    compute_buf_0, compute_buf_1,
-                    size_buf_0, size_buf_1,
-                    n_tokens, n_batch);
-        } else if (params.use_flash) {
-            logits = forward_batch_wo_cache_flash_attn(&model, ctx0, gf, tokens_input, n_tokens, n_batch);
-            loss   = cross_entropy_loss(ctx0, logits, target_probs);
-            ggml_build_forward_expand(gf, loss);
-            *gb = ggml_build_backward(ctx0, gf, true);
-        } else {
-            logits = forward_batch_wo_cache(&model, ctx0, gf, tokens_input, n_tokens, n_batch);
-            loss   = cross_entropy_loss(ctx0, logits, target_probs);
-            ggml_build_forward_expand(gf, loss);
-            *gb = ggml_build_backward(ctx0, gf, true);
-        }
-
-        ggml_graph_compute_helper(work_buffer, gf, params.n_threads);
-
-        size_t used_mem_before_opt = ggml_used_mem(ctx0);
-
-        float error_before_opt = ggml_get_f32_1d(loss, 0);
-
-        opt->params.adam.sched = (opt->iter < params.warmup)
-            ? (float) opt->iter / (float) params.warmup
-            : cosine_decay_restart(
-                params.cos_decay_steps,
-                params.cos_decay_alpha,
-                opt->iter - params.warmup,
-                params.cos_decay_restart);
-
-        printf("%s: opt->params.adam.sched %.5f\n", __func__, opt->params.adam.sched);
-
-        ggml_opt_resume_g(ctx0, opt, loss, gf, gb);
-
-        size_t used_mem_after_opt = ggml_used_mem(ctx0);
-
-        model.train_its = opt->iter;
-        model.train_samples += n_batch;
-        model.train_tokens  += n_batch * n_tokens;
-
-        ggml_graph_compute_helper(work_buffer, gf, params.n_threads);
-
-        float error_after_opt = ggml_get_f32_1d(loss, 0);
-
-        if (params.print_info_interval > 0 && ex % params.print_info_interval == 0) {
-            printf("Example %d, opt iter %d\n", ex, opt->iter);
-            printf("error_before_opt: %.6f\n", error_before_opt);
-            printf("error_after_opt:  %.6f\n", error_after_opt);
-            printf("used_mem_before_opt: %zu bytes\n", used_mem_before_opt);
-            printf("used_mem_after_opt:  %zu bytes\n", used_mem_after_opt);
-        }
-
-        if (params.print_details_interval > 0 && ex % params.print_details_interval == 0) {
-            // set_logits_masked(logits, token_notavail, -1e9);
-            for (int i=0; i<n_batch; ++i) {
-                init_sampler(&sampler, lctx);
-                for (int k=0; k<n_tokens; ++k) {
-                    int32_t token = sample(&sampler,
-                        (float *)       ((char *) logits->data + i*logits->nb[2] + k*logits->nb[1]),
-                        (llama_token *) ((char *) tokens_input->data + i*tokens_input->nb[1]),
-                        k);
-                    * ((int32_t *) ((char *) after_opt_best_samples->data + i*after_opt_best_samples->nb[1] + k*after_opt_best_samples->nb[0])) = token;
-                }
-            }
-
-            // printf("probabilities after optimization:\n");
-            // print_matrix(after_opt_probs);
-            printf("Example:\n---\n");
-            print_tokens_batch(lctx, tokens_input);
-            printf("\n---\n");
-
-            // printf("best samples after optimization:\n---\n");
-            printf("samples after optimization:\n---\n");
-            print_tokens_batch(lctx, after_opt_best_samples);
-            printf("\n---\n");
-        }
-
-        ggml_free(ctx0);
+        save_train_files(&save_data, train);
+        opt_cb_data.last_save_iter = opt->iter;
     }
 
-    if (params.n_examples > 0) {
-        save_checkpoint(&model, opt, params.fn_checkpoint_out);
+    if (alloc) {
+        ggml_allocr_free(alloc);
     }
 
-    if (strlen(params.fn_model_out) > 0) {
-        save_as_llama_model(&vocab, &model, params.fn_model_out);
-    }
-
-    {
-        int n_gen = params.n_predict;
-        int sample_ctx = n_tokens - n_tokens/8;
-
-        sampler.params.temp = 0.2f;
-        sampler.params.repeat_penalty = 1.1f;
-        sampler.params.mirostat = 2;
-        init_sampler(&sampler, lctx);
-
-        printf("Generating %d tokens.\n", n_gen);
-
-        struct ggml_tensor * tokens_input  = ggml_new_tensor_1d(model.ctx, GGML_TYPE_I32, n_tokens);
-        struct ggml_tensor * target_logits = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab,  n_tokens);
-        struct ggml_tensor * target_probs  = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab,  n_tokens);
-
-        get_example_targets(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), rand()%train_samples.size(), tokens_input, target_logits, target_probs);
-        for (int i=sample_ctx; i<n_tokens; ++i) {
-            ggml_set_i32_1d(tokens_input, i, n_vocab/2);
-        }
-
-        for (int i=0; i<sample_ctx-1; ++i) {
-            print_token(lctx, ggml_get_i32_1d(tokens_input, i));
-        }
-
-        printf("---\n");
-        for (int i=0; i<n_gen; ++i) {
-            struct ggml_init_params cparams = {
-                /*.mem_size   =*/ compute_size,
-                /*.mem_buffer =*/ compute_addr,
-                /*.no_alloc   =*/ false,
-            };
-            struct ggml_context * ctx0 = ggml_init(cparams);
-
-            ggml_cgraph gf = {};
-
-            int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
-
-            ggml_build_forward_expand(&gf, logits);
-            ggml_graph_compute_helper(work_buffer, &gf, params.n_threads);
-
-            //struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
-            //struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);
-
-            // set_logits_masked(logits, token_notavail, -1e9);
-            int token = sample(&sampler,
-                (float *) ((char *) logits->data + (sample_ctx-1)*logits->nb[1]),
-                (llama_token *) tokens_input->data,
-                sample_ctx-1);
-            //int token = ggml_get_i32_1d(best_samples, sample_ctx-1);
-
-            // print_row(probs, sample_at);
-            print_token(lctx, token);
-
-            lshift_examples(tokens_input, target_logits, target_probs, 1);
-            ggml_set_i32_1d(tokens_input, 0, 0);
-            ggml_set_i32_1d(tokens_input, sample_ctx-1, token);
-
-            ggml_free(ctx0);
-        }
-    }
-
-    delete[] compute_addr;
-    delete[] compute_buf_0;
-    delete[] compute_buf_1;
-
+    ggml_free(opt->ctx);
+    free_train_state(train);
+    ggml_free(model.ctx);
     llama_free(lctx);
     llama_free_model(lmodel);
-    ggml_free(model.ctx);
-
     return 0;
 }
diff --git a/flake.lock b/flake.lock
index 33164e096..a7777d05d 100644
--- a/flake.lock
+++ b/flake.lock
@@ -5,11 +5,11 @@
         "systems": "systems"
       },
       "locked": {
-        "lastModified": 1685518550,
-        "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=",
+        "lastModified": 1692799911,
+        "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=",
         "owner": "numtide",
         "repo": "flake-utils",
-        "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef",
+        "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1685931219,
-        "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=",
+        "lastModified": 1692913444,
+        "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "7409480d5c8584a1a83c422530419efe4afb0d19",
+        "rev": "18324978d632ffc55ef1d928e81630c620f4f447",
         "type": "github"
       },
       "original": {
diff --git a/flake.nix b/flake.nix
index 616b90252..cfc4776a4 100644
--- a/flake.nix
+++ b/flake.nix
@@ -6,6 +6,9 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
+        name = "llama.cpp";
+        src = ./.;
+        meta.mainProgram = "llama";
         inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
         buildInputs = with pkgs; [ openmpi ];
         osSpecific = with pkgs; buildInputs ++
@@ -21,11 +24,31 @@
               CoreGraphics
               CoreVideo
             ]
+          else if isDarwin then
+            with pkgs.darwin.apple_sdk.frameworks; [
+              Accelerate
+              CoreGraphics
+              CoreVideo
+            ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
+        nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ];
+        cudatoolkit_joined = with pkgs; symlinkJoin {
+          # HACK(Green-Sky): nix currently has issues with cmake findcudatoolkit
+          # see https://github.com/NixOS/nixpkgs/issues/224291
+          # copied from jaxlib
+          name = "${cudaPackages.cudatoolkit.name}-merged";
+          paths = [
+            cudaPackages.cudatoolkit.lib
+            cudaPackages.cudatoolkit.out
+          ] ++ lib.optionals (lib.versionOlder cudaPackages.cudatoolkit.version "11") [
+            # for some reason some of the required libs are in the targets/x86_64-linux
+            # directory; not sure why but this works around it
+            "${cudaPackages.cudatoolkit}/targets/${system}"
+          ];
+        };
         llama-python =
           pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
         postPatch = ''
@@ -36,37 +59,47 @@
         postInstall = ''
           mv $out/bin/main $out/bin/llama
           mv $out/bin/server $out/bin/llama-server
+          mkdir -p $out/include
+          cp ${src}/llama.h $out/include/
         '';
-        cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in {
+        cmakeFlags = [ "-DLLAMA_NATIVE=OFF" "-DLLAMA_BUILD_SERVER=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
+      in
+      {
         packages.default = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
+          inherit name src meta postPatch nativeBuildInputs postInstall;
           buildInputs = osSpecific;
           cmakeFlags = cmakeFlags
             ++ (if isAarch64 && isDarwin then [
-              "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-              "-DLLAMA_METAL=ON"
-            ] else [
-              "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=OpenBLAS"
+            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
+            "-DLLAMA_METAL=ON"
+          ] else [
+            "-DLLAMA_BLAS=ON"
+            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
         };
         packages.opencl = pkgs.stdenv.mkDerivation {
-          name = "llama.cpp";
-          src = ./.;
-          postPatch = postPatch;
-          nativeBuildInputs = nativeBuildInputs;
+          inherit name src meta postPatch nativeBuildInputs postInstall;
           buildInputs = with pkgs; buildInputs ++ [ clblast ];
           cmakeFlags = cmakeFlags ++ [
             "-DLLAMA_CLBLAST=ON"
           ];
-          postInstall = postInstall;
-          meta.mainProgram = "llama";
+        };
+        packages.cuda = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ cudatoolkit_joined ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_CUBLAS=ON"
+          ];
+        };
+        packages.rocm = pkgs.stdenv.mkDerivation {
+          inherit name src meta postPatch nativeBuildInputs postInstall;
+          buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ];
+          cmakeFlags = cmakeFlags ++ [
+            "-DLLAMA_HIPBLAS=1"
+            "-DCMAKE_C_COMPILER=hipcc"
+            "-DCMAKE_CXX_COMPILER=hipcc"
+            "-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
+          ];
         };
         apps.llama-server = {
           type = "app";
@@ -80,8 +113,17 @@
           type = "app";
           program = "${self.packages.${system}.default}/bin/llama";
         };
+        apps.quantize = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/quantize";
+        };
+        apps.train-text-from-scratch = {
+          type = "app";
+          program = "${self.packages.${system}.default}/bin/train-text-from-scratch";
+        };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
+          buildInputs = [ llama-python ];
           packages = nativeBuildInputs ++ osSpecific;
         };
       });
diff --git a/ggml-alloc.c b/ggml-alloc.c
index f06f9a3c1..805759db7 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -6,8 +6,29 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
+
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
 
 //#define GGML_ALLOCATOR_DEBUG
 
@@ -56,7 +77,7 @@ struct free_block {
     size_t size;
 };
 
-#define MAX_FREE_BLOCKS 128
+#define MAX_FREE_BLOCKS 256
 
 struct ggml_allocr {
     void * data;
@@ -67,8 +88,8 @@ struct ggml_allocr {
     struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE];
     size_t max_size;
     bool measure;
-    int parse_seq[GGML_MAX_NODES];
-    bool has_parse_seq;
+    int parse_seq[GGML_MAX_CONCUR];
+    int parse_seq_len;
 
 #ifdef GGML_ALLOCATOR_DEBUG
     struct ggml_tensor * allocated_tensors[1024];
@@ -98,15 +119,28 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(alloc);
 }
 
+// check if a tensor is allocated by this buffer
+static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+}
+
+static bool ggml_is_view(struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+#ifdef GGML_ALLOCATOR_DEBUG
+    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
+    GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
+#endif
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -130,14 +164,14 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
-            max_avail = MAX(max_avail, block->size);
         } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
-        return;
+            return;
         }
     }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
@@ -153,6 +187,7 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     }
 
     tensor->data = addr;
+    AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
@@ -172,19 +207,20 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     void * ptr = tensor->data;
 
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+    if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
         return;
     }
 
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
-    AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: alloc->data = %p alloc->data+alloc->size = %p alloc->data+alloc->max_size = %p\n", __func__, alloc->data, (char*)alloc->data + alloc->size, (char*)alloc->data + alloc->max_size);
 
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
@@ -238,15 +274,11 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t
     alloc->n_free_blocks++;
 }
 
-void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n) {
-    int pos = 0;
+void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) {
     for (int i = 0; i < n; i++) {
-        if (list[i] != -1) {
-            alloc->parse_seq[pos] = list[i];
-            pos++;
-        }
+        alloc->parse_seq[i] = list[i];
     }
-    alloc->has_parse_seq = true;
+    alloc->parse_seq_len = n;
 }
 
 void ggml_allocr_reset(struct ggml_allocr * alloc) {
@@ -269,9 +301,9 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
         /*.max_size      = */ 0,
         /*.measure       = */ false,
         /*.parse_seq     = */ {0},
-        /*.has_parse_seq = */ false,
+        /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
+        /*.allocated_tensors = */ {0},
 #endif
     };
 
@@ -280,17 +312,68 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// address and size of the buffer when measuring
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+// OS specific functions to allocate and free uncommitted virtual memory
+static void * alloc_vmem(size_t size) {
+#if defined(_WIN32)
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
+}
+
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 128GB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
+
     *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
-        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.data          = */ base_addr,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -298,9 +381,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
         /*.max_size      = */ 0,
         /*.measure       = */ true,
         /*.parse_seq     = */ {0},
-        /*.has_parse_seq = */ false,
+        /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
+        /*.allocated_tensors = */ {0},
 #endif
     };
 
@@ -310,6 +393,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
+    }
     free(alloc);
 }
 
@@ -319,11 +405,6 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 
 //////////// compute graph allocator
 
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
-}
-
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
     if (a->type != b->type) {
         return false;
@@ -339,28 +420,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -368,7 +427,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -378,7 +436,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_CONT:
             return true;
@@ -392,24 +449,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            size_t offset;
-            switch(node->op) {
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
+            assert(node->view_src->data != NULL);
+            node->data = (char *)node->view_src->data + node->view_offs;
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -420,8 +461,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if ((char *) parent->data < (char *) alloc->data ||
-                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                    if (ggml_allocr_is_own(alloc, parent) == false) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
@@ -429,7 +469,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct ggml_tensor * view_src = parent->view_src;
                             struct hash_node * view_src_hn = hash_get(ht, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
@@ -445,8 +485,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                         else {
                             AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
                             node->data = parent->data;
+                            return;
                         }
-                        return;
                     }
                 }
             }
@@ -455,7 +495,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocator_alloc_graph_tensors_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -471,7 +511,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             struct ggml_tensor * node = gf->nodes[i];
 
             if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
+                struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
             }
 
@@ -497,76 +537,92 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 allocate_node(alloc, input);
             }
         }
-        for (int ind = 0; ind < gf->n_nodes; ind++) {
-            int i;
-            if (alloc->has_parse_seq) {
-                i = alloc->parse_seq[ind];
-            } else {
-                i = ind;
-            }
-            struct ggml_tensor * node = gf->nodes[i];
+        // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
+        int last_barrier_pos = 0;
+        int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes;
 
-            // allocate parents (leafs)
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                allocate_node(alloc, parent);
-            }
+        for (int ind = 0; ind < n_nodes; ind++) {
+            // allocate a node if there is no parse_seq or this is not a barrier
+            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) {
+                int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind;
+                struct ggml_tensor * node = gf->nodes[i];
 
-            // allocate node
-            allocate_node(alloc, node);
+                // allocate parents (leafs)
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * parent = node->src[j];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    allocate_node(alloc, parent);
+                }
 
-            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                AT_PRINTF("%s", parent->name);
-                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                    AT_PRINTF(", ");
+                // allocate node
+                allocate_node(alloc, node);
+
+                AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * parent = node->src[j];
+                    if (parent == NULL) {
+                        break;
+                    }
+                    AT_PRINTF("%s", parent->name);
+                    if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                        AT_PRINTF(", ");
+                    }
                 }
+                AT_PRINTF("\n");
             }
-            AT_PRINTF("\n");
 
             // update parents
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
+            // update immediately if there is no parse_seq
+            // update only at barriers if there is parse_seq
+            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
+                int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
+                int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
+                for (int i = update_start; i < update_end; i++) {
+                    int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i;
+                    struct ggml_tensor * node = gf->nodes[node_i];
+
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * parent = node->src[j];
+                        if (parent == NULL) {
+                            break;
+                        }
+                        struct hash_node * p_hn = hash_get(ht, parent);
+                        p_hn->n_children -= 1;
+
+                        //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
+
+                        if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                            if (ggml_is_view(parent)) {
+                                struct ggml_tensor * view_src = parent->view_src;
+                                struct hash_node * view_src_hn = hash_get(ht, view_src);
+                                view_src_hn->n_views -= 1;
+                                AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
+                                if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
+                                    ggml_allocr_free_tensor(alloc, view_src);
+                                }
+                            }
+                            else {
+                                if (parent->data != node->data) {
+                                    ggml_allocr_free_tensor(alloc, parent);
+                                }
+                            }
+                        }
+                    }
                 }
-                struct hash_node * p_hn = hash_get(ht, parent);
-                p_hn->n_children -= 1;
-
-                //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
-                        struct ggml_tensor * view_src = get_view_source(parent);
-                        struct hash_node * view_src_hn = hash_get(ht, view_src);
-                        view_src_hn->n_views -= 1;
-                        AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views);
-                        if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, view_src);
-                        }
-                    }
-                    else {
-                        if (parent->data != node->data) {
-                            ggml_allocator_free_tensor(alloc, parent);
-                        }
-                    }
+                AT_PRINTF("\n");
+                if (alloc->parse_seq_len) {
+                    last_barrier_pos = ind + 1;
                 }
             }
-            AT_PRINTF("\n");
         }
         // free graph outputs here that wouldn't be freed otherwise because they have no children
         if (outputs != NULL && outputs[g] != NULL) {
             for (int i = 0; outputs[g][i] != NULL; i++) {
                 struct ggml_tensor * output = outputs[g][i];
                 AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
+                ggml_allocr_free_tensor(alloc, output);
             }
         }
     }
@@ -575,5 +631,9 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+}
+
+size_t ggml_allocr_max_size(struct ggml_allocr * alloc) {
+    return alloc->max_size;
 }
diff --git a/ggml-alloc.h b/ggml-alloc.h
index 14a4350ac..0c224f174 100644
--- a/ggml-alloc.h
+++ b/ggml-alloc.h
@@ -12,13 +12,14 @@ GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);
 
 // tell the allocator to parse nodes following the order described in the list
 // you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n);
+GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n);
 
 GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);
 GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);
 GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);
 GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);
 GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);
+GGML_API size_t ggml_allocr_max_size(struct ggml_allocr * alloc);
 
 
 #ifdef  __cplusplus
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 70a950bb5..989c419cd 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -1,3 +1,4 @@
+#include <algorithm>
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -6,15 +7,148 @@
 #include <atomic>
 #include <assert.h>
 
+#if defined(GGML_USE_HIPBLAS)
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasCreate hipblasCreate
+#define cublasGemmEx hipblasGemmEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#else
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
 #include <cuda_fp16.h>
+#endif // defined(GGML_USE_HIPBLAS)
 
 #include "ggml-cuda.h"
 #include "ggml.h"
 
-#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define CC_TURING   700
+#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define CC_VOLTA      700
+#define CC_OFFSET_AMD 1000000
+#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
+
+#if defined(GGML_USE_HIPBLAS)
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int&>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int&>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
+#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
+    c = __builtin_amdgcn_sdot4(a, b, c, false);
+#elif defined(__gfx1100__)
+    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+#elif defined(__gfx1010__) || defined(__gfx900__)
+    int tmp1;
+    int tmp2;
+    asm("\n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        "
+        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
+        : "v"(a), "v"(b)
+    );
+#else
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
+#endif
+    return c;
+}
+#endif // defined(GGML_USE_HIPBLAS)
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -26,8 +160,11 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
     do {                                                                                \
         cudaError_t err_ = (err);                                                       \
         if (err_ != cudaSuccess) {                                                      \
-            fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__,   \
+            int id;                                                                     \
+            cudaGetDevice(&id);                                                         \
+            fprintf(stderr, "\nCUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
                 cudaGetErrorString(err_));                                              \
+            fprintf(stderr, "current device: %d\n", id);                                \
             exit(1);                                                                    \
         }                                                                               \
     } while (0)
@@ -37,8 +174,11 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
     do {                                                                                \
         cublasStatus_t err_ = (err);                                                    \
         if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
+            int id;                                                                     \
+            cudaGetDevice(&id);                                                         \
             fprintf(stderr, "\ncuBLAS error %d at %s:%d: %s\n",                         \
                     err_, __FILE__, __LINE__, cublasGetStatusString(err_));             \
+            fprintf(stderr, "current device: %d\n", id);                                \
             exit(1);                                                                    \
         }                                                                               \
     } while (0)
@@ -47,12 +187,21 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
     do {                                                                                \
         cublasStatus_t err_ = (err);                                                    \
         if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
+            int id;                                                                     \
+            cudaGetDevice(&id);                                                         \
             fprintf(stderr, "\ncuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__);  \
+            fprintf(stderr, "current device: %d\n", id);                                \
             exit(1);                                                                    \
         }                                                                               \
     } while (0)
 #endif // CUDART_VERSION >= 11
 
+#if CUDART_VERSION >= 11100
+#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
+#else
+#define GGML_CUDA_ASSUME(x)
+#endif // CUDART_VERSION >= 11100
+
 #ifdef GGML_CUDA_F16
 typedef half dfloat; // dequantize float
 typedef half2 dfloat2;
@@ -89,15 +238,22 @@ static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t *
     return *((int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
 }
 
+template<typename T>
+using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream);
+typedef to_t_cuda_t<float> to_fp32_cuda_t;
+typedef to_t_cuda_t<half> to_fp16_cuda_t;
+
 typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
-typedef void (*to_fp32_cuda_t)(const void * __restrict__ x, float * __restrict__ y, int k, cudaStream_t stream);
 typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v);
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
-typedef void (*ggml_cuda_op_t)(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i, float * src0_ddf_i,
-    float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main);
+typedef void (*ggml_cuda_op_mul_mat_t)(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, const cudaStream_t & stream);
+typedef void (*ggml_cuda_op_flatten_t)(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream);
 
 // QK = number of values after dequantization
 // QR = QK / number of values before dequantization
@@ -205,11 +361,11 @@ typedef struct {
 #define QI4_K (QK_K / (4*QR4_K))
 #ifdef GGML_QKK_64
 typedef struct {
-    half    d[2];              // super-block scales/mins
+    half    dm[2];             // super-block scales/mins
     uint8_t scales[2];         // 4-bit block scales/mins
     uint8_t qs[QK_K/2];        // 4--bit quants
 } block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
 #else
 typedef struct {
     half2 dm;                  // super-block scale for quantized scales/mins
@@ -278,11 +434,33 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
 #endif
 
+#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
+#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
+#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
+
+#define MUL_MAT_SRC1_COL_STRIDE 128
+
+#define MAX_STREAMS 8
+static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { nullptr };
+
 struct ggml_tensor_extra_gpu {
     void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
-    cudaEvent_t events[GGML_CUDA_MAX_DEVICES]; // events for synchronizing multiple GPUs
+    cudaEvent_t events[GGML_CUDA_MAX_DEVICES][MAX_STREAMS]; // events for synchronizing multiple GPUs
 };
 
+// this is faster on Windows
+// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
+inline cudaError_t ggml_cuda_set_device(const int device) {
+    int current_device;
+    CUDA_CHECK(cudaGetDevice(&current_device));
+
+    if (device == current_device) {
+        return cudaSuccess;
+    }
+
+    return cudaSetDevice(device);
+}
+
 static int g_device_count = -1;
 static int g_main_device = 0;
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
@@ -290,13 +468,11 @@ static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 static bool g_mul_mat_q = true;
 
 static void * g_scratch_buffer = nullptr;
-static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default
+static size_t g_scratch_size = 0; // disabled by default
 static size_t g_scratch_offset = 0;
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-static cudaStream_t g_cudaStreams_main[GGML_CUDA_MAX_DEVICES] = { nullptr };
-
 static __global__ void add_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -346,58 +522,91 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) {
     dst[i] = x[i] / (1.0f + expf(-x[i]));
 }
 
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+template <int block_size>
 static __global__ void norm_f32(const float * x, float * dst, const int ncols) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
     const float eps = 1e-5f;
 
-    float mean = 0.0f;
-    float var = 0.0f;
+    float2 mean_var = make_float2(0.f, 0.f);
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[row*ncols + col];
-        mean += xi;
-        var += xi * xi;
+        mean_var.x += xi;
+        mean_var.y += xi * xi;
     }
 
     // sum up partial sums
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        mean += __shfl_xor_sync(0xffffffff, mean, mask, 32);
-        var += __shfl_xor_sync(0xffffffff, var, mask, 32);
+    mean_var = warp_reduce_sum(mean_var);
+    if (block_size > WARP_SIZE) {
+        __shared__ float2 s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        __syncthreads();
+        mean_var = s_sum[lane_id];
+        mean_var = warp_reduce_sum(mean_var);
     }
 
-    mean /= ncols;
-    var = var / ncols - mean * mean;
-    const float inv_var = rsqrtf(var + eps);
+    const float mean = mean_var.x / ncols;
+    const float var = mean_var.y / ncols - mean * mean;
+    const float inv_std = rsqrtf(var + eps);
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
-        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_var;
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
     }
 }
 
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+template <int block_size>
 static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
     const int row = blockIdx.x*blockDim.y + threadIdx.y;
     const int tid = threadIdx.x;
 
     float tmp = 0.0f; // partial sum for thread in warp
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         const float xi = x[row*ncols + col];
         tmp += xi * xi;
     }
 
     // sum up partial sums
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
     }
 
     const float mean = tmp / ncols;
     const float scale = rsqrtf(mean + eps);
 
-    for (int col = tid; col < ncols; col += WARP_SIZE) {
+    for (int col = tid; col < ncols; col += block_size) {
         dst[row*ncols + col] = scale * x[row*ncols + col];
     }
 }
@@ -424,8 +633,8 @@ static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
 
     const int vui = x[ib].qs[iqs];
 
@@ -467,8 +676,8 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
 static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const dfloat d = x[ib].dm.x;
-    const dfloat m = x[ib].dm.y;
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -506,7 +715,8 @@ static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const in
 
 //================================== k-quants
 
-static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float * __restrict__ yy) {
+template<typename dst_t>
+static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
     const int i   = blockIdx.x;
     const block_q2_K * x = (const block_q2_K *) vx;
@@ -518,10 +728,10 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
     const int is  = 8*n + l/16;
 
     const uint8_t q = x[i].qs[32*n + l];
-    float * y = yy + i*QK_K + 128*n;
+    dst_t * y = yy + i*QK_K + 128*n;
 
-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
     y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
@@ -530,16 +740,17 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float
     const int is = tid/16;  // 0 or 1
     const int il = tid%16;  // 0...15
     const uint8_t q = x[i].qs[il] >> (2*is);
-    float * y = yy + i*QK_K + 16*is + il;
-    float dall = x[i].dm.x;
-    float dmin = x[i].dm.y;
+    dst_t * y = yy + i*QK_K + 16*is + il;
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
     y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
     y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
 #endif
 
 }
 
-static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, float * __restrict__ yy) {
+template<typename dst_t>
+static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
     const int i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
@@ -563,7 +774,7 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, float
     float d_all = x[i].d;
     float dl = d_all * (us - 32);
 
-    float * y = yy + i*QK_K + 128*n + 32*j;
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
     const uint8_t * q = x[i].qs + 32*n;
     const uint8_t * hm = x[i].hmask;
 
@@ -575,7 +786,7 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, float
     const int im  = il/8;    // 0...1
     const int in  = il%8;    // 0...7
 
-    float * y = yy + i*QK_K + 16*is + il;
+    dst_t * y = yy + i*QK_K + 16*is + il;
 
     const uint8_t q = x[i].qs[il] >> (2*is);
     const uint8_t h = x[i].hmask[in] >> (2*is + im);
@@ -603,7 +814,8 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
 }
 #endif
 
-static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float * __restrict__ yy) {
+template<typename dst_t>
+static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q4_K * x = (const block_q4_K *) vx;
 
     const int i = blockIdx.x;
@@ -616,10 +828,10 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float
     const int is  = 2*il;
     const int n   = 4;
 
-    float * y = yy + i*QK_K + 64*il + n*ir;
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
 
-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
 
     const uint8_t * q = x[i].qs + 32*il + n*ir;
 
@@ -635,15 +847,16 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float
 #else
     const int tid = threadIdx.x;
     const uint8_t * q = x[i].qs;
-    float * y = yy + i*QK_K;
-    const float d = (float)x[i].d[0];
-    const float m = (float)x[i].d[1];
+    dst_t * y = yy + i*QK_K;
+    const float d = (float)x[i].dm[0];
+    const float m = (float)x[i].dm[1];
     y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
     y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
 #endif
 }
 
-static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float * __restrict__ yy) {
+template<typename dst_t>
+static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q5_K * x = (const block_q5_K *) vx;
 
     const int i = blockIdx.x;
@@ -655,10 +868,10 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float
     const int ir  = tid%16;   // ir is in 0...15
     const int is  = 2*il;     // is is in 0...6
 
-    float * y = yy + i*QK_K + 64*il + 2*ir;
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
 
-    const float dall = x[i].dm.x;
-    const float dmin = x[i].dm.y;
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
 
     const uint8_t * ql = x[i].qs + 32*il + 2*ir;
     const uint8_t * qh = x[i].qh + 2*ir;
@@ -683,13 +896,14 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float
     const int is = tid/16; // 0 or 1
     const uint8_t h = x[i].qh[in] >> im;
     const float d = x[i].d;
-    float * y = yy + i*QK_K + tid;
+    dst_t * y = yy + i*QK_K + tid;
     y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
     y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
 #endif
 }
 
-static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, float * __restrict__ yy) {
+template<typename dst_t>
+static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const block_q6_K * x = (const block_q6_K *) vx;
 
     const int i = blockIdx.x;
@@ -701,7 +915,7 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, float
     const int il  = tid - 32*ip; // 0...32
     const int is  = 8*ip + il/16;
 
-    float * y = yy + i*QK_K + 128*ip + il;
+    dst_t * y = yy + i*QK_K + 128*ip + il;
 
     const float d = x[i].d;
 
@@ -720,7 +934,7 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, float
     const int ip  = tid/16;         // 0 or 1
     const int il  = tid - 16*ip;    // 0...15
 
-    float * y = yy + i*QK_K + 16*ip + il;
+    dst_t * y = yy + i*QK_K + 16*ip + il;
 
     const float d = x[i].d;
 
@@ -770,8 +984,8 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
         const float   * y = yy + i * QK_K + y_offset;
         const uint8_t * q = x[i].qs + q_offset;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
         aux[0] = a[0] & 0x0f0f0f0f;
@@ -991,8 +1205,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         const float   * y1 = yy + i*QK_K + y_offset;
         const float   * y2 = y1 + 128;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux[0] = a[im+0] & kmask1;
@@ -1054,8 +1268,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux16[0] = a[0] & 0x0f0f;
         aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].d[0];
-        const float m = (float)x[i].d[1];
+        const float d = (float)x[i].dm[0];
+        const float m = (float)x[i].dm[1];
         float sum = 0.f;
         for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
             sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
@@ -1124,8 +1338,8 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
         const float   * y1  = yy + i*QK_K + y_offset;
         const float   * y2  = y1 + 128;
 
-        const float dall = x[i].dm.x;
-        const float dmin = x[i].dm.y;
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
 
         const uint16_t * a = (const uint16_t *)x[i].scales;
         aux[0] = a[im+0] & kmask1;
@@ -1313,6 +1527,14 @@ static __device__ void convert_f16(const void * vx, const int ib, const int iqs,
     v.y = x[ib + iqs + 1];
 }
 
+static __device__ void convert_f32(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const float * x = (const float *) vx;
+
+    // automatic half -> float type cast if dfloat == float
+    v.x = x[ib + iqs + 0];
+    v.y = x[ib + iqs + 1];
+}
+
 static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) {
     const int ix = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -1348,12 +1570,12 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
         return;
     }
 
-    y[ib].ds.x = d;
-    y[ib].ds.y = sum;
+    reinterpret_cast<half&>(y[ib].ds.x) = d;
+    reinterpret_cast<half&>(y[ib].ds.y) = sum;
 }
 
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
-static __global__ void dequantize_block(const void * __restrict__ vx, float * __restrict__ y, const int k) {
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
     const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
 
     if (i >= k) {
@@ -1956,10 +2178,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI4_0;
     const int kqsx = k % QI4_0;
@@ -2050,10 +2272,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI4_1;
     const int kqsx = k % QI4_1;
@@ -2142,10 +2364,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI5_0;
     const int kqsx = k % QI5_0;
@@ -2256,10 +2478,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset < nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset < nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI5_1;
     const int kqsx = k % QI5_1;
@@ -2346,7 +2568,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
         u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
     }
 
-    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, bq8_1->ds.x);
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
 }
 
 template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
@@ -2362,10 +2584,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI8_0;
     const int kqsx = k % QI8_0;
@@ -2432,7 +2654,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR2_K; ++ i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@@ -2453,10 +2675,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI2_K;
     const int kqsx = k % QI2_K;
@@ -2551,7 +2773,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR3_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
     }
 
     return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@@ -2574,10 +2796,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI3_K;
     const int kqsx = k % QI3_K;
@@ -2720,7 +2942,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
 
     for (int i = 0; i < QR4_K; ++i) {
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2half(bq8i->ds);
 
         const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
         u[2*i+0] = q8[0];
@@ -2744,11 +2966,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     aux16[0] = a[0] & 0x0f0f;
     aux16[1] = (a[0] >> 4) & 0x0f0f;
 
-    const float dall = bq4_K->d[0];
-    const float dmin = bq4_K->d[1];
+    const float dall = bq4_K->dm[0];
+    const float dmin = bq4_K->dm[1];
 
-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2float(bq8_1[0].ds);
+    const float d8_2 = __low2float(bq8_1[1].ds);
 
     const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
     const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@@ -2792,10 +3014,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI4_K; // == 0 if QK_K == 256
     const int kqsx = k % QI4_K; // == k if QK_K == 256
@@ -2828,7 +3050,11 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
+#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
+#else
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
+#endif
     }
 
 #pragma unroll
@@ -2901,7 +3127,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR5_K; ++i) {
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds.x;
+        d8[i] = __low2float(bq8i->ds);
 
         const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
         u[2*i+0] = q8[0];
@@ -2919,8 +3145,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 
     const float d = bq5_K->d;
 
-    const float d8_1 = bq8_1[0].ds.x;
-    const float d8_2 = bq8_1[1].ds.x;
+    const float d8_1 = __low2half(bq8_1[0].ds);
+    const float d8_2 = __low2half(bq8_1[1].ds);
 
     const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
     const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
@@ -2969,10 +3195,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI5_K; // == 0 if QK_K == 256
     const int kqsx = k % QI5_K; // == k if QK_K == 256
@@ -3018,7 +3244,9 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
+#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
+#endif
     }
 
 #pragma unroll
@@ -3075,7 +3303,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 #pragma unroll
     for (int i = 0; i < QR6_K; ++i) {
         u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = bq8_1[bq8_offset + 2*i].ds.x;
+        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
     }
 
     return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@@ -3096,10 +3324,10 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
     const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
     int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
 
-    __builtin_assume(i_offset >= 0);
-    __builtin_assume(i_offset <  nwarps);
-    __builtin_assume(k >= 0);
-    __builtin_assume(k <  WARP_SIZE);
+    GGML_CUDA_ASSUME(i_offset >= 0);
+    GGML_CUDA_ASSUME(i_offset <  nwarps);
+    GGML_CUDA_ASSUME(k >= 0);
+    GGML_CUDA_ASSUME(k <  WARP_SIZE);
 
     const int kbx  = k / QI6_K; // == 0 if QK_K == 256
     const int kqsx = k % QI6_K; // == k if QK_K == 256
@@ -3243,7 +3471,7 @@ static __device__ __forceinline__ void mul_mat_q(
                     *dsi_dst = *dsi_src;
                 } else {
                     float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = (*dsi_src).x;
+                    *dfi_dst = __low2half(*dsi_src);
                 }
             }
 
@@ -3287,6 +3515,12 @@ static __device__ __forceinline__ void mul_mat_q(
     }
 }
 
+#define  MMQ_X_Q4_0_RDNA2  64
+#define  MMQ_Y_Q4_0_RDNA2  128
+#define NWARPS_Q4_0_RDNA2  8
+#define  MMQ_X_Q4_0_RDNA1  64
+#define  MMQ_Y_Q4_0_RDNA1  64
+#define NWARPS_Q4_0_RDNA1  8
 #define  MMQ_X_Q4_0_AMPERE 64
 #define  MMQ_Y_Q4_0_AMPERE 128
 #define NWARPS_Q4_0_AMPERE 4
@@ -3294,11 +3528,32 @@ static __device__ __forceinline__ void mul_mat_q(
 #define  MMQ_Y_Q4_0_PASCAL 64
 #define NWARPS_Q4_0_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q4_0(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q4_0(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_0_RDNA2;
+    const int nwarps = NWARPS_Q4_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_0_RDNA1;
+    const int nwarps = NWARPS_Q4_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
+        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q4_0_AMPERE;
     const int mmq_y  =  MMQ_Y_Q4_0_AMPERE;
     const int nwarps = NWARPS_Q4_0_AMPERE;
@@ -3318,9 +3573,15 @@ template <bool need_check> static __global__ void mul_mat_q4_0(
 #else
     (void) vec_dot_q4_0_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q4_1_RDNA2  64
+#define  MMQ_Y_Q4_1_RDNA2  128
+#define NWARPS_Q4_1_RDNA2  8
+#define  MMQ_X_Q4_1_RDNA1  64
+#define  MMQ_Y_Q4_1_RDNA1  64
+#define NWARPS_Q4_1_RDNA1  8
 #define  MMQ_X_Q4_1_AMPERE 64
 #define  MMQ_Y_Q4_1_AMPERE 128
 #define NWARPS_Q4_1_AMPERE 4
@@ -3329,14 +3590,33 @@ template <bool need_check> static __global__ void mul_mat_q4_0(
 #define NWARPS_Q4_1_PASCAL 8
 
 template <bool need_check> static __global__ void
-#if __CUDA_ARCH__ < CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
     __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_TURING
+#endif // __CUDA_ARCH__ < CC_VOLTA
     mul_mat_q4_1(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_1_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_1_RDNA2;
+    const int nwarps = NWARPS_Q4_1_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_1_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_1_RDNA1;
+    const int nwarps = NWARPS_Q4_1_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
+        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q4_1_AMPERE;
     const int mmq_y  =  MMQ_Y_Q4_1_AMPERE;
     const int nwarps = NWARPS_Q4_1_AMPERE;
@@ -3356,9 +3636,15 @@ template <bool need_check> static __global__ void
 #else
     (void) vec_dot_q4_1_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q5_0_RDNA2  64
+#define  MMQ_Y_Q5_0_RDNA2  128
+#define NWARPS_Q5_0_RDNA2  8
+#define  MMQ_X_Q5_0_RDNA1  64
+#define  MMQ_Y_Q5_0_RDNA1  64
+#define NWARPS_Q5_0_RDNA1  8
 #define  MMQ_X_Q5_0_AMPERE 128
 #define  MMQ_Y_Q5_0_AMPERE 64
 #define NWARPS_Q5_0_AMPERE 4
@@ -3366,11 +3652,32 @@ template <bool need_check> static __global__ void
 #define  MMQ_Y_Q5_0_PASCAL 64
 #define NWARPS_Q5_0_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q5_0(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q5_0(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_0_RDNA2;
+    const int nwarps = NWARPS_Q5_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_0_RDNA1;
+    const int nwarps = NWARPS_Q5_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
+        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q5_0_AMPERE;
     const int mmq_y  =  MMQ_Y_Q5_0_AMPERE;
     const int nwarps = NWARPS_Q5_0_AMPERE;
@@ -3390,9 +3697,15 @@ template <bool need_check> static __global__ void mul_mat_q5_0(
 #else
     (void) vec_dot_q5_0_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q5_1_RDNA2  64
+#define  MMQ_Y_Q5_1_RDNA2  128
+#define NWARPS_Q5_1_RDNA2  8
+#define  MMQ_X_Q5_1_RDNA1  64
+#define  MMQ_Y_Q5_1_RDNA1  64
+#define NWARPS_Q5_1_RDNA1  8
 #define  MMQ_X_Q5_1_AMPERE 128
 #define  MMQ_Y_Q5_1_AMPERE 64
 #define NWARPS_Q5_1_AMPERE 4
@@ -3400,11 +3713,32 @@ template <bool need_check> static __global__ void mul_mat_q5_0(
 #define  MMQ_Y_Q5_1_PASCAL 64
 #define NWARPS_Q5_1_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q5_1(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q5_1(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_1_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_1_RDNA2;
+    const int nwarps = NWARPS_Q5_1_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_1_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_1_RDNA1;
+    const int nwarps = NWARPS_Q5_1_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
+        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q5_1_AMPERE;
     const int mmq_y  =  MMQ_Y_Q5_1_AMPERE;
     const int nwarps = NWARPS_Q5_1_AMPERE;
@@ -3424,9 +3758,15 @@ template <bool need_check> static __global__ void mul_mat_q5_1(
 #else
     (void) vec_dot_q5_1_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q8_0_RDNA2  64
+#define  MMQ_Y_Q8_0_RDNA2  128
+#define NWARPS_Q8_0_RDNA2  8
+#define  MMQ_X_Q8_0_RDNA1  64
+#define  MMQ_Y_Q8_0_RDNA1  64
+#define NWARPS_Q8_0_RDNA1  8
 #define  MMQ_X_Q8_0_AMPERE 128
 #define  MMQ_Y_Q8_0_AMPERE 64
 #define NWARPS_Q8_0_AMPERE 4
@@ -3434,11 +3774,32 @@ template <bool need_check> static __global__ void mul_mat_q5_1(
 #define  MMQ_Y_Q8_0_PASCAL 64
 #define NWARPS_Q8_0_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q8_0(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    mul_mat_q8_0(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q8_0_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q8_0_RDNA2;
+    const int nwarps = NWARPS_Q8_0_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q8_0_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q8_0_RDNA1;
+    const int nwarps = NWARPS_Q8_0_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
+        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q8_0_AMPERE;
     const int mmq_y  =  MMQ_Y_Q8_0_AMPERE;
     const int nwarps = NWARPS_Q8_0_AMPERE;
@@ -3458,9 +3819,15 @@ template <bool need_check> static __global__ void mul_mat_q8_0(
 #else
     (void) vec_dot_q8_0_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q2_K_RDNA2  64
+#define  MMQ_Y_Q2_K_RDNA2  128
+#define NWARPS_Q2_K_RDNA2  8
+#define  MMQ_X_Q2_K_RDNA1  128
+#define  MMQ_Y_Q2_K_RDNA1  32
+#define NWARPS_Q2_K_RDNA1  8
 #define  MMQ_X_Q2_K_AMPERE 64
 #define  MMQ_Y_Q2_K_AMPERE 128
 #define NWARPS_Q2_K_AMPERE 4
@@ -3468,11 +3835,32 @@ template <bool need_check> static __global__ void mul_mat_q8_0(
 #define  MMQ_Y_Q2_K_PASCAL 64
 #define NWARPS_Q2_K_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q2_K(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q2_K(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q2_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q2_K_RDNA2;
+    const int nwarps = NWARPS_Q2_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q2_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q2_K_RDNA1;
+    const int nwarps = NWARPS_Q2_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
+        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q2_K_AMPERE;
     const int mmq_y  =  MMQ_Y_Q2_K_AMPERE;
     const int nwarps = NWARPS_Q2_K_AMPERE;
@@ -3492,9 +3880,15 @@ template <bool need_check> static __global__ void mul_mat_q2_K(
 #else
     (void) vec_dot_q2_K_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q3_K_RDNA2  128
+#define  MMQ_Y_Q3_K_RDNA2  64
+#define NWARPS_Q3_K_RDNA2  8
+#define  MMQ_X_Q3_K_RDNA1  32
+#define  MMQ_Y_Q3_K_RDNA1  128
+#define NWARPS_Q3_K_RDNA1  8
 #define  MMQ_X_Q3_K_AMPERE 128
 #define  MMQ_Y_Q3_K_AMPERE 128
 #define NWARPS_Q3_K_AMPERE 4
@@ -3503,14 +3897,33 @@ template <bool need_check> static __global__ void mul_mat_q2_K(
 #define NWARPS_Q3_K_PASCAL 8
 
 template <bool need_check> static __global__ void
-#if __CUDA_ARCH__ < CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
     __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_TURING
+#endif // __CUDA_ARCH__ < CC_VOLTA
     mul_mat_q3_K(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q3_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q3_K_RDNA2;
+    const int nwarps = NWARPS_Q3_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q3_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q3_K_RDNA1;
+    const int nwarps = NWARPS_Q3_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
+        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q3_K_AMPERE;
     const int mmq_y  =  MMQ_Y_Q3_K_AMPERE;
     const int nwarps = NWARPS_Q3_K_AMPERE;
@@ -3530,9 +3943,15 @@ template <bool need_check> static __global__ void
 #else
     (void) vec_dot_q3_K_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q4_K_RDNA2  64
+#define  MMQ_Y_Q4_K_RDNA2  128
+#define NWARPS_Q4_K_RDNA2  8
+#define  MMQ_X_Q4_K_RDNA1  32
+#define  MMQ_Y_Q4_K_RDNA1  64
+#define NWARPS_Q4_K_RDNA1  8
 #define  MMQ_X_Q4_K_AMPERE 64
 #define  MMQ_Y_Q4_K_AMPERE 128
 #define NWARPS_Q4_K_AMPERE 4
@@ -3541,14 +3960,33 @@ template <bool need_check> static __global__ void
 #define NWARPS_Q4_K_PASCAL 8
 
 template <bool need_check> static __global__ void
-#if __CUDA_ARCH__ < CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
     __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_TURING
+#endif // __CUDA_ARCH__ < CC_VOLTA
     mul_mat_q4_K(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q4_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q4_K_RDNA2;
+    const int nwarps = NWARPS_Q4_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q4_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q4_K_RDNA1;
+    const int nwarps = NWARPS_Q4_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
+        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q4_K_AMPERE;
     const int mmq_y  =  MMQ_Y_Q4_K_AMPERE;
     const int nwarps = NWARPS_Q4_K_AMPERE;
@@ -3568,9 +4006,15 @@ template <bool need_check> static __global__ void
 #else
     (void) vec_dot_q4_K_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q5_K_RDNA2  64
+#define  MMQ_Y_Q5_K_RDNA2  128
+#define NWARPS_Q5_K_RDNA2  8
+#define  MMQ_X_Q5_K_RDNA1  32
+#define  MMQ_Y_Q5_K_RDNA1  64
+#define NWARPS_Q5_K_RDNA1  8
 #define  MMQ_X_Q5_K_AMPERE 64
 #define  MMQ_Y_Q5_K_AMPERE 128
 #define NWARPS_Q5_K_AMPERE 4
@@ -3578,11 +4022,32 @@ template <bool need_check> static __global__ void
 #define  MMQ_Y_Q5_K_PASCAL 64
 #define NWARPS_Q5_K_PASCAL 8
 
-template <bool need_check> static __global__ void mul_mat_q5_K(
+template <bool need_check> static __global__ void
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+mul_mat_q5_K(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q5_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q5_K_RDNA2;
+    const int nwarps = NWARPS_Q5_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q5_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q5_K_RDNA1;
+    const int nwarps = NWARPS_Q5_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
+        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q5_K_AMPERE;
     const int mmq_y  =  MMQ_Y_Q5_K_AMPERE;
     const int nwarps = NWARPS_Q5_K_AMPERE;
@@ -3602,9 +4067,15 @@ template <bool need_check> static __global__ void mul_mat_q5_K(
 #else
     (void) vec_dot_q5_K_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
+#define  MMQ_X_Q6_K_RDNA2  64
+#define  MMQ_Y_Q6_K_RDNA2  128
+#define NWARPS_Q6_K_RDNA2  8
+#define  MMQ_X_Q6_K_RDNA1  32
+#define  MMQ_Y_Q6_K_RDNA1  64
+#define NWARPS_Q6_K_RDNA1  8
 #define  MMQ_X_Q6_K_AMPERE 64
 #define  MMQ_Y_Q6_K_AMPERE 64
 #define NWARPS_Q6_K_AMPERE 4
@@ -3613,14 +4084,33 @@ template <bool need_check> static __global__ void mul_mat_q5_K(
 #define NWARPS_Q6_K_PASCAL 8
 
 template <bool need_check> static __global__ void
-#if __CUDA_ARCH__ < CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#elif __CUDA_ARCH__ < CC_VOLTA
     __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_TURING
+#endif // __CUDA_ARCH__ < CC_VOLTA
     mul_mat_q6_K(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
 
-#if __CUDA_ARCH__ >= CC_TURING
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    const int mmq_x  =  MMQ_X_Q6_K_RDNA2;
+    const int mmq_y  =  MMQ_Y_Q6_K_RDNA2;
+    const int nwarps = NWARPS_Q6_K_RDNA2;
+#else
+    const int mmq_x  =  MMQ_X_Q6_K_RDNA1;
+    const int mmq_y  =  MMQ_Y_Q6_K_RDNA1;
+    const int nwarps = NWARPS_Q6_K_RDNA1;
+#endif // defined(RDNA3) || defined(RDNA2)
+
+    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
+        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
+        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
+
+#elif __CUDA_ARCH__ >= CC_VOLTA
     const int mmq_x  =  MMQ_X_Q6_K_AMPERE;
     const int mmq_y  =  MMQ_Y_Q6_K_AMPERE;
     const int nwarps = NWARPS_Q6_K_AMPERE;
@@ -3640,7 +4130,7 @@ template <bool need_check> static __global__ void
 #else
     (void) vec_dot_q6_K_q8_1_mul_mat;
     assert(false);
-#endif // __CUDA_ARCH__ >= CC_TURING
+#endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
 template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
@@ -3885,8 +4375,10 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
 }
 
 // rope == RoPE == rotary positional embedding
-static __global__ void rope_f32(const float * x, float * dst, const int ncols, const float p0,
-                                const float p_delta, const int p_delta_rows, const float theta_scale) {
+
+template<typename T, bool has_pos>
+static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale,
+                            const int p_delta_rows, const float theta_scale) {
     const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (col >= ncols) {
@@ -3895,8 +4387,11 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c
 
     const int row = blockDim.x*blockIdx.x + threadIdx.x;
     const int i = row*ncols + col;
+    const int i2 = row/p_delta_rows;
 
-    const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2);
+    const int p = has_pos ? pos[i2] : 0;
+    const float p0 = p*freq_scale;
+    const float theta = p0*powf(theta_scale, col/2);
     const float sin_theta = sinf(theta);
     const float cos_theta = cosf(theta);
 
@@ -3907,7 +4402,34 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c
     dst[i + 1] = x0*sin_theta + x1*cos_theta;
 }
 
-static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) {
+template<typename T, bool has_pos>
+static __global__ void rope_neox(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale,
+                                 const int p_delta_rows, const float theta_scale) {
+    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i = row*ncols + col/2;
+    const int i2 = row/p_delta_rows;
+
+    const int p = has_pos ? pos[i2] : 0;
+    const float p0 = p*freq_scale;
+    const float theta = p0*powf(theta_scale, col/2);
+    const float sin_theta = sinf(theta);
+    const float cos_theta = cosf(theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + ncols/2];
+
+    dst[i + 0]       = x0*cos_theta - x1*sin_theta;
+    dst[i + ncols/2] = x0*sin_theta + x1*cos_theta;
+}
+
+static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const int32_t * pos, const float freq_scale,
+                                    const int p_delta_rows, const float theta_scale, const int n_ctx) {
     const int col = blockDim.x*blockIdx.x + threadIdx.x;
     const int half_n_dims = ncols/4;
 
@@ -3917,10 +4439,13 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol
 
     const int row = blockDim.y*blockIdx.y + threadIdx.y;
     const int i = row*ncols + col;
+    const int i2 = row/p_delta_rows;
 
     const float col_theta_scale = powf(theta_scale, col);
+     // FIXME: this is likely wrong
+    const int p = pos != nullptr ? pos[i2] : 0;
 
-    const float theta = p*col_theta_scale;
+    const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale;
     const float sin_theta = sinf(theta);
     const float cos_theta = cosf(theta);
 
@@ -3930,7 +4455,7 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol
     dst[i + 0]           = x0*cos_theta - x1*sin_theta;
     dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta;
 
-    const float block_theta = block_p*col_theta_scale;
+    const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale;
     const float sin_block_theta = sinf(block_theta);
     const float cos_block_theta = cosf(block_theta);
 
@@ -4057,14 +4582,24 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
 
 static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols);
+    }
 }
 
 static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
     GGML_ASSERT(ncols % WARP_SIZE == 0);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    rms_norm_f32<<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
 }
 
 static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) {
@@ -4074,32 +4609,38 @@ static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, con
     quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
 }
 
-static void dequantize_row_q4_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q4_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK4_1, QR4_1, dequantize_q4_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q5_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q5_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK5_0, QR5_0, dequantize_q5_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q5_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q5_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK5_1, QR5_1, dequantize_q5_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q8_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
     dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q2_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
 #if QK_K == 256
     dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
@@ -4108,7 +4649,8 @@ static void dequantize_row_q2_K_cuda(const void * vx, float * y, const int k, cu
 #endif
 }
 
-static void dequantize_row_q3_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
 #if QK_K == 256
     dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
@@ -4117,12 +4659,14 @@ static void dequantize_row_q3_K_cuda(const void * vx, float * y, const int k, cu
 #endif
 }
 
-static void dequantize_row_q4_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
     dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
 }
 
-static void dequantize_row_q5_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
 #if QK_K == 256
     dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
@@ -4131,7 +4675,8 @@ static void dequantize_row_q5_K_cuda(const void * vx, float * y, const int k, cu
 #endif
 }
 
-static void dequantize_row_q6_K_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+template<typename dst_t>
+static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
     const int nb = k / QK_K;
 #if QK_K == 256
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
@@ -4322,6 +4867,11 @@ static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, c
     dequantize_block<1, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
+static void convert_fp32_to_fp16_cuda(const void * vx, half * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    dequantize_block<1, 1, convert_f32><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
 static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
@@ -4331,6 +4881,35 @@ static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, floa
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
+static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_row_q5_0_cuda;
+        case GGML_TYPE_Q5_1:
+            return dequantize_row_q5_1_cuda;
+        case GGML_TYPE_Q8_0:
+            return dequantize_row_q8_0_cuda;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_F32:
+            return convert_fp32_to_fp16_cuda;
+        default:
+            return nullptr;
+    }
+}
+
 static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -4369,7 +4948,15 @@ static void ggml_mul_mat_q4_0_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA2;
+        nwarps = NWARPS_Q4_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA1;
+        nwarps = NWARPS_Q4_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q4_0_AMPERE;
         mmq_y  =  MMQ_Y_Q4_0_AMPERE;
         nwarps = NWARPS_Q4_0_AMPERE;
@@ -4406,7 +4993,15 @@ static void ggml_mul_mat_q4_1_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA2;
+        nwarps = NWARPS_Q4_1_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA1;
+        nwarps = NWARPS_Q4_1_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q4_1_AMPERE;
         mmq_y  =  MMQ_Y_Q4_1_AMPERE;
         nwarps = NWARPS_Q4_1_AMPERE;
@@ -4443,7 +5038,15 @@ static void ggml_mul_mat_q5_0_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA2;
+        nwarps = NWARPS_Q5_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA1;
+        nwarps = NWARPS_Q5_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q5_0_AMPERE;
         mmq_y  =  MMQ_Y_Q5_0_AMPERE;
         nwarps = NWARPS_Q5_0_AMPERE;
@@ -4480,7 +5083,15 @@ static void ggml_mul_mat_q5_1_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA2;
+        nwarps = NWARPS_Q5_1_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA1;
+        nwarps = NWARPS_Q5_1_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q5_1_AMPERE;
         mmq_y  =  MMQ_Y_Q5_1_AMPERE;
         nwarps = NWARPS_Q5_1_AMPERE;
@@ -4517,7 +5128,15 @@ static void ggml_mul_mat_q8_0_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA2;
+        nwarps = NWARPS_Q8_0_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA1;
+        nwarps = NWARPS_Q8_0_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q8_0_AMPERE;
         mmq_y  =  MMQ_Y_Q8_0_AMPERE;
         nwarps = NWARPS_Q8_0_AMPERE;
@@ -4554,7 +5173,15 @@ static void ggml_mul_mat_q2_K_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA2;
+        nwarps = NWARPS_Q2_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA1;
+        nwarps = NWARPS_Q2_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q2_K_AMPERE;
         mmq_y  =  MMQ_Y_Q2_K_AMPERE;
         nwarps = NWARPS_Q2_K_AMPERE;
@@ -4586,12 +5213,22 @@ static void ggml_mul_mat_q3_K_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
 
+#if QK_K == 256
+
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA2;
+        nwarps = NWARPS_Q3_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA1;
+        nwarps = NWARPS_Q3_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q3_K_AMPERE;
         mmq_y  =  MMQ_Y_Q3_K_AMPERE;
         nwarps = NWARPS_Q3_K_AMPERE;
@@ -4617,6 +5254,7 @@ static void ggml_mul_mat_q3_K_q8_1_cuda(
         mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
             (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
     }
+#endif
 }
 
 static void ggml_mul_mat_q4_K_q8_1_cuda(
@@ -4628,7 +5266,15 @@ static void ggml_mul_mat_q4_K_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA2;
+        nwarps = NWARPS_Q4_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA1;
+        nwarps = NWARPS_Q4_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q4_K_AMPERE;
         mmq_y  =  MMQ_Y_Q4_K_AMPERE;
         nwarps = NWARPS_Q4_K_AMPERE;
@@ -4665,7 +5311,15 @@ static void ggml_mul_mat_q5_K_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA2;
+        nwarps = NWARPS_Q5_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA1;
+        nwarps = NWARPS_Q5_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q5_K_AMPERE;
         mmq_y  =  MMQ_Y_Q5_K_AMPERE;
         nwarps = NWARPS_Q5_K_AMPERE;
@@ -4702,7 +5356,15 @@ static void ggml_mul_mat_q6_K_q8_1_cuda(
     const int compute_capability = g_compute_capabilities[id];
 
     int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_TURING) {
+    if (compute_capability >= CC_RDNA2) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA2;
+        nwarps = NWARPS_Q6_K_RDNA2;
+    } else if (compute_capability >= CC_OFFSET_AMD) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA1;
+        nwarps = NWARPS_Q6_K_RDNA1;
+    } else if (compute_capability >= CC_VOLTA) {
         mmq_x  =  MMQ_X_Q6_K_AMPERE;
         mmq_y  =  MMQ_Y_Q6_K_AMPERE;
         nwarps = NWARPS_Q6_K_AMPERE;
@@ -4774,21 +5436,41 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons
     scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
 }
 
-static void rope_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0,
-                          const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
-    GGML_ASSERT(nrows % 2 == 0);
-    const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1);
+template<typename T>
+static void rope_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
+                          const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
+    GGML_ASSERT(ncols % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nrows, num_blocks_x, 1);
-    rope_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale);
+    if (pos == nullptr) {
+        rope<T, false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+    } else {
+        rope<T, true><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+    }
 }
 
-static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float block_p, const float theta_scale, cudaStream_t stream) {
-    GGML_ASSERT(nrows % 4 == 0);
-    const dim3 block_dims(4*CUDA_ROPE_BLOCK_SIZE, 1, 1);
-    const int num_blocks_x = (ncols + 4*CUDA_ROPE_BLOCK_SIZE - 1) / (4*CUDA_ROPE_BLOCK_SIZE);
+template<typename T>
+static void rope_neox_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
+                          const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
+    GGML_ASSERT(ncols % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nrows, num_blocks_x, 1);
+    if (pos == nullptr) {
+        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+    } else {
+        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale);
+    }
+}
+
+static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
+                              const int p_delta_rows, const float theta_scale, const int n_ctx, cudaStream_t stream) {
+    GGML_ASSERT(ncols % 4 == 0);
+    const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
+    const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
     const dim3 block_nums(num_blocks_x, nrows, 1);
-    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, p, block_p, theta_scale);
+    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale, n_ctx);
 }
 
 static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
@@ -4914,29 +5596,42 @@ void ggml_init_cublas() {
     static bool initialized = false;
 
     if (!initialized) {
+
+#ifdef __HIP_PLATFORM_AMD__
+        // Workaround for a rocBLAS bug when using multiple graphics cards:
+        // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
+        rocblas_initialize();
+        CUDA_CHECK(cudaDeviceSynchronize());
+#endif
+
         CUDA_CHECK(cudaGetDeviceCount(&g_device_count));
         GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
         int64_t total_vram = 0;
-        fprintf(stderr, "%s: found %d CUDA devices:\n", __func__, g_device_count);
-        for (int id = 0; id < g_device_count; ++id) {
+        fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
+        for (int64_t id = 0; id < g_device_count; ++id) {
             cudaDeviceProp prop;
             CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-            fprintf(stderr, "  Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
+            fprintf(stderr, "  Device %ld: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
 
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
-
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            g_compute_capabilities[id] = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
+#else
             g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         }
-        for (int id = 0; id < g_device_count; ++id) {
+        for (int64_t id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
         }
 
-        for (int id = 0; id < g_device_count; ++id) {
-            CUDA_CHECK(cudaSetDevice(id));
+        for (int64_t id = 0; id < g_device_count; ++id) {
+            CUDA_CHECK(ggml_cuda_set_device(id));
 
-            // create main stream
-            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams_main[id], cudaStreamNonBlocking));
+            // create cuda streams
+            for (int64_t is = 0; is < MAX_STREAMS; ++is) {
+                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking));
+            }
 
             // create cublas handle
             CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
@@ -5005,7 +5700,8 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
     if (src->backend == GGML_BACKEND_CPU) {
         kind = cudaMemcpyHostToDevice;
         src_ptr = (char *) src->data;
-    } else if (src->backend == GGML_BACKEND_GPU) {
+    } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
+        GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
         kind = cudaMemcpyDeviceToDevice;
         struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
         int id;
@@ -5044,240 +5740,209 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
 }
 
 inline void ggml_cuda_op_add(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddq_i != nullptr || src0_ddf_i != nullptr);
-    GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i  != nullptr);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
 
-    // compute
     if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        add_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne00*i01_diff, ne10*ne11, cudaStream_main);
+        add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        add_f16_f32_f16_cuda((half *) src0_ddq_i, src1_ddf_i, (half *) dst_ddf_i, ne00*i01_diff, cudaStream_main);
+        add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream);
     } else {
         GGML_ASSERT(false);
     }
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) i02;
-    (void) i1;
 }
 
 inline void ggml_cuda_op_mul(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i  != nullptr);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
 
-    mul_f32_cuda(src0_ddf_i, src1_ddf_i, dst_ddf_i, ne00*i01_diff, ne10*ne11, cudaStream_main);
+    mul_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream);
 
     (void) dst;
-    (void) src0_ddq_i;
-    (void) i02;
-    (void) i1;
 }
 
 inline void ggml_cuda_op_gelu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
-
-    // compute
-    gelu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main);
+    gelu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_silu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
-
-    // compute
-    silu_f32_cuda(src0_ddf_i, dst_ddf_i, ne00*i01_diff, cudaStream_main);
+    silu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_norm(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t nrows = ggml_nrows(src0);
 
-    // compute
-    norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
+    norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_rms_norm(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t nrows = ggml_nrows(src0);
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    // compute
-    rms_norm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, eps, cudaStream_main);
+    rms_norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_mul_mat_q(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
-
-    GGML_ASSERT(src0_ddq_i != nullptr);
-    GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, const cudaStream_t & stream) {
 
     const int64_t ne00 = src0->ne[0];
 
     const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
     GGML_ASSERT(ne10 % QK8_1 == 0);
 
     const int64_t ne0 = dst->ne[0];
 
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t row_diff = row_high - row_low;
 
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into
-    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : i01_diff;
-
-    const int64_t padded_row_size = ne10 % MATRIX_ROW_PADDING == 0 ?
-        ne10 : ne10 - ne10 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
-    size_t as;
-    void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*ne11*sizeof(block_q8_1)/QK8_1, &as);
-    quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne10, ne11, padded_row_size, cudaStream_main);
+    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
 
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-            ggml_mul_mat_q4_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q4_1:
-            ggml_mul_mat_q4_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q5_0:
-            ggml_mul_mat_q5_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q5_1:
-            ggml_mul_mat_q5_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q8_0:
-            ggml_mul_mat_q8_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q2_K:
-            ggml_mul_mat_q2_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q3_K:
-            ggml_mul_mat_q3_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q4_K:
-            ggml_mul_mat_q4_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q5_K:
-            ggml_mul_mat_q5_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         case GGML_TYPE_Q6_K:
-            ggml_mul_mat_q6_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, i01_diff, ne11, padded_row_size, nrows_dst, cudaStream_main);
+            ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
             break;
         default:
             GGML_ASSERT(false);
             break;
     }
 
-    ggml_cuda_pool_free(src1_q8_1, as);
-
     (void) src1;
     (void) dst;
-    (void) src0_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_ddf_i;
 }
 
 static int64_t get_row_rounding(ggml_type type) {
-    int max_compute_capability = INT_MIN;
-    for (int id = 0; id < g_device_count; ++id) {
-        if (max_compute_capability < g_compute_capabilities[id]
-                && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
-            max_compute_capability = g_compute_capabilities[id];
+    int64_t min_compute_capability = INT_MAX;
+    int64_t max_compute_capability = INT_MIN;
+    for (int64_t id = 0; id < g_device_count; ++id) {
+        if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
+            if (min_compute_capability > g_compute_capabilities[id]) {
+                min_compute_capability = g_compute_capabilities[id];
+            }
+            if (max_compute_capability < g_compute_capabilities[id]) {
+                max_compute_capability = g_compute_capabilities[id];
+            }
         }
     }
 
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
     switch(type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
-            return max_compute_capability >= CC_TURING ? 128 : 64;
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
+        case GGML_TYPE_F16:
+            return 1;
+        case GGML_TYPE_Q2_K:
+            return max_compute_capability >= CC_RDNA2 ? 128 : 32;
+        case GGML_TYPE_Q3_K:
+            return min_compute_capability < CC_RDNA2 ? 128 : 64;
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
+        default:
+            GGML_ASSERT(false);
+    }
+#else
+    switch(type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
@@ -5288,222 +5953,271 @@ static int64_t get_row_rounding(ggml_type type) {
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
-            return max_compute_capability >= CC_TURING ? 128 : 64;
+            return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
         default:
             GGML_ASSERT(false);
     }
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 }
 
-inline void ggml_cuda_op_mul_mat_vec(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
-
-    GGML_ASSERT(src0_ddq_i != nullptr);
-    GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+inline void ggml_cuda_op_mul_mat_vec_q(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, const cudaStream_t & stream) {
 
     const int64_t ne00 = src0->ne[0];
-    const int64_t nrows = i01_high - i01_low;
+    const int64_t row_diff = row_high - row_low;
 
-#ifdef GGML_CUDA_FORCE_DMMV
-    const bool use_mul_mat_vec_q = false;
-    (void) g_compute_capabilities[0];
-#else
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-
-    bool mul_mat_vec_q_implemented =
-        src0->type == GGML_TYPE_Q4_0 ||
-        src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 ||
-        src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0;
-#if QK_K == 256
-    mul_mat_vec_q_implemented = mul_mat_vec_q_implemented ||
-        src0->type == GGML_TYPE_Q2_K ||
-        src0->type == GGML_TYPE_Q3_K ||
-        src0->type == GGML_TYPE_Q4_K ||
-        src0->type == GGML_TYPE_Q5_K ||
-        src0->type == GGML_TYPE_Q6_K;
-#endif // QK_K == 256
-
-    const bool use_mul_mat_vec_q = g_compute_capabilities[id] >= MIN_CC_DP4A && mul_mat_vec_q_implemented;
-#endif
-
-    if (use_mul_mat_vec_q) {
-        const int64_t padded_row_size = ne00 % MATRIX_ROW_PADDING == 0 ?
-            ne00 : ne00 - ne00 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
-        size_t as;
-        void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*sizeof(block_q8_1)/QK8_1, &as);
-        quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne00, 1, padded_row_size, cudaStream_main);
-
-        switch (src0->type) {
-            case GGML_TYPE_Q4_0:
-                mul_mat_vec_q4_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q4_1:
-                mul_mat_vec_q4_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_0:
-                mul_mat_vec_q5_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_1:
-                mul_mat_vec_q5_1_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q8_0:
-                mul_mat_vec_q8_0_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q2_K:
-                mul_mat_vec_q2_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q3_K:
-                mul_mat_vec_q3_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q4_K:
-                mul_mat_vec_q4_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_K:
-                mul_mat_vec_q5_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q6_K:
-                mul_mat_vec_q6_K_q8_1_cuda(src0_ddq_i, src1_q8_1, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-
-        ggml_cuda_pool_free(src1_q8_1, as);
-    } else {
-        // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_F16
-        size_t ash;
-        dfloat * src1_dfloat = nullptr; // dfloat == half
-
-        bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-            src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-            src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-        if (src1_convert_f16) {
-            src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
-            ggml_cpy_f32_f16_cuda((char *) src1_ddf_i, (char *) src1_dfloat, ne00,
-                                    ne00, 1, sizeof(float), 0, 0,
-                                    ne00, 1, sizeof(half),  0, 0, cudaStream_main);
-        }
-#else
-        dfloat * src1_dfloat = src1_ddf_i; // dfloat == float, no conversion
-#endif // GGML_CUDA_F16
-
-        switch (src0->type) {
-            case GGML_TYPE_Q4_0:
-                dequantize_mul_mat_vec_q4_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q4_1:
-                dequantize_mul_mat_vec_q4_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_0:
-                dequantize_mul_mat_vec_q5_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_1:
-                dequantize_mul_mat_vec_q5_1_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q8_0:
-                dequantize_mul_mat_vec_q8_0_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q2_K:
-                dequantize_mul_mat_vec_q2_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q3_K:
-                dequantize_mul_mat_vec_q3_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q4_K:
-                dequantize_mul_mat_vec_q4_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q5_K:
-                dequantize_mul_mat_vec_q5_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_Q6_K:
-                dequantize_mul_mat_vec_q6_K_cuda(src0_ddq_i, src1_ddf_i, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            case GGML_TYPE_F16:
-                convert_mul_mat_vec_f16_cuda(src0_ddq_i, src1_dfloat, dst_ddf_i, ne00, nrows, cudaStream_main);
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-
-#ifdef GGML_CUDA_F16
-        if (src1_convert_f16) {
-            ggml_cuda_pool_free(src1_dfloat, ash);
-        }
-#endif // GGML_CUDA_F16
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
     }
 
     (void) src1;
     (void) dst;
-    (void) src0_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_ddf_i;
+    (void) src1_ncols;
+    (void) src1_padded_row_size;
+}
+
+inline void ggml_cuda_op_dequantize_mul_mat_vec(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, const cudaStream_t & stream) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_CUDA_F16
+    size_t ash;
+    dfloat * src1_dfloat = nullptr; // dfloat == half
+
+    bool src1_convert_f16 = src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+    if (src1_convert_f16) {
+        src1_dfloat = (half *) ggml_cuda_pool_malloc(ne00*sizeof(half), &ash);
+        ggml_cpy_f32_f16_cuda((const char *) src1_ddf_i, (char *) src1_dfloat, ne00,
+                                ne00, 1, sizeof(float), 0, 0,
+                                ne00, 1, sizeof(half),  0, 0, stream);
+    }
+#else
+    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
+#endif // GGML_CUDA_F16
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_F16:
+            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+
+#ifdef GGML_CUDA_F16
+    if (src1_convert_f16) {
+        ggml_cuda_pool_free(src1_dfloat, ash);
+    }
+#endif // GGML_CUDA_F16
+
+    (void) src1;
+    (void) dst;
+    (void) src1_ddq_i;
+    (void) src1_ncols;
+    (void) src1_padded_row_size;
 }
 
 inline void ggml_cuda_op_mul_mat_cublas(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, const cudaStream_t & stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
+    GGML_ASSERT(src0_dd_i != nullptr);
     GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(dst_dd_i != nullptr);
 
-    const float alpha = 1.0f;
-    const float beta = 0.0f;
 
     const int64_t ne00 = src0->ne[0];
 
     const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
 
     const int64_t ne0 = dst->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t row_diff = row_high - row_low;
 
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
 
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : i01_diff;
+    int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
 
-    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], cudaStream_main));
-    CUBLAS_CHECK(
-        cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
-                i01_diff, ne11, ne10,
-                &alpha, src0_ddf_i, ne00,
-                        src1_ddf_i, ne10,
-                &beta,  dst_ddf_i,  ldc));
+    const int compute_capability = g_compute_capabilities[id];
+
+    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
+        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
+        half * src0_as_f16 = nullptr;
+        size_t src0_as = 0;
+        if (src0->type != GGML_TYPE_F16) {
+            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
+            GGML_ASSERT(to_fp16_cuda != nullptr);
+            size_t ne = row_diff*ne00;
+            src0_as_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &src0_as);
+            to_fp16_cuda(src0_dd_i, src0_as_f16, ne, stream);
+        }
+        const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16;
+
+        half * src1_as_f16 = nullptr;
+        size_t src1_as = 0;
+        if (src1->type != GGML_TYPE_F16) {
+            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+            GGML_ASSERT(to_fp16_cuda != nullptr);
+            size_t ne = src1_ncols*ne10;
+            src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &src1_as);
+            to_fp16_cuda(src1_ddf_i, src1_as_f16, ne, stream);
+        }
+        const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16;
+
+        size_t dst_as = 0;
+        half * dst_f16 = (half *) ggml_cuda_pool_malloc(row_diff*src1_ncols * sizeof(half), &dst_as);
+
+        const half alpha_f16 = 1.0f;
+        const half beta_f16 = 0.0f;
+
+        CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
+        CUBLAS_CHECK(
+            cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                    row_diff, src1_ncols, ne10,
+                    &alpha_f16, src0_ptr, CUDA_R_16F, ne00,
+                                src1_ptr, CUDA_R_16F, ne10,
+                    &beta_f16,   dst_f16, CUDA_R_16F, ldc,
+                    CUBLAS_COMPUTE_16F,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+        to_fp32_cuda(dst_f16, dst_dd_i, row_diff*src1_ncols, stream);
+
+        ggml_cuda_pool_free(dst_f16, dst_as);
+
+        if (src0_as != 0) {
+            ggml_cuda_pool_free(src0_as_f16, src0_as);
+        }
+
+        if (src1_as != 0) {
+            ggml_cuda_pool_free(src1_as_f16, src1_as);
+        }
+    }
+    else {
+        float * src0_ddq_as_f32 = nullptr;
+        size_t src0_as = 0;
+
+        if (src0->type != GGML_TYPE_F32) {
+            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
+            GGML_ASSERT(to_fp32_cuda != nullptr);
+            src0_ddq_as_f32 = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_as); // NOLINT
+            to_fp32_cuda(src0_dd_i, src0_ddq_as_f32, row_diff*ne00, stream);
+        }
+        const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32;
+
+        const float alpha = 1.0f;
+        const float beta = 0.0f;
+
+        CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
+        CUBLAS_CHECK(
+            cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+                    row_diff, src1_ncols, ne10,
+                    &alpha, src0_ddf_i, ne00,
+                            src1_ddf_i,  ne10,
+                    &beta,  dst_dd_i,   ldc));
+
+        if (src0_as != 0) {
+            ggml_cuda_pool_free(src0_ddq_as_f32, src0_as);
+        }
+    }
 
     (void) dst;
-    (void) src0_ddq_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_ddq_i;
+    (void) src1_padded_row_size;
 }
 
 inline void ggml_cuda_op_rope(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t ne2 = dst->ne[2];
+    const int64_t nrows = ggml_nrows(src0);
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_dims = ((int32_t *) dst->op_params)[1];
     const int mode   = ((int32_t *) dst->op_params)[2];
     const int n_ctx  = ((int32_t *) dst->op_params)[3];
@@ -5515,38 +6229,55 @@ inline void ggml_cuda_op_rope(
 
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
-    const bool is_glm = mode & 4;
+    const int32_t * pos = nullptr;
+    if ((mode & 1) == 0) {
+        GGML_ASSERT(src1->type == GGML_TYPE_I32);
+        GGML_ASSERT(src1->ne[0] == ne2);
+        pos = (const int32_t *) src1_dd;
+    }
+
+    const bool is_neox = mode & 2;
+    const bool is_glm  = mode & 4;
 
     // compute
     if (is_glm) {
-        const float p = (((mode & 1) == 0 ? n_past + i02 : i02)) * freq_scale;
-        const float id_p = min(p, n_ctx - 2.f);
-        const float block_p = max(p - (n_ctx - 2.f), 0.f);
-        rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main);
+        GGML_ASSERT(false);
+        rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, n_ctx, main_stream);
+    } else if (is_neox) {
+        GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet");
+        if (src0->type == GGML_TYPE_F32) {
+            rope_neox_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_neox_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+        } else {
+            GGML_ASSERT(false);
+        }
     } else {
-        const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale;
-        rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main);
+        if (src0->type == GGML_TYPE_F32) {
+            rope_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream);
+        } else {
+            GGML_ASSERT(false);
+        }
     }
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_alibi(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t nrows = ggml_nrows(src0);
 
     const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
@@ -5561,334 +6292,393 @@ inline void ggml_cuda_op_alibi(
     const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
-    // compute
-    alibi_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_heads_log2_floor, m0, m1, cudaStream_main);
+    alibi_f32_cuda(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream);
 
     (void) src1;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_diag_mask_inf(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int nrows0 = ggml_nrows(src0);
 
     const int n_past = ((int32_t *) dst->op_params)[0];
 
-    // compute
-    diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main);
+    diag_mask_inf_f32_cuda(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_soft_max(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
+    const int64_t nrows = ggml_nrows(src0);
 
-    // compute
-    soft_max_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, cudaStream_main);
+    soft_max_f32_cuda(src0_dd, dst_dd, ne00, nrows, main_stream);
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
 inline void ggml_cuda_op_scale(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, char * src0_ddq_i,
-    float * src0_ddf_i, float * src1_ddf_i, float * dst_ddf_i, int64_t i02, int64_t i01_low, int64_t i01_high, int i1,
-    cudaStream_t & cudaStream_main){
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
 
-    GGML_ASSERT(src0_ddf_i != nullptr);
-    GGML_ASSERT(dst_ddf_i != nullptr);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     const float scale = ((float *) src1->data)[0];
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t i01_diff = i01_high - i01_low;
-
-    // compute
-    scale_f32_cuda(src0_ddf_i, dst_ddf_i, scale, ne00*i01_diff, cudaStream_main);
+    scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
     CUDA_CHECK(cudaGetLastError());
 
     (void) src1;
     (void) dst;
-    (void) src0_ddq_i;
-    (void) src1_ddf_i;
-    (void) i02;
-    (void) i1;
+    (void) src1_dd;
 }
 
-static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-                         ggml_cuda_op_t op, bool src0_needs_f32, bool flatten_rows) {
+static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) {
+    const int64_t nrows0 = ggml_nrows(src0);
+
+    const bool use_src1 = src1 != nullptr;
+    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
+
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_GPU_SPLIT);
+
+    struct ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
+    struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
+    struct ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
+
+    const bool src0_on_device =             src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
+    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
+    const bool  dst_on_device =              dst->backend == GGML_BACKEND_GPU;
+
+    const bool src1_stays_on_host = use_src1 && dst->op == GGML_OP_SCALE;
+
+    // dd = data device
+    float * src0_ddf = nullptr;
+    float * src1_ddf = nullptr;
+    float *  dst_ddf = nullptr;
+
+    // as = actual size
+    size_t src0_asf = 0;
+    size_t src1_asf = 0;
+    size_t  dst_asf = 0;
+
+    ggml_cuda_set_device(g_main_device);
+    const cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
+
+    if (src0_on_device) {
+        src0_ddf = (float *) src0_extra->data_device[g_main_device];
+    } else {
+        src0_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_asf);
+        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream));
+    }
+
+    if (use_src1 && !src1_stays_on_host) {
+        if (src1_on_device) {
+            src1_ddf = (float *) src1_extra->data_device[g_main_device];
+        } else {
+            src1_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf);
+            CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream));
+        }
+    }
+    if (dst_on_device) {
+        dst_ddf = (float *) dst_extra->data_device[g_main_device];
+    } else {
+        dst_ddf = (float *) ggml_cuda_pool_malloc(ggml_nbytes(dst), &dst_asf);
+    }
+
+    // do the computation
+    op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    // copy dst to host if necessary
+    if (!dst_on_device) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream));
+    }
+
+    if (src0_asf > 0) {
+        ggml_cuda_pool_free(src0_ddf, src0_asf);
+    }
+    if (src1_asf > 0) {
+        ggml_cuda_pool_free(src1_ddf, src1_asf);
+    }
+    if (dst_asf > 0) {
+        ggml_cuda_pool_free(dst_ddf, dst_asf);
+    }
+
+    if (dst->backend == GGML_BACKEND_CPU) {
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
+}
+
+static void ggml_cuda_set_peer_access(const int n_tokens) {
+    static bool peer_access_enabled = false;
+
+    const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
+
+    if (peer_access_enabled == enable_peer_access) {
+        return;
+    }
+
+#ifdef NDEBUG
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(ggml_cuda_set_device(id));
+
+        for (int id_other = 0; id_other < g_device_count; ++id_other) {
+            if (id == id_other) {
+                continue;
+            }
+            if (id != g_main_device && id_other != g_main_device) {
+                continue;
+            }
+
+            int can_access_peer;
+            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
+            if (can_access_peer) {
+                if (enable_peer_access) {
+                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+                } else {
+                    CUDA_CHECK(cudaDeviceDisablePeerAccess(id_other));
+                }
+            }
+        }
+    }
+#endif // NDEBUG
+
+    peer_access_enabled = enable_peer_access;
+}
+
+static void ggml_cuda_op_mul_mat(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
+    const bool convert_src1_to_q8_1) {
+
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
     const int64_t ne03 = src0->ne[3];
     const int64_t nrows0 = ggml_nrows(src0);
 
-    const bool use_src1 = src1 != nullptr;
-    const int64_t ne10 = use_src1 ? src1->ne[0] : 1;
-    const int64_t ne11 = use_src1 ? src1->ne[1] : 1;
-    const int64_t ne12 = use_src1 ? src1->ne[2] : 1;
-    const int64_t ne13 = use_src1 ? src1->ne[3] : 1;
-    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int64_t nrows1 = ggml_nrows(src1);
 
     GGML_ASSERT(ne03 == ne13);
 
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
+    const int nb2 = dst->nb[2];
+    const int nb3 = dst->nb[3];
+
+    ggml_cuda_set_peer_access(ne11);
 
     GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
+    GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
 
-    // strides for iteration over dims 3 and 2
-    const int64_t num_iters_0 = ne02 >= ne12 ? ne02*ne03 : ne12*ne13;
-    const int64_t num_iters = flatten_rows ? 1 : num_iters_0;
-    const int64_t stride_mod = flatten_rows ? num_iters_0 : 1;
-    const int64_t src0_stride = ne00 * ne01 * stride_mod;
-    const int64_t src1_stride = ne10 * ne11 * stride_mod;
-    const int64_t dst_stride = ne0 * ne1 * stride_mod;
+    GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
 
-    const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;
-    const int64_t i03_max = flatten_rows ? 1 : ne03;
-    const int64_t i02_max = flatten_rows ? 1 : (ne02 >= ne12 ? ne02 : ne12);
-    const int64_t i02_divisor = ne02 >= ne12 ? 1 : ne12 / ne02;
-    GGML_ASSERT(!(flatten_rows && ne02 < ne12));
+    const int64_t i02_divisor = ne12 / ne02;
 
     const size_t src0_ts = ggml_type_size(src0->type);
     const size_t src0_bs = ggml_blck_size(src0->type);
+    const size_t q8_1_ts = sizeof(block_q8_1);
+    const size_t q8_1_bs = QK8_1;
 
-    struct ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    struct ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    struct ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *) dst->extra;
+    struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    struct ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
     const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
-    const bool src0_is_f32 = src0->type == GGML_TYPE_F32;
 
-    const bool src1_is_contiguous = use_src1 && ggml_is_contiguous(src1);
-    const bool src1_stays_on_host = use_src1 && (
-        dst->op == GGML_OP_SCALE || dst->op == GGML_OP_DIAG_MASK_INF || dst->op == GGML_OP_ROPE);
+    const bool src1_is_contiguous = ggml_is_contiguous(src1);
+    const int64_t src1_padded_col_size = ne10 % MATRIX_ROW_PADDING == 0 ?
+        ne10 : ne10 - ne10 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
 
     const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+    GGML_ASSERT(!(split && ne02 > 1));
+    GGML_ASSERT(!(split && ne03 > 1));
     GGML_ASSERT(!(split && ne02 < ne12));
 
-    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
-
     // dd = data device
-    char  * src0_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // quantized
-    float * src0_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float
-    float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr};
-    float *  dst_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr};
+    char  *  src0_dd[GGML_CUDA_MAX_DEVICES] = {nullptr};
+    float * src1_ddf[GGML_CUDA_MAX_DEVICES] = {nullptr}; // float
+    char  * src1_ddq[GGML_CUDA_MAX_DEVICES] = {nullptr}; // q8_1
+    float *   dst_dd[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-    // asq = actual size quantized, asf = actual size float
-    size_t src0_asq[GGML_CUDA_MAX_DEVICES] = {0};
-    size_t src0_asf[GGML_CUDA_MAX_DEVICES] = {0};
+    // as = actual size
+    size_t  src0_as[GGML_CUDA_MAX_DEVICES] = {0};
     size_t src1_asf[GGML_CUDA_MAX_DEVICES] = {0};
-    size_t  dst_asf[GGML_CUDA_MAX_DEVICES] = {0};
+    size_t src1_asq[GGML_CUDA_MAX_DEVICES] = {0};
+    size_t   dst_as[GGML_CUDA_MAX_DEVICES] = {0};
 
-    // if multiple devices are used they need to wait for the main device
-    // here an event is recorded that signifies that the main device has finished calculating the input data
-    if (split && g_device_count > 1) {
-        CUDA_CHECK(cudaSetDevice(g_main_device));
-        CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device], g_cudaStreams_main[g_main_device]));
-    }
+    int64_t  row_low[GGML_CUDA_MAX_DEVICES];
+    int64_t row_high[GGML_CUDA_MAX_DEVICES];
 
-    for (int id = 0; id < g_device_count; ++id) {
-        if (!split && id != g_main_device) {
-            continue;
-        }
+    for (int64_t id = 0; id < g_device_count; ++id) {
+        // by default, use all rows
+        row_low[id]  = 0;
+        row_high[id] = ne01;
 
-        const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU && id == g_main_device;
-        const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device;
-
-        int64_t row_low, row_high;
+        // for multi GPU, get the row boundaries from tensor split
+        // and round to mul_mat_q tile sizes
         if (split) {
             const int64_t rounding = get_row_rounding(src0->type);
 
-            row_low = id == 0 ? 0 : nrows0*g_tensor_split[id];
-            row_low -= row_low % rounding;
-
-            if (id == g_device_count - 1) {
-                row_high = nrows0;
-            } else {
-                row_high = nrows0*g_tensor_split[id + 1];
-                row_high -= row_high % rounding;
+            if (id != 0) {
+                row_low[id]  = ne01*g_tensor_split[id];
+                row_low[id] -= row_low[id] % rounding;
+            }
+
+            if (id != g_device_count - 1) {
+                row_high[id]  = ne01*g_tensor_split[id + 1];
+                row_high[id] -= row_high[id] % rounding;
             }
-        } else {
-            row_low = 0;
-            row_high = nrows0*i02_divisor;
         }
-        if (row_low == row_high) {
+    }
+
+    for (int64_t id = 0; id < g_device_count; ++id) {
+        if ((!split && id != g_main_device) || row_low[id] == row_high[id]) {
             continue;
         }
 
-        int64_t row_diff = row_high - row_low;
+        const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
+        const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
 
-        cudaSetDevice(id);
-        cudaStream_t cudaStream_main = g_cudaStreams_main[id];
-
-        // wait for main GPU data if necessary
-        if (split && id != g_main_device) {
-            CUDA_CHECK(cudaStreamWaitEvent(cudaStream_main, src0_extra->events[g_main_device]));
-        }
+        ggml_cuda_set_device(id);
+        const cudaStream_t stream = g_cudaStreams[id][0];
 
         if (src0_on_device && src0_is_contiguous) {
-            if (src0_is_f32) {
-                src0_ddf[id] = (float *) src0_extra->data_device[id];
-            } else {
-                src0_ddq[id] = (char *) src0_extra->data_device[id];
-            }
+            src0_dd[id] = (char *) src0_extra->data_device[id];
         } else {
-            if (src0_is_f32) {
-                src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]);
-            } else {
-                src0_ddq[id] = (char *) ggml_cuda_pool_malloc(row_diff*ne00 * src0_ts/src0_bs, &src0_asq[id]);
+            const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0);
+            src0_dd[id] = (char *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_as[id]);
+        }
+
+        if (src1_on_device && src1_is_contiguous) {
+            src1_ddf[id] = (float *) src1_extra->data_device[id];
+        } else {
+            src1_ddf[id] = (float *) ggml_cuda_pool_malloc(ggml_nbytes(src1), &src1_asf[id]);
+        }
+
+        if (convert_src1_to_q8_1) {
+            src1_ddq[id] = (char *) ggml_cuda_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]);
+
+            if (split && src1_on_device && src1_is_contiguous) {
+                quantize_row_q8_1_cuda(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream);
+                CUDA_CHECK(cudaGetLastError());
             }
         }
 
-        if (src0_needs_f32 && !src0_is_f32) {
-            src0_ddf[id] = (float *) ggml_cuda_pool_malloc(row_diff*ne00 * sizeof(float), &src0_asf[id]);
-        }
-
-        if (use_src1 && !src1_stays_on_host) {
-            if (src1_on_device && src1_is_contiguous) {
-                src1_ddf[id] = (float *) src1_extra->data_device[id];
-            } else {
-                src1_ddf[id] = (float *) ggml_cuda_pool_malloc(num_iters*src1_stride * sizeof(float), &src1_asf[id]);
-            }
-        }
         if (dst_on_device) {
-            dst_ddf[id] = (float *) dst_extra->data_device[id];
+            dst_dd[id] = (float *) dst_extra->data_device[id];
         } else {
-            size_t size_dst_ddf = split ? row_diff*ne1 * sizeof(float) : num_iters*dst_stride * sizeof(float);
-            dst_ddf[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_asf[id]);
+            const size_t size_dst_ddf = split ? (row_high[id]-row_low[id])*ne1*sizeof(float) : ggml_nbytes(dst);
+            dst_dd[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_as[id]);
         }
+    }
 
-        for (int64_t i03 = 0; i03 < i03_max; i03++) {
-            const int64_t i13 = i03 % ne13;
-            for (int64_t i02 = 0; i02 < i02_max; i02++) {
-                const int64_t i12 = i02 % ne12;
+    // if multiple devices are used they need to wait for the main device
+    // here an event is recorded that signals that the main device has finished calculating the input data
+    if (split && g_device_count > 1) {
+        CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+        CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0]));
+    }
 
-                const int64_t i0 = i03*i02_max + i02;
+    const int64_t src1_col_stride = split && g_device_count > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
+    for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
+        const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0;
+        const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
 
-                // i0 values that contain the lower/upper rows for a split tensor when using multiple GPUs
-                const int64_t i0_offset_low = row_low/rows_per_iter;
-                const int64_t i0_offset_high = row_high/rows_per_iter;
+        for (int64_t id = 0; id < g_device_count; ++id) {
+            if ((!split && id != g_main_device) || row_low[id] == row_high[id]) {
+                continue;
+            }
 
-                int64_t i01_low = 0;
-                int64_t i01_high = rows_per_iter;
-                if (split) {
-                    if (i0 < i0_offset_low || i0 > i0_offset_high) {
-                        continue;
-                    }
-                    if (i0 == i0_offset_low) {
-                        i01_low = row_low % rows_per_iter;
-                    }
-                    if (i0 == i0_offset_high) {
-                        i01_high = row_high % rows_per_iter;
-                    }
-                }
+            const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
+            const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
+            const int64_t row_diff = row_high[id] - row_low[id];
 
-                // There is possibly a bug in the Windows nvcc compiler regarding instruction reordering or optimizing out local variables.
-                // Removing the first assert or changing the order of the arguments causes the second assert to fail.
-                // Removing both asserts results in i01_high becoming 0 which in turn results in garbage output.
-                // The root cause seems to be a problem with i0_offset_high becoming 0 when it should always be >0 (for single GPU).
-                GGML_ASSERT(i01_low == 0 || g_device_count > 1);
-                GGML_ASSERT(i01_high == rows_per_iter || g_device_count > 1);
+            ggml_cuda_set_device(id);
+            const cudaStream_t stream = g_cudaStreams[id][is];
 
-                const int64_t i01_diff = i01_high - i01_low;
-                if (i01_diff == 0) {
-                    continue;
-                }
-                const int64_t i11 = i13*ne12 + i12;
+            // wait for main GPU data if necessary
+            if (split && (id != g_main_device || is != 0)) {
+                CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[g_main_device][0], 0));
+            }
+
+            for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
+                const int64_t i03 = i0 / ne12;
+                const int64_t i02 = i0 % ne12;
+
+                const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
 
                 // for split tensors the data begins at i0 == i0_offset_low
-                char  * src0_ddq_i = src0_ddq[id] + (i0/i02_divisor - i0_offset_low)*src0_stride*src0_ts/src0_bs;
-                float * src0_ddf_i = src0_ddf[id] + (i0/i02_divisor - i0_offset_low)*src0_stride;
-                float * src1_ddf_i = src1_ddf[id] + i11*src1_stride;
-                float * dst_ddf_i  =  dst_ddf[id] + (i0             - i0_offset_low)*dst_stride;
-
-                // for split tensors the data pointer needs to be rounded down
-                // to the bin edge for i03, i02 bins beyond the first
-                if (i0 - i0_offset_low > 0) {
-                    GGML_ASSERT(!flatten_rows);
-                    src0_ddq_i -= (row_low % ne01)*ne00 * src0_ts/src0_bs;
-                    src0_ddf_i -= (row_low % ne01)*ne00;
-                    dst_ddf_i  -= (row_low % ne0)*ne1;
-                }
+                char  *  src0_dd_i =  src0_dd[id] + (i0/i02_divisor) * ne01*ne00*src0_ts/src0_bs;
+                float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10;
+                char  * src1_ddq_i = src1_ddq[id] +  src1_ddq_i_offset;
+                float *   dst_dd_i =   dst_dd[id] + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);
 
                 // the main device memory buffer can be on VRAM scratch, with space for all partial results
                 // in that case an offset on dst_ddf_i is needed
                 if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) {
-                    dst_ddf_i += i01_low; // offset is 0 if no tensor split
+                    dst_dd_i += row_low[id]; // offset is 0 if no tensor split
                 }
 
                 // copy src0, src1 to device if necessary
-                if (use_src1 && !src1_stays_on_host) {
-                    if (src1->backend == GGML_BACKEND_CPU) {
-                        GGML_ASSERT(!flatten_rows || nrows0 == ggml_nrows(src1));
-                        int64_t nrows1 = flatten_rows ? nrows0 : ne11;
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, nrows1, cudaStream_main));
-                    } else if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
-                        if (id != g_main_device) {
-                            GGML_ASSERT(!flatten_rows);
+                if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
+                    if (id != g_main_device) {
+                        if (convert_src1_to_q8_1) {
+                            char * src1_ddq_i_source = src1_ddq[g_main_device] + src1_ddq_i_offset;
+                            CUDA_CHECK(cudaMemcpyAsync(src1_ddq_i, src1_ddq_i_source, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs,
+                                                    cudaMemcpyDeviceToDevice, stream));
+                        } else {
                             float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device];
-                            src1_ddf_i_source += i11*src1_stride;
-                            CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_stride*sizeof(float),
-                                                    cudaMemcpyDeviceToDevice, cudaStream_main));
+                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
+                            CUDA_CHECK(cudaMemcpyAsync(src1_ddf_i, src1_ddf_i_source, src1_ncols*ne10*sizeof(float),
+                                                    cudaMemcpyDeviceToDevice, stream));
                         }
-                    } else if (src1_on_device && !src1_is_contiguous) {
-                        GGML_ASSERT(!split);
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf_i, src1, i03, i02, 0, ne11, cudaStream_main));
-                    } else {
-                        GGML_ASSERT(false);
                     }
+                } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) {
+                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
+                                   src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
+                } else {
+                    GGML_ASSERT(false);
                 }
 
-                if ((!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
-                    if (src0_is_f32) {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
-                    } else {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
-                    }
-                }
-
-                // convert src0 to f32 if it is necessary for the ggml_cuda_op
-                if (src0_needs_f32 && !src0_is_f32) {
-                    to_fp32_cuda(src0_ddq_i, src0_ddf_i, i01_diff*ne00, cudaStream_main);
+                if (convert_src1_to_q8_1 && src1->backend == GGML_BACKEND_CPU) {
+                    quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
 
+                if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
+                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, row_low[id], row_high[id], stream));
+                }
+
                 // do the computation
-                op(src0, src1, dst, src0_ddq_i, src0_ddf_i, src1_ddf_i, dst_ddf_i, i02, i01_low, i01_high, i11, cudaStream_main);
+                op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
+                   row_low[id], row_high[id], src1_ncols, src1_padded_col_size, stream);
                 CUDA_CHECK(cudaGetLastError());
 
                 // copy dst to host or other device if necessary
@@ -5910,95 +6700,86 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                         // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
                         // Instead they need to be copied to the correct slice in ne0 = dst row index.
                         // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
-                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i01_low*sizeof(float) + i02*nb2 + i03*nb3);
-                        CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_ddf_i, i01_diff*sizeof(float),
-                                                     i01_diff*sizeof(float), ne1, kind, cudaStream_main));
+                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0 + row_low[id];
+                        CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float), dst_dd_i, row_diff*sizeof(float),
+                                                    row_diff*sizeof(float), src1_ncols, kind, stream));
                     } else {
                         float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
-                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_ddf_i, dst_stride*sizeof(float), kind, cudaStream_main));
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0;
+                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), kind, stream));
                     }
                 }
 
-                // signify to main device that other device is done
-                if (split && g_device_count > 1 && id != g_main_device) {
-                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id], cudaStream_main));
+                // add event for the main device to wait on until other device is done
+                if (split && (id != g_main_device || is != 0)) {
+                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
                 }
             }
         }
     }
 
-    // wait until each device is finished, then free their buffers
-    for (int id = 0; id < g_device_count; ++id) {
-        if (src0_asq[id] == 0 && src0_asf[id] == 0 && src1_asf[id] == 0 && dst_asf[id] == 0) {
-            continue;
-        }
+    for (int64_t id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(ggml_cuda_set_device(id));
 
-        CUDA_CHECK(cudaSetDevice(id));
-
-        if (src0_asq[id] > 0) {
-            ggml_cuda_pool_free(src0_ddq[id], src0_asq[id]);
-        }
-        if (src0_asf[id] > 0) {
-            ggml_cuda_pool_free(src0_ddf[id], src0_asf[id]);
+        // free buffers again when done
+        if (src0_as[id] > 0) {
+            ggml_cuda_pool_free(src0_dd[id], src0_as[id]);
         }
         if (src1_asf[id] > 0) {
             ggml_cuda_pool_free(src1_ddf[id], src1_asf[id]);
         }
-        if (dst_asf[id] > 0) {
-            ggml_cuda_pool_free(dst_ddf[id], dst_asf[id]);
+        if (src1_asq[id] > 0) {
+            ggml_cuda_pool_free(src1_ddq[id], src1_asq[id]);
+        }
+        if (dst_as[id] > 0) {
+            ggml_cuda_pool_free(dst_dd[id], dst_as[id]);
         }
     }
 
     // main device waits for all other devices to be finished
     if (split && g_device_count > 1) {
-        CUDA_CHECK(cudaSetDevice(g_main_device));
-        for (int id = 0; id < g_device_count; ++id) {
-            if (id != g_main_device && src0_extra->events[id]) {
-                CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams_main[g_main_device], src0_extra->events[id]));
+        int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
+        is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS;
+
+        CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+        for (int64_t id = 0; id < g_device_count; ++id) {
+            for (int64_t is = 0; is < is_max; ++is) {
+                CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
             }
         }
     }
 
     if (dst->backend == GGML_BACKEND_CPU) {
-        CUDA_CHECK(cudaSetDevice(g_main_device));
+        CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         CUDA_CHECK(cudaDeviceSynchronize());
     }
 }
 
-void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    // ggml_cuda_add permits f16 dst even though this could in theory cause problems with the pointer arithmetic in ggml_cuda_op.
-    // Due to flatten_rows == true this does in practice not make a difference however.
-    // Better solution would be nice but right now that would require disproportionate changes.
-    GGML_ASSERT(
-        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) &&
-        src1->type == GGML_TYPE_F32 &&
-        (dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16));
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_add, false, true);
+static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
 }
 
-void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul, true, false); // TODO ggml_cuda_op needs modification for flatten
+static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul);
 }
 
-void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_gelu, true, true);
+static void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu);
 }
 
-void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_silu, true, true);
+static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu);
 }
 
-void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_norm, true, true);
+static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm);
 }
 
-void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rms_norm, true, true);
+static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
 }
 
 bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
@@ -6008,17 +6789,13 @@ bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_te
     const int64_t ne1 = dst->ne[1];
 
     // TODO: find the optimal values for these
-    if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
-        src1->type == GGML_TYPE_F32 &&
-        dst->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
-        return true;
-    }
-
-    return false;
+    return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
+            src1->type == GGML_TYPE_F32 &&
+             dst->type == GGML_TYPE_F32 &&
+            (ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
 }
 
-void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
+static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
     GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
@@ -6032,8 +6809,8 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
 
     const int64_t ne12 = src1->ne[2];
 
-    CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
     struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
@@ -6044,10 +6821,10 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, cudaStream_main);
+    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
 }
 
-void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
+static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
     GGML_ASSERT(!ggml_is_contiguous(src0) && ggml_is_contiguous(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
     GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
@@ -6063,8 +6840,8 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int64_t nb01 = src0->nb[1];
     const int64_t nb02 = src0->nb[2];
 
-    CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
     struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     void * src0_ddq = src0_extra->data_device[g_main_device];
@@ -6075,38 +6852,49 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
-    const int row_stride_x = nb01 / sizeof(half);
-    const int channel_stride_x = nb02 / sizeof(half);
+    const int64_t row_stride_x = nb01 / sizeof(half);
+    const int64_t channel_stride_x = nb02 / sizeof(half);
 
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, cudaStream_main);
+    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
 }
 
-void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     bool all_on_device = (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
         src1->backend == GGML_BACKEND_GPU && dst->backend == GGML_BACKEND_GPU;
 
+    int64_t min_compute_capability = INT_MAX;
+    for (int64_t id = 0; id < g_device_count; ++id) {
+        if (min_compute_capability > g_compute_capabilities[id]
+                && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
+            min_compute_capability = g_compute_capabilities[id];
+        }
+    }
+
     if (all_on_device && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
     } else if (all_on_device && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && src1->ne[1] == 1) {
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
     }else if (src0->type == GGML_TYPE_F32) {
-        ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
         if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0) {
-            ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_vec, false, false);
-        } else {
-            int min_compute_capability = INT_MAX;
-            for (int id = 0; id < g_device_count; ++id) {
-                if (min_compute_capability > g_compute_capabilities[id]
-                        && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
-                    min_compute_capability = g_compute_capabilities[id];
-                }
-            }
 
-            if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) {
-                ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_q, false, false);
+#ifdef GGML_CUDA_FORCE_DMMV
+            const bool use_mul_mat_vec_q = false;
+#else
+            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+#endif // GGML_CUDA_FORCE_DMMV
+
+            if (use_mul_mat_vec_q) {
+                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
             } else {
-                ggml_cuda_op(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, true, false);
+                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
+            }
+        } else {
+            if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) {
+                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
+            } else {
+                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
             }
         }
     } else {
@@ -6114,12 +6902,11 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_
     }
 }
 
-void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_scale, true, true);
+static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
 }
 
-void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -6145,8 +6932,8 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
     const int64_t nb11 = src1->nb[1];
     const int64_t nb12 = src1->nb[2];
 
-    CUDA_CHECK(cudaSetDevice(g_main_device));
-    cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
 
     const struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     const struct ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
@@ -6156,53 +6943,49 @@ void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tens
 
     if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_f32_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
-                              ne10, ne11, nb10, nb11, nb12, cudaStream_main);
+                              ne10, ne11, nb10, nb11, nb12, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02,
-                              ne10, ne11, nb10, nb11, nb12, cudaStream_main);
+                              ne10, ne11, nb10, nb11, nb12, main_stream);
     } else {
+        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
         GGML_ASSERT(false);
     }
 
     (void) dst;
 }
 
-void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_cpy(src0, dst, nullptr);
     (void) src1;
 }
 
-void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_diag_mask_inf, true, true);
+static void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_diag_mask_inf);
 }
 
-void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_soft_max, true, true);
+static void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_soft_max);
 }
 
-void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-
-    const int mode = ((int32_t *) dst->op_params)[2];
-    const bool is_glm = mode & 4;
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm
+static void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rope);
 }
 
-void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_alibi, true, true);
+static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi);
 }
 
-void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     (void) src0;
     (void) src1;
     (void) dst;
 }
 
 void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
-    int nrows = ggml_nrows(tensor);
+    const int64_t nrows = ggml_nrows(tensor);
 
     const int64_t ne0 = tensor->ne[0];
 
@@ -6212,14 +6995,14 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
     struct ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu;
     memset(extra, 0, sizeof(*extra));
 
-    for (int id = 0; id < g_device_count; ++id) {
+    for (int64_t id = 0; id < g_device_count; ++id) {
         if (backend == GGML_BACKEND_GPU && id != g_main_device) {
             continue;
         }
 
-        cudaSetDevice(id);
+        ggml_cuda_set_device(id);
 
-        int row_low, row_high;
+        int64_t row_low, row_high;
         if (backend == GGML_BACKEND_GPU) {
             row_low = 0;
             row_high = nrows;
@@ -6269,7 +7052,9 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         extra->data_device[id] = buf;
 
         if (backend == GGML_BACKEND_GPU_SPLIT) {
-            CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id], cudaEventDisableTiming));
+            for (int64_t is = 0; is < MAX_STREAMS; ++is) {
+                CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
+            }
         }
     }
 
@@ -6283,15 +7068,17 @@ void ggml_cuda_free_data(struct ggml_tensor * tensor) {
 
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
 
-    for (int id = 0; id < g_device_count; ++id) {
+    for (int64_t id = 0; id < g_device_count; ++id) {
         if (extra->data_device[id] != nullptr) {
-            CUDA_CHECK(cudaSetDevice(id));
+            CUDA_CHECK(ggml_cuda_set_device(id));
             CUDA_CHECK(cudaFree(extra->data_device[id]));
         }
 
-        if (extra->events[id] != nullptr) {
-            CUDA_CHECK(cudaSetDevice(id));
-            CUDA_CHECK(cudaEventDestroy(extra->events[id]));
+        for (int64_t is = 0; is < MAX_STREAMS; ++is) {
+            if (extra->events[id][is] != nullptr) {
+                CUDA_CHECK(ggml_cuda_set_device(id));
+                CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
+            }
         }
     }
 
@@ -6314,11 +7101,13 @@ static struct ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
     return extra;
 }
 
-void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace, bool no_alloc) {
+static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace, bool no_alloc) {
     if (scratch && g_scratch_size == 0) {
         return;
     }
 
+    tensor->backend = GGML_BACKEND_GPU;
+
     // recursively assign CUDA buffers until a compute tensor is found
     if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
         const ggml_op src0_op = tensor->src[0]->op;
@@ -6330,8 +7119,6 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
         ggml_cuda_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc);
     }
 
-    tensor->backend = GGML_BACKEND_GPU;
-
     if (scratch && no_alloc) {
         return;
     }
@@ -6343,7 +7130,7 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
         force_inplace;
     const size_t size = ggml_nbytes(tensor);
 
-    CUDA_CHECK(cudaSetDevice(g_main_device));
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
     if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
         struct ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
@@ -6392,6 +7179,7 @@ void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset)
         return;
     }
     if (g_scratch_buffer == nullptr) {
+        ggml_cuda_set_device(g_main_device);
         CUDA_CHECK(cudaMalloc(&g_scratch_buffer, g_scratch_size));
     }
 
@@ -6415,6 +7203,15 @@ void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset)
     tensor->extra = extra;
 }
 
+void ggml_cuda_copy_to_device(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(ggml_is_contiguous(tensor));
+
+    struct ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
+    CUDA_CHECK(ggml_cuda_set_device(g_main_device));
+    CUDA_CHECK(cudaMemcpy(extra->data_device[g_main_device], tensor->data, ggml_nbytes(tensor), cudaMemcpyHostToDevice));
+}
+
 void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
     ggml_cuda_assign_buffers_impl(tensor, true, false, false);
 }
@@ -6431,7 +7228,7 @@ void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor) {
     ggml_cuda_assign_buffers_impl(tensor, false, true, false);
 }
 
-void ggml_cuda_set_main_device(int main_device) {
+void ggml_cuda_set_main_device(const int main_device) {
     if (main_device >= g_device_count) {
         fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n",
                 main_device, g_device_count, g_main_device);
@@ -6445,12 +7242,17 @@ void ggml_cuda_set_main_device(int main_device) {
     }
 }
 
-void ggml_cuda_set_mul_mat_q(bool mul_mat_q) {
+void ggml_cuda_set_mul_mat_q(const bool mul_mat_q) {
     g_mul_mat_q = mul_mat_q;
 }
 
-void ggml_cuda_set_scratch_size(size_t scratch_size) {
-    g_scratch_size = scratch_size;
+void ggml_cuda_set_scratch_size(const size_t scratch_size) {
+    // this is a hack to not completely break llama.cpp when using multiple models or contexts simultaneously
+    // it still won't always work as expected, but it's better than nothing
+    if (scratch_size > g_scratch_size) {
+        ggml_cuda_free_scratch();
+    }
+    g_scratch_size = std::max(g_scratch_size, scratch_size);
 }
 
 void ggml_cuda_free_scratch() {
diff --git a/ggml-cuda.h b/ggml-cuda.h
index f66bb1678..fda704b66 100644
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@@ -2,6 +2,14 @@
 
 #include "ggml.h"
 
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -23,6 +31,7 @@ GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tens
 
 GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
 GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
+GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
 
 GGML_API void   ggml_cuda_set_main_device(int main_device);
 GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
diff --git a/ggml-metal.h b/ggml-metal.h
index 00202b787..790cf0bf7 100644
--- a/ggml-metal.h
+++ b/ggml-metal.h
@@ -19,11 +19,14 @@
 
 #pragma once
 
+#include "ggml.h"
+
 #include <stddef.h>
 #include <stdbool.h>
 
 // max memory buffers that can be mapped to the device
 #define GGML_METAL_MAX_BUFFERS 16
+#define GGML_METAL_MAX_COMMAND_BUFFERS 32
 
 struct ggml_tensor;
 struct ggml_cgraph;
@@ -32,6 +35,8 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 struct ggml_metal_context;
 
 // number of command buffers to use
diff --git a/ggml-metal.m b/ggml-metal.m
index 835c5f297..92956ed97 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -12,9 +12,13 @@
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
 #ifdef GGML_METAL_NDEBUG
-#define metal_printf(...)
+#define GGML_METAL_LOG_INFO(...)
+#define GGML_METAL_LOG_WARN(...)
+#define GGML_METAL_LOG_ERROR(...)
 #else
-#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 #endif
 
 #define UNUSED(x) (void)(x)
@@ -33,12 +37,15 @@ struct ggml_metal_buffer {
 struct ggml_metal_context {
     int n_cb;
 
-    float * logits;
-
     id<MTLDevice>       device;
     id<MTLCommandQueue> queue;
     id<MTLLibrary>      library;
 
+    id<MTLCommandBuffer>         command_buffers [GGML_METAL_MAX_COMMAND_BUFFERS];
+    id<MTLComputeCommandEncoder> command_encoders[GGML_METAL_MAX_COMMAND_BUFFERS];
+
+    dispatch_queue_t d_queue;
+
     int n_buffers;
     struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
 
@@ -59,10 +66,14 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(relu);
     GGML_METAL_DECL_KERNEL(gelu);
     GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(soft_max_4);
     GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DECL_KERNEL(get_rows_f32);
     GGML_METAL_DECL_KERNEL(get_rows_f16);
     GGML_METAL_DECL_KERNEL(get_rows_q4_0);
     GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q8_0);
     GGML_METAL_DECL_KERNEL(get_rows_q2_K);
     GGML_METAL_DECL_KERNEL(get_rows_q3_K);
     GGML_METAL_DECL_KERNEL(get_rows_q4_K);
@@ -70,27 +81,36 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(norm);
-    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f32_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row);
+    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(rope_f32);
+    GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
     GGML_METAL_DECL_KERNEL(cpy_f32_f16);
     GGML_METAL_DECL_KERNEL(cpy_f32_f32);
     GGML_METAL_DECL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DECL_KERNEL(concat);
+    GGML_METAL_DECL_KERNEL(sqr);
 
 #undef GGML_METAL_DECL_KERNEL
 };
@@ -106,60 +126,103 @@ static NSString * const msl_library_source = @"see metal.metal";
 @implementation GGMLMetalClass
 @end
 
+ggml_log_callback ggml_metal_log_callback = NULL;
+void * ggml_metal_log_user_data = NULL;
+
+void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_metal_log_callback  = log_callback;
+    ggml_metal_log_user_data = user_data;
+}
+
+static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){
+    if (ggml_metal_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
+        } else {
+            char* buffer2 = malloc(len+1);
+            vsnprintf(buffer2, len+1, format, args);
+            buffer2[len] = 0;
+            ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
+            free(buffer2);
+        }
+        va_end(args);
+    }
+}
+
+
+
 struct ggml_metal_context * ggml_metal_init(int n_cb) {
-    fprintf(stderr, "%s: allocating\n", __func__);
+    GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
 
+    id <MTLDevice> device;
+    NSString * s;
+
+#if TARGET_OS_OSX
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
+    for (device in devices) {
+        s = [device name];
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
+    }
+#endif
+
+    // Pick and show default Metal device
+    device = MTLCreateSystemDefaultDevice();
+    s = [device name];
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
+
+    // Configure context
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
-
-    ctx->n_cb   = n_cb;
-    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->device = device;
+    ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
     ctx->n_buffers = 0;
     ctx->concur_list_len = 0;
 
+    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-#if 0
-    // compile from source string and show compile log
+    // load library
     {
-        NSError * error = nil;
-
-        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
-        if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            return NULL;
-        }
-    }
+        NSBundle * bundle = nil;
+#ifdef SWIFT_PACKAGE
+        bundle = SWIFTPM_MODULE_BUNDLE;
 #else
-    UNUSED(msl_library_source);
-
-    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
-    {
+        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
         NSError * error = nil;
+        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (libPath != nil) {
+            NSURL * libURL = [NSURL fileURLWithPath:libPath];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
+            ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+        } else {
+            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
-        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
-
-        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
-        if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
-            return NULL;
-        }
+            NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
+            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
+            if (error) {
+                GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                return NULL;
+            }
 
+            MTLCompileOptions* options = nil;
 #ifdef GGML_QKK_64
-        MTLCompileOptions* options = [MTLCompileOptions new];
-        options.preprocessorMacros = @{ @"QK_K" : @(64) };
-        ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
-#else
-        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+            options = [MTLCompileOptions new];
+            options.preprocessorMacros = @{ @"QK_K" : @(64) };
 #endif
+            ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+        }
+
         if (error) {
-            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
             return NULL;
         }
     }
-#endif
 
     // load kernels
     {
@@ -167,9 +230,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(name) \
         ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
         ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
+        GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
+                (int) ctx->pipeline_##name.threadExecutionWidth); \
         if (error) { \
-            fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+          GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
             return NULL; \
         }
 
@@ -182,10 +247,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(relu);
         GGML_METAL_ADD_KERNEL(gelu);
         GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(soft_max_4);
         GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf_8);
+        GGML_METAL_ADD_KERNEL(get_rows_f32);
         GGML_METAL_ADD_KERNEL(get_rows_f16);
         GGML_METAL_ADD_KERNEL(get_rows_q4_0);
         GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q8_0);
         GGML_METAL_ADD_KERNEL(get_rows_q2_K);
         GGML_METAL_ADD_KERNEL(get_rows_q3_K);
         GGML_METAL_ADD_KERNEL(get_rows_q4_K);
@@ -193,55 +262,151 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(norm);
-        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(mul_mv_f32_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row);
+        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
+        if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
+            GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+        }
+        GGML_METAL_ADD_KERNEL(rope_f32);
+        GGML_METAL_ADD_KERNEL(rope_f16);
         GGML_METAL_ADD_KERNEL(alibi_f32);
         GGML_METAL_ADD_KERNEL(cpy_f32_f16);
         GGML_METAL_ADD_KERNEL(cpy_f32_f32);
         GGML_METAL_ADD_KERNEL(cpy_f16_f16);
+        GGML_METAL_ADD_KERNEL(concat);
+        GGML_METAL_ADD_KERNEL(sqr);
 
 #undef GGML_METAL_ADD_KERNEL
     }
 
-    fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-    fprintf(stderr, "%s: hasUnifiedMemory             = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
-    if (ctx->device.maxTransferRate != 0) {
-        fprintf(stderr, "%s: maxTransferRate              = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
-    } else {
-        fprintf(stderr, "%s: maxTransferRate              = built-in GPU\n", __func__);
+#if TARGET_OS_OSX
+    // print MTL GPU family:
+    GGML_METAL_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
+
+    // determine max supported GPU family
+    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+    for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
+        if ([ctx->device supportsFamily:i]) {
+            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - MTLGPUFamilyApple1 + 1, i);
+            break;
+        }
     }
 
+    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    if (ctx->device.maxTransferRate != 0) {
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+    } else {
+        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
+    }
+#endif
+
     return ctx;
 }
 
 void ggml_metal_free(struct ggml_metal_context * ctx) {
-    fprintf(stderr, "%s: deallocating\n", __func__);
+    GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
+#define GGML_METAL_DEL_KERNEL(name) \
+    [ctx->function_##name release]; \
+    [ctx->pipeline_##name release];
+
+    GGML_METAL_DEL_KERNEL(add);
+    GGML_METAL_DEL_KERNEL(add_row);
+    GGML_METAL_DEL_KERNEL(mul);
+    GGML_METAL_DEL_KERNEL(mul_row);
+    GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(silu);
+    GGML_METAL_DEL_KERNEL(relu);
+    GGML_METAL_DEL_KERNEL(gelu);
+    GGML_METAL_DEL_KERNEL(soft_max);
+    GGML_METAL_DEL_KERNEL(soft_max_4);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf_8);
+    GGML_METAL_DEL_KERNEL(get_rows_f32);
+    GGML_METAL_DEL_KERNEL(get_rows_f16);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DEL_KERNEL(get_rows_q8_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q2_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q3_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DEL_KERNEL(rms_norm);
+    GGML_METAL_DEL_KERNEL(norm);
+    GGML_METAL_DEL_KERNEL(mul_mv_f32_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row);
+    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
+    if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
+        GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+    }
+    GGML_METAL_DEL_KERNEL(rope_f32);
+    GGML_METAL_DEL_KERNEL(rope_f16);
+    GGML_METAL_DEL_KERNEL(alibi_f32);
+    GGML_METAL_DEL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DEL_KERNEL(cpy_f32_f32);
+    GGML_METAL_DEL_KERNEL(cpy_f16_f16);
+    GGML_METAL_DEL_KERNEL(concat);
+    GGML_METAL_DEL_KERNEL(sqr);
+
+#undef GGML_METAL_DEL_KERNEL
+
     for (int i = 0; i < ctx->n_buffers; ++i) {
         [ctx->buffers[i].metal release];
     }
+
+    [ctx->library release];
+    [ctx->queue release];
+    [ctx->device release];
+
+    dispatch_release(ctx->d_queue);
+
     free(ctx);
 }
 
 void * ggml_metal_host_malloc(size_t n) {
     void * data = NULL;
-    const int result = posix_memalign((void **) &data, getpagesize(), n);
+    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
     if (result != 0) {
-        fprintf(stderr, "%s: error: posix_memalign failed\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
         return NULL;
     }
 
@@ -253,7 +418,7 @@ void ggml_metal_host_free(void * data) {
 }
 
 void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
-    ctx->n_cb = n_cb;
+    ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
 }
 
 int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
@@ -269,7 +434,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
 // Metal buffer based on the host memory pointer
 //
 static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
-    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+    //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
 
     const int64_t tsize = ggml_nbytes(t);
 
@@ -277,16 +442,17 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
     for (int i = 0; i < ctx->n_buffers; ++i) {
         const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
 
+        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
         if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
-            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
 
             return ctx->buffers[i].metal;
         }
     }
 
-    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
 
     return nil;
 }
@@ -298,7 +464,7 @@ bool ggml_metal_add_buffer(
                          size_t   size,
                          size_t   max_size) {
     if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
-        fprintf(stderr, "%s: too many buffers\n", __func__);
+        GGML_METAL_LOG_ERROR("%s: error: too many buffers\n", __func__);
         return false;
     }
 
@@ -308,12 +474,12 @@ bool ggml_metal_add_buffer(
             const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
 
             if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                GGML_METAL_LOG_ERROR("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
                 return false;
             }
         }
 
-        const size_t size_page = getpagesize();
+        const size_t size_page = sysconf(_SC_PAGESIZE);
 
         size_t size_aligned = size;
         if ((size_aligned % size_page) != 0) {
@@ -329,11 +495,11 @@ bool ggml_metal_add_buffer(
             ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
                 return false;
             }
 
-            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
 
             ++ctx->n_buffers;
         } else {
@@ -353,28 +519,32 @@ bool ggml_metal_add_buffer(
                 ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
                     return false;
                 }
 
-                fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
                 if (i + size_step < size) {
-                    fprintf(stderr, "\n");
+                    GGML_METAL_LOG_INFO("\n");
                 }
 
                 ++ctx->n_buffers;
             }
         }
 
-        fprintf(stderr, ", (%8.2f / %8.2f)",
+#if TARGET_OS_OSX
+        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
                 ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
                 ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
         if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
+            GGML_METAL_LOG_WARN(", warning: current allocated size is greater than the recommended max working set size\n", __func__);
         } else {
-            fprintf(stderr, "\n");
+            GGML_METAL_LOG_INFO("\n");
         }
+#else
+        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
+#endif
     }
 
     return true;
@@ -383,8 +553,6 @@ bool ggml_metal_add_buffer(
 void ggml_metal_set_tensor(
         struct ggml_metal_context * ctx,
         struct ggml_tensor * t) {
-    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
-
     size_t offs;
     id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
 
@@ -394,8 +562,6 @@ void ggml_metal_set_tensor(
 void ggml_metal_get_tensor(
         struct ggml_metal_context * ctx,
         struct ggml_tensor * t) {
-    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
-
     size_t offs;
     id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
 
@@ -490,14 +656,14 @@ void ggml_metal_graph_find_concurrency(
     }
 
     if (ctx->concur_list_len > GGML_MAX_CONCUR) {
-        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+        GGML_METAL_LOG_WARN("%s: too many elements for metal ctx->concur_list!\n", __func__);
     }
 }
 
 void ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
-    metal_printf("%s: evaluating graph\n", __func__);
+    @autoreleasepool {
 
     // if there is ctx->concur_list, dispatch concurrently
     // else fallback to serial dispatch
@@ -513,32 +679,28 @@ void ggml_metal_graph_compute(
 
     const int n_cb = ctx->n_cb;
 
-    NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb];
-
     for (int i = 0; i < n_cb; ++i) {
-        command_buffers[i] = [ctx->queue commandBuffer];
+        ctx->command_buffers[i] = [ctx->queue commandBuffer];
 
         // enqueue the command buffers in order to specify their execution order
-        [command_buffers[i] enqueue];
-    }
+        [ctx->command_buffers[i] enqueue];
 
-    // TODO: is this the best way to start threads?
-    dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+        ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc];
+    }
 
     for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
         const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
 
-        dispatch_async(queue, ^{
+        dispatch_async(ctx->d_queue, ^{
             size_t offs_src0 = 0;
             size_t offs_src1 = 0;
             size_t offs_dst  = 0;
 
-            id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
+            id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
+            id<MTLComputeCommandEncoder> encoder = ctx->command_encoders[cb_idx];
 
-            id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
-
-            const int node_start =                                  (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
 
             for (int ind = node_start; ind < node_end; ++ind) {
                 const int i = has_concur ? ctx->concur_list[ind] : ind;
@@ -548,7 +710,7 @@ void ggml_metal_graph_compute(
                     continue;
                 }
 
-                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+                //GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
 
                 struct ggml_tensor * src0 = gf->nodes[i]->src[0];
                 struct ggml_tensor * src1 = gf->nodes[i]->src[1];
@@ -592,17 +754,17 @@ void ggml_metal_graph_compute(
                 id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
                 id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
 
-                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
                 //if (src0) {
-                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //    GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
                 //            ggml_is_contiguous(src0), src0->name);
                 //}
                 //if (src1) {
-                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //    GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
                 //            ggml_is_contiguous(src1), src1->name);
                 //}
                 //if (dst) {
-                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //    GGML_METAL_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
                 //            dst->name);
                 //}
 
@@ -615,11 +777,60 @@ void ggml_metal_graph_compute(
                         {
                             // noop
                         } break;
+                    case GGML_OP_CONCAT:
+                        {
+
+                            int64_t nb = ne00;
+                            [encoder setComputePipelineState:ctx->pipeline_concat];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
+
+                            const int nth = MIN(1024, ne0);
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
                     case GGML_OP_ADD:
                         {
-                            if (ggml_nelements(src1) == ne10) {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
+
+                            bool bcast_row = false;
+
+                            int64_t nb = ne00;
+
+                            if (ggml_nelements(src1) == ne10 && ne00 % 4 == 0) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
+
+                                nb = ne00 / 4;
                                 [encoder setComputePipelineState:ctx->pipeline_add_row];
+
+                                bcast_row = true;
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_add];
                             }
@@ -627,15 +838,53 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
 
-                            const int64_t n = ggml_nelements(dst);
+                            if (bcast_row) {
+                                const int64_t n = ggml_nelements(dst)/4;
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } else {
+                                const int nth = MIN(1024, ne0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL:
                         {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+                            GGML_ASSERT(ggml_is_contiguous(src1));
+
+                            // utilize float4
+                            GGML_ASSERT(ne00 % 4 == 0);
+                            const int64_t nb = ne00/4;
+
                             if (ggml_nelements(src1) == ne10) {
                                 // src1 is a row
+                                GGML_ASSERT(ne11 == 1);
                                 [encoder setComputePipelineState:ctx->pipeline_mul_row];
                             } else {
                                 [encoder setComputePipelineState:ctx->pipeline_mul];
@@ -643,14 +892,16 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&nb     length:sizeof(nb) atIndex:3];
 
-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
                     case GGML_OP_SCALE:
                         {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
                             const float scale = *(const float *) src1->data;
 
                             [encoder setComputePipelineState:ctx->pipeline_scale];
@@ -658,7 +909,7 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
 
-                            const int64_t n = ggml_nelements(dst);
+                            const int64_t n = ggml_nelements(dst)/4;
 
                             [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
@@ -670,7 +921,7 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
@@ -690,27 +941,41 @@ void ggml_metal_graph_compute(
                                     [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                     [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                    const int64_t n = ggml_nelements(dst);
+                                    const int64_t n = ggml_nelements(dst)/4;
 
                                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                                 } break;
                             default:
                                 {
-                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                                     GGML_ASSERT(false);
                                 }
                         } break;
+                    case GGML_OP_SQR:
+                        {
+                            GGML_ASSERT(ggml_is_contiguous(src0));
+
+                            [encoder setComputePipelineState:ctx->pipeline_sqr];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                            const int64_t n = ggml_nelements(dst);
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
                     case GGML_OP_SOFT_MAX:
                         {
-                            const int nth = 32;
+                            const int nth = MIN(32, ne00);
 
-                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            if (ne00%4 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
@@ -718,68 +983,121 @@ void ggml_metal_graph_compute(
                         {
                             const int n_past = ((int32_t *)(dst->op_params))[0];
 
-                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            if (ne00%8 == 0) {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            }
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
                             [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
                             [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
                             [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            if (ne00%8 == 0) {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
+                            else {
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            }
                         } break;
                     case GGML_OP_MUL_MAT:
                         {
-                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
-
                             GGML_ASSERT(ne00 == ne10);
-                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
-                            uint gqa = ne12/ne02;
                             GGML_ASSERT(ne03 == ne13);
 
+                            const uint gqa = ne12/ne02;
+
+                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+                            // to the matrix-vector kernel
+                            int ne11_mm_min = 1;
+
+#if 0
+                            // the numbers below are measured on M2 Ultra for 7B and 13B models
+                            // these numbers do not translate to other devices or model sizes
+                            // TODO: need to find a better approach
+                            if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
+                                    case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
+                                    case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q4_0:
+                                    case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
+                                    case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
+                                    case GGML_TYPE_Q5_0:                          // not tested yet
+                                    case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
+                                    case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
+                                    case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
+                                    default:             ne11_mm_min = 1;  break;
+                                }
+                            }
+#endif
+
                             // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                             // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            if (ggml_is_contiguous(src0) &&
-                                ggml_is_contiguous(src1) &&
+                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                !ggml_is_transposed(src0) &&
+                                !ggml_is_transposed(src1) &&
                                 src1t == GGML_TYPE_F32 &&
-                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                                ne00%32 == 0 &&
-                                ne11 > 1) {
-                                    switch (src0->type) {
-                                        case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
-                                        case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
-                                        case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
-                                        case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
-                                        case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
-                                        case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
-                                        case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
-                                        case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
-                                        default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
-                                    }
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
-                                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
-                                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
-                                    [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
-                                    [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
-                                    [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
-                                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                                ne00 % 32 == 0 &&
+                                ne11 > ne11_mm_min) {
+                                //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+                                switch (src0->type) {
+                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
+                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
+                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
+                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
+                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
+                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
+                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
+                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                 }
-                            else {
+                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
+                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
+                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:13];
+                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            } else {
                                 int nth0 = 32;
                                 int nth1 = 1;
+                                int nrows = 1;
+                                //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
                                 // use custom matrix x vector kernel
                                 switch (src0t) {
+                                    case GGML_TYPE_F32:
+                                        {
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32];
+                                            nrows = 4;
+                                        } break;
                                     case GGML_TYPE_F16:
                                         {
-                                            nth0 = 64;
+                                            nth0 = 32;
                                             nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                            if (ne11 * ne12 < 4) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row];
+                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4];
+                                                nrows = ne11;
+                                            } else {
+                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32];
+                                                nrows = 4;
+                                            }
                                         } break;
                                     case GGML_TYPE_Q4_0:
                                         {
@@ -788,7 +1106,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 8;
                                             nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_0_f32];
                                         } break;
                                     case GGML_TYPE_Q4_1:
                                         {
@@ -797,7 +1115,16 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 8;
                                             nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q8_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q8_0_f32];
                                         } break;
                                     case GGML_TYPE_Q2_K:
                                         {
@@ -806,7 +1133,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
                                         {
@@ -815,16 +1142,16 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
                                         {
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
+                                            nth0 = 4; //1;
+                                            nth1 = 8; //32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
                                         {
@@ -833,7 +1160,7 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
                                         {
@@ -842,11 +1169,11 @@ void ggml_metal_graph_compute(
 
                                             nth0 = 2;
                                             nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
                                         } break;
                                     default:
                                         {
-                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
                                             GGML_ASSERT(false && "not implemented");
                                         }
                                 };
@@ -868,36 +1195,41 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
-                                [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
+                                [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
-                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
+                                    src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                                 }
                                 else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                                 else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
-                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                    int64_t ny = (ne11 + nrows - 1)/nrows;
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
                             }
                         } break;
                     case GGML_OP_GET_ROWS:
                         {
                             switch (src0->type) {
-                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
+                                case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f32];  break;
+                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
                                 case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
                                 case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
                                 case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
                                 case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
                                 case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
@@ -909,9 +1241,9 @@ void ggml_metal_graph_compute(
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:5];
 
                             const int64_t n = ggml_nelements(src1);
 
@@ -922,7 +1254,7 @@ void ggml_metal_graph_compute(
                             float eps;
                             memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 512;
+                            const int nth = MIN(512, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -938,16 +1270,17 @@ void ggml_metal_graph_compute(
                         } break;
                     case GGML_OP_NORM:
                         {
-                            const float eps = 1e-5f;
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
 
-                            const int nth = 256;
+                            const int nth = MIN(256, ne00);
 
                             [encoder setComputePipelineState:ctx->pipeline_norm];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
-                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
                             [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
@@ -958,17 +1291,16 @@ void ggml_metal_graph_compute(
                         {
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
+                            const int nth = MIN(1024, ne00);
+
                             const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
                             const int n_head = ((int32_t *) dst->op_params)[1];
                             float max_bias;
                             memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-                            if (__builtin_popcount(n_head) != 1) {
-                                GGML_ASSERT(false && "only power-of-two n_head implemented");
-                            }
-
                             const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
                             const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+                            const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
 
                             [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -989,12 +1321,18 @@ void ggml_metal_graph_compute(
                             [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
                             [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
                             [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
-                            const int nth = 32;
+                            [encoder setBytes:&m0   length:sizeof(   float) atIndex:18];
+                            [encoder setBytes:&m1   length:sizeof(   float) atIndex:19];
+                            [encoder setBytes:&n_heads_log2_floor   length:sizeof(int) atIndex:20];
+
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_ROPE:
                         {
+                            GGML_ASSERT(ne10 == ne02);
+
+                            const int nth = MIN(1024, ne00);
+
                             const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_dims = ((int32_t *) dst->op_params)[1];
                             const int mode   = ((int32_t *) dst->op_params)[2];
@@ -1004,38 +1342,44 @@ void ggml_metal_graph_compute(
                             memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
                             memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
-                            [encoder setComputePipelineState:ctx->pipeline_rope];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
+                            switch (src0->type) {
+                                case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
+                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_rope_f16]; break;
+                                default: GGML_ASSERT(false);
+                            };
 
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:2];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
+                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     case GGML_OP_DUP:
                     case GGML_OP_CPY:
                     case GGML_OP_CONT:
                         {
-                            const int nth = 32;
+                            const int nth = MIN(1024, ne00);
 
                             switch (src0t) {
                                 case GGML_TYPE_F32:
@@ -1057,30 +1401,30 @@ void ggml_metal_graph_compute(
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         } break;
                     default:
                         {
-                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
                             GGML_ASSERT(false);
                         }
                 }
@@ -1096,17 +1440,19 @@ void ggml_metal_graph_compute(
     }
 
     // wait for all threads to finish
-    dispatch_barrier_sync(queue, ^{});
-
-    [command_buffers[n_cb - 1] waitUntilCompleted];
+    dispatch_barrier_sync(ctx->d_queue, ^{});
 
     // check status of command buffers
     // needed to detect if the device ran out-of-memory for example (#1881)
     for (int i = 0; i < n_cb; i++) {
-        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status];
+        [ctx->command_buffers[i] waitUntilCompleted];
+
+        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
         if (status != MTLCommandBufferStatusCompleted) {
-            fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
             GGML_ASSERT(false);
         }
     }
+
+    }
 }
diff --git a/ggml-metal.metal b/ggml-metal.metal
index ce3541f4b..b6288db28 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -13,34 +13,87 @@ typedef struct {
 
 #define QK4_1 32
 typedef struct {
-    half d;          // delta
-    half m;          // min
+    half d;                 // delta
+    half m;                 // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
 
+#define QK8_0 32
+typedef struct {
+    half    d;         // delta
+    int8_t  qs[QK8_0]; // quants
+} block_q8_0;
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
 kernel void kernel_add(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig];
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant  int64_t & nb00,
+        constant  int64_t & nb01,
+        constant  int64_t & nb02,
+        constant  int64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant  int64_t & nb10,
+        constant  int64_t & nb11,
+        constant  int64_t & nb12,
+        constant  int64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant  int64_t & nb0,
+        constant  int64_t & nb1,
+        constant  int64_t & nb2,
+        constant  int64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
+
+        src0_ptr += ntg.x*nb00;
+        src1_ptr += ntg.x*nb10;
+        dst_ptr  += ntg.x*nb0;
+    }
 }
 
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_add_row(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant    int64_t & nb [[buffer(27)]],
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] + src1[tpig % ne00];
+    dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
 
 kernel void kernel_mul(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * src1[tpig];
 }
@@ -48,27 +101,27 @@ kernel void kernel_mul(
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_mul_row(
-        device const float * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant    int64_t & nb,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * src1[tpig % ne00];
+    dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
 
 kernel void kernel_scale(
-        device const float * src0,
-        device       float * dst,
+        device const float4 * src0,
+        device       float4 * dst,
         constant     float & scale,
         uint tpig[[thread_position_in_grid]]) {
     dst[tpig] = src0[tpig] * scale;
 }
 
 kernel void kernel_silu(
-        device const float * src0,
-        device       float * dst,
+        device const float4 * src0,
+        device       float4 * dst,
         uint tpig[[thread_position_in_grid]]) {
-    float x = src0[tpig];
+    device const float4 & x = src0[tpig];
     dst[tpig] = x / (1.0f + exp(-x));
 }
 
@@ -79,15 +132,27 @@ kernel void kernel_relu(
     dst[tpig] = max(0.0f, src0[tpig]);
 }
 
+kernel void kernel_sqr(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src0[tpig];
+}
+
 constant float GELU_COEF_A    = 0.044715f;
 constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
 
 kernel void kernel_gelu(
-    device const float * src0,
-    device       float * dst,
+    device const float4 * src0,
+    device       float4 * dst,
     uint tpig[[thread_position_in_grid]]) {
-    float x = src0[tpig];
-    dst[tpig] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+    device const float4 & x = src0[tpig];
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
 kernel void kernel_soft_max(
@@ -96,7 +161,6 @@ kernel void kernel_soft_max(
         constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
-        threadgroup float  * buf [[threadgroup(0)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -108,55 +172,67 @@ kernel void kernel_soft_max(
     device       float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
 
     // parallel max
-    buf[tpitg[0]] = -INFINITY;
-    for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
-        buf[tpitg[0]] = MAX(buf[tpitg[0]], psrc0[i00]);
+    float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY;
+    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
+        lmax = MAX(lmax, psrc0[i00]);
     }
-
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg[0]/2; i > 0; i /= 2) {
-        if (tpitg[0] < i) {
-            buf[tpitg[0]] = MAX(buf[tpitg[0]], buf[tpitg[0] + i]);
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    // broadcast
-    if (tpitg[0] == 0) {
-        buf[0] = buf[0];
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    const float max = buf[0];
+    const float max = simd_max(lmax);
 
     // parallel sum
-    buf[tpitg[0]] = 0.0f;
+    float lsum = 0.0f;
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
-        buf[tpitg[0]] += exp(psrc0[i00] - max);
+        const float exp_psrc0 = exp(psrc0[i00] - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // whish to compute it twice.
+        pdst[i00] = exp_psrc0;
     }
 
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg[0]/2; i > 0; i /= 2) {
-        if (tpitg[0] < i) {
-            buf[tpitg[0]] += buf[tpitg[0] + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-
-    // broadcast
-    if (tpitg[0] == 0) {
-        buf[0] = buf[0];
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    const float sum = buf[0];
+    const float sum = simd_sum(lsum);
 
     for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
-        pdst[i00] = exp(psrc0[i00] - max) / sum;
+        pdst[i00] /= sum;
+    }
+}
+
+kernel void kernel_soft_max_4(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    device       float4 * pdst4 = (device       float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    // parallel max
+    float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY;
+    for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        lmax4 = fmax(lmax4, psrc4[i00]);
+    }
+    float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    const float max = simd_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        const float4 exp_psrc4 = exp(psrc4[i00] - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    const float sum = simd_sum(lsum);
+
+    for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
+        pdst4[i00] /= sum;
     }
 }
 
@@ -175,6 +251,33 @@ kernel void kernel_diag_mask_inf(
         dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
     } else {
         dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+     }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        device const float4 * src0,
+        device       float4 * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne01,
+        constant        int & n_past,
+        uint3 tpig[[thread_position_in_grid]]) {
+
+    const int64_t i = 2*tpig[0];
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int64_t i4 = 4*i;
+    const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    const int64_t i01 = i4/(ne00);      i4 -= i01*ne00;
+    const int64_t i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        dst[i+1][k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            dst[i][k] = -INFINITY;
+        }
     }
 }
 
@@ -203,25 +306,17 @@ kernel void kernel_norm(
         }
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
-    // broadcast
-    if (tpitg == 0) {
-        sum[0] /= ne00;
-    }
+    const float mean  = sum[0] / ne00;
+
+    // recenter and VARIANCE
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    const float mean  = sum[0];
-
-    // recenter
     device float * y = dst + tgpig*ne00;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = x[i00] - mean;
-    }
-
-    // VARIANCE
-    // parallel sum
     sum[tpitg] = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = x[i00] - mean;
         sum[tpitg] += y[i00] * y[i00];
     }
+
     // reduce
     threadgroup_barrier(mem_flags::mem_threadgroup);
     for (uint i = ntg/2; i > 0; i /= 2) {
@@ -230,12 +325,7 @@ kernel void kernel_norm(
         }
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
-    // broadcast
-    if (tpitg == 0) {
-        sum[0] /= ne00;
-    }
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    const float variance = sum[0];
+    const float variance = sum[0] / ne00;
 
     const float scale = 1.0f/sqrt(variance + eps);
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
@@ -243,7 +333,6 @@ kernel void kernel_norm(
     }
 }
 
-
 kernel void kernel_rms_norm(
         device const  void * src0,
         device       float * dst,
@@ -334,8 +423,8 @@ inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thre
 }
 
 // putting them in the kernel cause a significant performance penalty
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_DST 4        // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2  // number of SIMD groups in a thread group
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 //Note: This is a template, but strictly speaking it only applies to
 //      quantizations where the block size is 32. It also does not
@@ -346,18 +435,23 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
                     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne10, int64_t ne12, int64_t ne0, int64_t ne1, uint gqa,
                     uint3 tgpig, uint tiisg, uint sgitg) {
     const int nb = ne00/QK4_0;
+
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
-    const int first_row = (r0 * nsg + sgitg) * nr;
-    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
-    device const block_q_type * x = (device const block_q_type *) src0 + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-    float yl[16];       // src1 vector cache
-    float sumf[nr]={0.f};
 
-    const int ix = tiisg/2;
-    const int il = 8*(tiisg%2);
+    const int first_row = (r0 * nsg + sgitg) * nr;
+
+    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+
+    device const block_q_type * x = (device const block_q_type *) src0 + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16]; // src1 vector cache
+    float sumf[nr] = {0.f};
+
+    const int ix = (tiisg/2);
+    const int il = (tiisg%2)*8;
 
     device const float * yb = y + ix * QK4_0 + il;
 
@@ -368,6 +462,7 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
             sumy += yb[i] + yb[i+1];
             yl[i+0] = yb[i+ 0];
             yl[i+1] = yb[i+ 1]/256.f;
+
             sumy += yb[i+16] + yb[i+17];
             yl[i+8] = yb[i+16]/16.f;
             yl[i+9] = yb[i+17]/4096.f;
@@ -383,12 +478,12 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
         }
     }
 }
 
-kernel void kernel_mul_mat_q4_0_f32(
+kernel void kernel_mul_mv_q4_0_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -401,12 +496,12 @@ kernel void kernel_mul_mat_q4_0_f32(
         constant   int64_t & ne1[[buffer(16)]],
         constant   uint    & gqa[[buffer(17)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
     mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
-kernel void kernel_mul_mat_q4_1_f32(
+kernel void kernel_mul_mv_q4_1_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -424,7 +519,73 @@ kernel void kernel_mul_mat_q4_1_f32(
      mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
 }
 
-kernel void kernel_mul_mat_f16_f32(
+#define NB_Q8_0 8
+
+kernel void kernel_mul_mv_q8_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01[[buffer(4)]],
+        constant   int64_t & ne02[[buffer(5)]],
+        constant   int64_t & ne10[[buffer(9)]],
+        constant   int64_t & ne12[[buffer(11)]],
+        constant   int64_t & ne0[[buffer(15)]],
+        constant   int64_t & ne1[[buffer(16)]],
+        constant   uint    & gqa[[buffer(17)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = ne00/QK8_0;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * nsg + sgitg) * nr;
+    const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
+    device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[NB_Q8_0];
+    float sumf[nr]={0.f};
+
+    const int ix = tiisg/4;
+    const int il = tiisg%4;
+
+    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
+
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
+            float sumq = 0.f;
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*x[ib+row*nb].d;
+        }
+
+        yb += NB_Q8_0 * nw;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+        }
+    }
+}
+
+#define N_F32_F32 4
+
+kernel void kernel_mul_mv_f32_f32(
         device const  char * src0,
         device const  char * src1,
         device       float * dst,
@@ -442,11 +603,79 @@ kernel void kernel_mul_mat_f16_f32(
         constant  uint64_t & nb12,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
-        threadgroup float  * sum [[threadgroup(0)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3  tpig[[thread_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3  tptg[[threads_per_threadgroup]]) {
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t rb = tgpig.y*N_F32_F32;
+    const int64_t im = tgpig.z;
+
+    device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        device const float4 * x4 = (device const float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
+            device const float4 * y4 = (device const float4 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
+
+kernel void kernel_mul_mv_f16_f32_1row(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
 
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
@@ -455,26 +684,145 @@ kernel void kernel_mul_mat_f16_f32(
     device const half  * x = (device const half  *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
     device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
 
-    sum[tpitg.x] = 0.0f;
-
-    for (int i = tpitg.x; i < ne00; i += tptg.x) {
-        sum[tpitg.x] += (float) x[i] * (float) y[i];
-    }
-
-    // accumulate the sum from all threads in the threadgroup
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = tptg.x/2; i > 0; i /= 2) {
-        if (tpitg.x < i) {
-            sum[tpitg.x] += sum[tpitg.x + i];
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = tiisg; i < ne00; i += 32) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        device const half4  * x4 = (device const half4  *) x;
+        device const float4 * y4 = (device const float4 *) y;
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k];
+        }
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
         }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    if (tpitg.x == 0) {
-        dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0];
+}
+
+#define N_F16_F32 4
+
+kernel void kernel_mul_mv_f16_f32(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t rb = tgpig.y*N_F16_F32;
+    const int64_t im = tgpig.z;
+
+    device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        device const half4 * x4 = (device const half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const float  * y  = (device const float  *) (src1 + r1*nb11 + im*nb12);
+            device const float4 * y4 = (device const float4 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
     }
 }
 
+// Assumes row size (ne00) is a multiple of 4
+kernel void kernel_mul_mv_f16_f32_l4(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int nrows = ne11;
+    const int64_t r0 = tgpig.x;
+    const int64_t im = tgpig.z;
+
+    device const half4 * x4 = (device const half4 *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
+
+        float sumf = 0;
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
+        }
+
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
 
 kernel void kernel_alibi_f32(
         device const float * src0,
@@ -495,7 +843,9 @@ kernel void kernel_alibi_f32(
         constant  uint64_t & nb1,
         constant  uint64_t & nb2,
         constant  uint64_t & nb3,
-        constant      float & m0,
+        constant     float & m0,
+        constant     float & m1,
+        constant       int & n_heads_log2_floor,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
@@ -511,70 +861,131 @@ kernel void kernel_alibi_f32(
     const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
 
     device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-    float m_k = pow(m0, i2 + 1);
+    float m_k;
+    if (i2 < n_heads_log2_floor) {
+        m_k = pow(m0, i2 + 1);
+    } else {
+        m_k = pow(m1, 2 * (i2 - n_heads_log2_floor) + 1);
+    }
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
         device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
         dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
     }
 }
 
+typedef void (rope_t)(
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]);
+
+template<typename T>
 kernel void kernel_rope(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant       int & n_past,
-        constant       int & n_dims,
-        constant       int & mode,
-        constant     float & freq_base,
-        constant     float & freq_scale,
-        uint3 tpig[[thread_position_in_grid]]) {
-    const int64_t i3 = tpig[2];
-    const int64_t i2 = tpig[1];
-    const int64_t i1 = tpig[0];
+        device const    void * src0,
+        device const int32_t * src1,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & mode,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]) {
+    const int64_t i3 = tgpig[2];
+    const int64_t i2 = tgpig[1];
+    const int64_t i1 = tgpig[0];
 
     const bool is_neox = mode & 2;
-    const float theta_scale = pow(freq_base, -2.0f/n_dims);
 
-    const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+    device const int32_t * pos = src1;
 
-    float theta = freq_scale * (float)p;
+    const int64_t p = pos[i2];
+
+    const float theta_0 = freq_scale * (float)p;
+    const float inv_ndims = -1.f/n_dims;
 
     if (!is_neox) {
-        for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
+
+            const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
             const float cos_theta = cos(theta);
             const float sin_theta = sin(theta);
 
-            theta *= theta_scale;
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-            device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       float * dst_data  = (device float *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-            const float x0 = src[0];
-            const float x1 = src[1];
+            const T x0 = src[0];
+            const T x1 = src[1];
 
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         }
     } else {
-        // TODO: implement
+        for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
+            for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
+
+                const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
+                const float cos_theta = cos(theta);
+                const float sin_theta = sin(theta);
+
+                const int64_t i0 = ib*n_dims + ic/2;
+
+                device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+                const float x0 = src[0];
+                const float x1 = src[n_dims/2];
+
+                dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            }
+        }
     }
 }
 
+template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
+template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
+
 kernel void kernel_cpy_f16_f16(
         device const half * src0,
         device       half * dst,
@@ -700,6 +1111,62 @@ kernel void kernel_cpy_f32_f32(
     }
 }
 
+kernel void kernel_concat(
+    device const char * src0,
+    device const char * src1,
+    device       char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne10,
+    constant   int64_t & ne11,
+    constant   int64_t & ne12,
+    constant   int64_t & ne13,
+    constant  uint64_t & nb10,
+    constant  uint64_t & nb11,
+    constant  uint64_t & nb12,
+    constant  uint64_t & nb13,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03 * nb03 + i02 * nb02 + i01 * nb01 + tpitg.x*nb00;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + tpitg.x*nb0;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        if (i02 < ne02) {
+            ((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
+            src0_ptr += ntg.x*nb00;
+        } else {
+            ((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
+            src1_ptr += ntg.x*nb10;
+        }
+        dst_ptr += ntg.x*nb0;
+    }
+}
+
 //============================================ k-quants ======================================================
 
 #ifndef QK_K
@@ -792,7 +1259,7 @@ static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
 
 //====================================== dot products =========================
 
-kernel void kernel_mul_mat_q2_K_f32(
+kernel void kernel_mul_mv_q2_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -936,7 +1403,7 @@ kernel void kernel_mul_mat_q2_K_f32(
 }
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -963,31 +1430,40 @@ kernel void kernel_mul_mat_q3_K_f32(
     device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
     device const float     * yy = (device const float      *) src1 + r1*ne10 + r2*ne00*ne1;
 
-    float yl[16];
+    float yl[32];
 
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
+    //const uint16_t kmask1 = 0x3030;
+    //const uint16_t kmask2 = 0x0f0f;
 
-    const int tid = tiisg/2;
-    const int ix  = tiisg%2;
-    const int ip  = tid/8;          // 0 or 1
-    const int il  = tid/2 - 4*ip;   // 0...3
+    const int tid = tiisg/4;
+    const int ix  = tiisg%4;
+    const int ip  = tid/4;          // 0 or 1
+    const int il  = 2*((tid%4)/2);  // 0 or 2
     const int ir  = tid%2;
     const int n   = 8;
     const int l0  = n*ir;
 
-    const uint16_t m1 = 1 << (4*ip + il);
-    const uint16_t m2 = m1 << 8;
+    // One would think that the Metal compiler would figure out that ip and il can only have
+    // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
+    // with these two tales.
+    //
+    // Possible masks for the high bit
+    const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200},  // ip = 0, il = 0
+                           {0x0004, 0x0400, 0x0008, 0x0800},  // ip = 0, il = 2
+                           {0x0010, 0x1000, 0x0020, 0x2000},  // ip = 1, il = 0
+                           {0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
+
+    // Possible masks for the low 2 bits
+    const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
+
+    const ushort4 hm = mm[2*ip + il/2];
 
     const int shift = 2*il;
-    const uint16_t qm1 = 0x0003 << shift;
-    const uint16_t qm2 = 0x0300 << shift;
-    const int32_t v1 = 4 << shift;
-    const int32_t v2 = 1024 << shift;
+    const float    v1 = il == 0 ? 4.f : 64.f;
+    const float    v2 = 4.f * v1;
 
     const uint16_t s_shift1 = 4*ip;
-    const uint16_t s_shift2 = s_shift1 + 2*(il/2);
-    const int ik = 4 + (il%2);
+    const uint16_t s_shift2 = s_shift1 + il;
 
     const int q_offset = 32*ip + l0;
     const int y_offset = 128*ip + 32*il + l0;
@@ -996,12 +1472,19 @@ kernel void kernel_mul_mat_q3_K_f32(
 
     device const float * y1 = yy + ix*QK_K + y_offset;
 
-    float sumf1[2] = {0.f}, sumf2[2] = {0.f};
-    for (int i = ix; i < nb; i += 2) {
+    uint32_t scales32, aux32;
+    thread uint16_t * scales16 = (thread uint16_t *)&scales32;
+    thread const int8_t * scales = (thread const int8_t *)&scales32;
+
+    float sumf1[2] = {0.f};
+    float sumf2[2] = {0.f};
+    for (int i = ix; i < nb; i += 4) {
 
         for (int l = 0; l < 8; ++l) {
-            yl[l+0] = y1[l+ 0];
-            yl[l+8] = y1[l+16];
+            yl[l+ 0] = y1[l+ 0];
+            yl[l+ 8] = y1[l+16];
+            yl[l+16] = y1[l+32];
+            yl[l+24] = y1[l+48];
         }
 
         device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
@@ -1012,27 +1495,43 @@ kernel void kernel_mul_mat_q3_K_f32(
         for (int row = 0; row < 2; ++row) {
 
             const float d_all = (float)dh[0];
-            const char2 scales = as_type<char2>((uint16_t)(((a[il] >> s_shift1) & kmask2) | (((a[ik] >> s_shift2) & kmask1) << 4)));
 
-            float s1 = 0, s2 = 0;
-            for (int l = 0; l < n; l += 2) {
-                const uint16_t qs = q[l/2];
-                s1 += yl[l+0] * ((int32_t)(qs & qm1) - ((h[l/2] & m1) ? 0 : v1));
-                s2 += yl[l+1] * ((int32_t)(qs & qm2) - ((h[l/2] & m2) ? 0 : v2));
-            }
-            float d = d_all * (s1 + 1.f/256.f * s2);
-            sumf1[row] += d * scales[0];
-            sumf2[row] += d;
+            scales16[0] = a[4];
+            scales16[1] = a[5];
+            aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
+            scales16[0] = a[il+0];
+            scales16[1] = a[il+1];
+            scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
 
-            s1 = s2 = 0;
+            float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
             for (int l = 0; l < n; l += 2) {
-                const uint16_t qs = q[l/2+8];
-                s1 += yl[l+8] * ((int32_t)(qs & qm1) - ((h[l/2+8] & m1) ? 0 : v1));
-                s2 += yl[l+9] * ((int32_t)(qs & qm2) - ((h[l/2+8] & m2) ? 0 : v2));
+                const int32_t qs = q[l/2];
+                s1 += yl[l+0] * (qs & qm[il/2][0]);
+                s2 += yl[l+1] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
+                s4 += yl[l+16] * (qs & qm[il/2][2]);
+                s5 += yl[l+17] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
             }
-            d = d_all * (s1 + 1.f/256.f * s2);
-            sumf1[row] += d * scales[1];
-            sumf2[row] += d;
+            float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
+            float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
+            sumf1[row] += d1 * (scales[0] - 32);
+            sumf2[row] += d2 * (scales[2] - 32);
+
+            s1 = s2 = s3 = s4 = s5 = s6 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const int32_t qs = q[l/2+8];
+                s1 += yl[l+8] * (qs & qm[il/2][0]);
+                s2 += yl[l+9] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
+                s4 += yl[l+24] * (qs & qm[il/2][2]);
+                s5 += yl[l+25] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
+            }
+            d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
+            d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
+            sumf1[row] += d1 * (scales[1] - 32);
+            sumf2[row] += d2 * (scales[3] - 32);
 
             q  += step;
             h  += step;
@@ -1041,20 +1540,22 @@ kernel void kernel_mul_mat_q3_K_f32(
 
         }
 
-        y1 += 2 * QK_K;
+        y1 += 4 * QK_K;
 
     }
 
     for (int row = 0; row < 2; ++row) {
-        const float sumf = (sumf1[row] - 32.f*sumf2[row]) / (1 << shift);
-        const float tot = simd_sum(sumf);
-        if (tiisg == 0) {
-            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = tot;
+        const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
+        sumf1[row] = simd_sum(sumf);
+    }
+    if (tiisg == 0) {
+        for (int row = 0; row < 2; ++row) {
+            dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row];
         }
     }
 }
 #else
-kernel void kernel_mul_mat_q3_K_f32(
+kernel void kernel_mul_mv_q3_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1125,18 +1626,18 @@ kernel void kernel_mul_mat_q3_K_f32(
 #endif
 
 #if QK_K == 256
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
         constant   int64_t & ne00,
-        constant   int64_t & ne01[[buffer(4)]],
-        constant   int64_t & ne02[[buffer(5)]],
-        constant   int64_t & ne10[[buffer(9)]],
-        constant   int64_t & ne12[[buffer(11)]],
-        constant   int64_t & ne0[[buffer(15)]],
-        constant   int64_t & ne1[[buffer(16)]],
-        constant   uint    & gqa[[buffer(17)]],
+        constant   int64_t & ne01 [[buffer(4)]],
+        constant   int64_t & ne02 [[buffer(5)]],
+        constant   int64_t & ne10 [[buffer(9)]],
+        constant   int64_t & ne12 [[buffer(11)]],
+        constant   int64_t & ne0  [[buffer(15)]],
+        constant   int64_t & ne1  [[buffer(16)]],
+        constant   uint    & gqa  [[buffer(17)]],
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
@@ -1154,7 +1655,8 @@ kernel void kernel_mul_mat_q4_K_f32(
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int r2 = tgpig.z;
-    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int first_row = r0 * N_DST;
     const int ib_row = first_row * nb;
     const uint offset0 = r2/gqa*(nb*ne0);
     device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
@@ -1230,7 +1732,7 @@ kernel void kernel_mul_mat_q4_K_f32(
     }
 }
 #else
-kernel void kernel_mul_mat_q4_K_f32(
+kernel void kernel_mul_mv_q4_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1319,7 +1821,7 @@ kernel void kernel_mul_mat_q4_K_f32(
 }
 #endif
 
-kernel void kernel_mul_mat_q5_K_f32(
+kernel void kernel_mul_mv_q5_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1403,17 +1905,25 @@ kernel void kernel_mul_mat_q5_K_f32(
             sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
             sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
 
-            float4 acc = {0.f, 0.f, 0.f, 0.f};
+            float4 acc1 = {0.f};
+            float4 acc2 = {0.f};
             for (int l = 0; l < n; ++l) {
                 uint8_t h = qh[l];
-                acc[0] += yl[l+0] * ((uint16_t)(q1[l] & 0x0F) + (h & hm1 ? 16 : 0));
-                acc[1] += yl[l+8] * ((uint16_t)(q1[l] & 0xF0) + (h & hm2 ? 256 : 0));
-                acc[2] += yh[l+0] * ((uint16_t)(q2[l] & 0x0F) + (h & hm3 ? 16 : 0));
-                acc[3] += yh[l+8] * ((uint16_t)(q2[l] & 0xF0) + (h & hm4 ? 256 : 0));
+                acc1[0] += yl[l+0] * (q1[l] & 0x0F);
+                acc1[1] += yl[l+8] * (q1[l] & 0xF0);
+                acc1[2] += yh[l+0] * (q2[l] & 0x0F);
+                acc1[3] += yh[l+8] * (q2[l] & 0xF0);
+                acc2[0] += h & hm1 ? yl[l+0] : 0.f;
+                acc2[1] += h & hm2 ? yl[l+8] : 0.f;
+                acc2[2] += h & hm3 ? yh[l+0] : 0.f;
+                acc2[3] += h & hm4 ? yh[l+8] : 0.f;
             }
             const float dall = dh[0];
             const float dmin = dh[1];
-            sumf[row] += dall * (acc[0] * sc8[0] + acc[1] * sc8[1] * 1.f/16.f + acc[2] * sc8[4] + acc[3] * sc8[5] * 1.f/16.f) -
+            sumf[row] += dall * (sc8[0] * (acc1[0] +  16.f*acc2[0]) +
+                                 sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
+                                 sc8[4] * (acc1[2] +  16.f*acc2[2]) +
+                                 sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
             q1 += step;
@@ -1484,7 +1994,7 @@ kernel void kernel_mul_mat_q5_K_f32(
 
 }
 
-kernel void kernel_mul_mat_q6_K_f32(
+kernel void kernel_mul_mv_q6_K_f32(
         device const  void * src0,
         device const float * src1,
         device       float * dst,
@@ -1586,6 +2096,15 @@ kernel void kernel_mul_mat_q6_K_f32(
 
 //============================= templates and their specializations =============================
 
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    float4x4 temp = *(((device float4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
 template <typename type4x4>
 void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
     half4x4 temp = *(((device half4x4 *)src));
@@ -1597,28 +2116,40 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
 template <typename type4x4>
 void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 1);
-    const half d = il ? (xb->d / 16.h) : xb->d;
-    const half m = il ? (-8.h * 16.h) : -8.h;
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
     const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = il ? 0xF000 : 0x0F00;
+    const ushort mask1 = mask0 << 8;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)] = (((qs[i] & mask0)) + m) * d;
-        reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) + m) * d;
+        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
+        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
     }
 }
 
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
-    const half d = il ? (xb->d / 16.h) : xb->d;
-    const half m = xb->m;
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
     const ushort mask0 = il ? 0x00F0 : 0x000F;
-    const ushort mask1 = il ? 0xF000 : 0x0F00;
+    const ushort mask1 = mask0 << 8;
 
     for (int i=0;i<8;i++) {
-        reg[i/2][2*(i%2)] = (((qs[i] & mask0)) * d) + m;
-        reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) * d) + m;
+        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
+        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const half d = xb->d;
+
+    for (int i=0;i<16;i++) {
+        reg[i/4][i%4] = (qs[i + 16*il] * d);
     }
 }
 
@@ -1644,7 +2175,7 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
-    const float d_all = (float)(xb->d);
+    const half d_all = xb->d;
     device const uint8_t * q = (device const uint8_t *)xb->qs;
     device const uint8_t * h = (device const uint8_t *)xb->hmask;
     device const int8_t * scales = (device const int8_t *)xb->scales;
@@ -1657,16 +2188,18 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
                                  ((il/4)>0 ? 12  : 3);
     uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
     uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
-    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2) : \
-                                 (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h);
+    const half ml = 4.h * dl;
 
-    il = (il/2)%4;
-    float   coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
 
     for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i] & m) ? 0 : 4.f/coef));
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
     }
 #else
     float    kcoef = il&1 ? 1.f/16.f : 1.f;
@@ -1681,26 +2214,31 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
 #endif
 }
 
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
 template <typename type4x4>
 void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
-    device const uint8_t * q = xb->qs;
+    device const uchar * q = xb->qs;
 
 #if QK_K == 256
-    const float d = (float)(xb->d);
-    const float min = (float)(xb->dmin);
     short is = (il/4) * 2;
     q = q + (il/4) * 32 + 16 * (il&1);
-    il = il%4;
-    const uchar4 sc = get_scale_min_k4(is, xb->scales);
-    const float dl = il<2 ? d * sc[0]   : d * sc[2]/16.h;
-    const float ml = il<2 ? min * sc[1] : min * sc[3];
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const half d   = il < 2 ? xb->d : xb->d / 16.h;
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 #else
     q = q + 16 * (il&1);
     device const uint8_t * s = xb->scales;
     device const half2 * dh = (device const half2 *)xb->d;
     const float2 d = (float2)dh[0];
     const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1 ]* (s[1]>>4);
+    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
 #endif
     const ushort mask = il<2 ? 0x0F : 0xF0;
     for (int i = 0; i < 16; ++i) {
@@ -1714,19 +2252,19 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     device const uint8_t * qh = xb->qh;
 
 #if QK_K == 256
-    const float d = (float)(xb->d);
-    const float min = (float)(xb->dmin);
     short is = (il/4) * 2;
     q  = q + 32 * (il/4) + 16 * (il&1);
     qh = qh + 16 * (il&1);
     uint8_t ul = 1 << (il/2);
-    il = il%4;
-    const uchar4 sc = get_scale_min_k4(is, xb->scales);
-    const float dl = il<2 ? d * sc[0]   : d * sc[2]/16.h;
-    const float ml = il<2 ? min * sc[1] : min * sc[3];
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const half d = il < 2 ? xb->d : xb->d / 16.h;
+    const half min = xb->dmin;
+    const half dl = d * sc[0];
+    const half ml = min * sc[1];
 
-    const ushort mask   = il<2 ? 0x0F : 0xF0;
-    const float  qh_val = il<2 ? 16.f : 256.f;
+    const ushort mask = il<2 ? 0x0F : 0xF0;
+    const half qh_val = il<2 ? 16.h : 256.h;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
     }
@@ -1745,7 +2283,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
 
 template <typename type4x4>
 void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
-    const float d_all = (float)(xb->d);
+    const half d_all = xb->d;
     device const uint8_t * ql = (device const uint8_t *)xb->ql;
     device const uint8_t * qh = (device const uint8_t *)xb->qh;
     device const int8_t * scales = (device const int8_t *)xb->scales;
@@ -1753,19 +2291,21 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 #if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
-    float sc = scales[(il%2) + 2 * ((il/2))];
-    il = (il/2)%4;
+    half sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
 #else
     ql = ql + 16 * (il&1);
-    float sc = scales[il];
+    half sc = scales[il];
 #endif
+    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
+    const half        coef = il>1 ? 1.f/16.h          : 1.h;
+    const half ml = d_all * sc * 32.h;
+    const half dl = d_all * sc * coef;
     for (int i = 0; i < 16; ++i) {
-        uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
-        uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-        const float coef = il>1 ? 1.f/16.f          : 1.f;
-        float q = il&1 ? ((ql[i]&kmask2)|((qh[i]&kmask1)<<2)) - 32.f/coef : \
-                         ((ql[i]&kmask2)|((qh[i]&kmask1)<<4)) - 32.f/coef;
-        reg[i/4][i%4] = d_all * sc * q * coef;
+        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
+                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
+        reg[i/4][i%4] = dl * q - ml;
     }
 }
 
@@ -1792,7 +2332,7 @@ kernel void kernel_get_rows(
 }
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix A
+#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
 #define BLOCK_SIZE_K 32
 #define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
 #define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
@@ -1805,35 +2345,40 @@ kernel void kernel_get_rows(
 // each block_q contains 16*nl weights
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm(device const  uchar * src0,
-                           device const  float * src1,
-                           device        float * dst,
-                           constant    int64_t & ne00,
-                           constant    int64_t & ne02,
-                           constant    int64_t & nb01,
-                           constant    int64_t & nb02,
-                           constant    int64_t & ne12,
-                           constant    int64_t & ne0,
-                           constant    int64_t & ne1,
-                           constant    uint & gqa,
-                           threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                           uint3                 tgpig[[threadgroup_position_in_grid]],
-                           uint                  tiitg[[thread_index_in_threadgroup]],
-                           uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+                          device const  uchar * src1,
+                          device        float * dst,
+                          constant    int64_t & ne00,
+                          constant    int64_t & ne02,
+                          constant    int64_t & nb01,
+                          constant    int64_t & nb02,
+                          constant    int64_t & ne12,
+                          constant    int64_t & nb10,
+                          constant    int64_t & nb11,
+                          constant    int64_t & nb12,
+                          constant    int64_t & ne0,
+                          constant    int64_t & ne1,
+                          constant       uint & gqa,
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    threadgroup half * sa = ((threadgroup half *)shared_memory);
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
     threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
 
     const uint r0 = tgpig.y;
     const uint r1 = tgpig.x;
     const uint im = tgpig.z;
+
     // if this block is of 64x32 shape or smaller
     short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
     short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
     // a thread shouldn't load data outside of the matrix
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
     short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
-    simdgroup_half8x8 ma[4];
+    simdgroup_half8x8  ma[4];
     simdgroup_float8x8 mb[2];
     simdgroup_float8x8 c_res[8];
     for (int i = 0; i < 8; i++){
@@ -1841,32 +2386,41 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     short il = (tiitg % THREAD_PER_ROW);
-    uint offset0 = im/gqa*nb02; ushort offset1 = il/nl;
-    device const block_q  * x = (device const block_q  *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
-    device const float * y = src1 + (r1 * BLOCK_SIZE_N + thread_col) * ne00 \
-                             + BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL) + im * ne00 * ne1;
+
+    uint   offset0 = im/gqa*nb02;
+    ushort offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
-        //load data and store to threadgroup memory
+        // load data and store to threadgroup memory
         half4x4 temp_a;
         dequantize_func(x, il, temp_a);
         threadgroup_barrier(mem_flags::mem_threadgroup);
+
         #pragma unroll(16)
         for (int i = 0; i < 16; i++) {
             *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
-            + 16 * (tiitg % THREAD_PER_ROW) + 8 * (i / 8)) \
-            + (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
         }
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) \
-                = *((device float2x4 *)y);
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2+nl-1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        //load matrices from threadgroup memory and conduct outer products
+
+        // load matrices from threadgroup memory and conduct outer products
         threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
         threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
         #pragma unroll(4)
         for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
             #pragma unroll(4)
@@ -1881,6 +2435,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
             lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
             lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
             #pragma unroll(8)
             for (int i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
@@ -1889,25 +2444,26 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     }
 
     if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float *C = dst + BLOCK_SIZE_M * r0 + 32 * (sgitg&1) \
-                          + (BLOCK_SIZE_N * r1 + 16 * (sgitg>>1)) * ne0 + im*ne1*ne0;
+        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
+                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
         for (int i = 0; i < 8; i++) {
             simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
         for (int i = 0; i < 8; i++) {
             simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
-        if (sgitg==0) {
+
+        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+        if (sgitg == 0) {
             for (int i = 0; i < n_rows; i++) {
-                for (int j = tiitg; j< n_cols; j += BLOCK_SIZE_N) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
                     *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
                 }
             }
@@ -1924,22 +2480,39 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
                           constant uint64_t &, constant uint64_t &, uint, uint, uint);
 
-template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows<half4x4, 1, dequantize_f16>;
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_t kernel_get_rows<float4x4,   1, dequantize_f32>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_t kernel_get_rows<half4x4,    1, dequantize_f16>;
 template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
 template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
 template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
 
-typedef void (mat_mm_t)(device const uchar *, device const float *, device float *, constant int64_t &,\
-                             constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \
-                             constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint);
+typedef void (mat_mm_t)(
+        device const  uchar * src0,
+        device const  uchar * src1,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant    int64_t & nb01,
+        constant    int64_t & nb02,
+        constant    int64_t & ne12,
+        constant    int64_t & nb10,
+        constant    int64_t & nb11,
+        constant    int64_t & nb12,
+        constant    int64_t & ne0,
+        constant    int64_t & ne1,
+        constant       uint & gqa,
+        threadgroup uchar *, uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
 template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
 template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp
index eb214a836..4a331f24a 100644
--- a/ggml-opencl.cpp
+++ b/ggml-opencl.cpp
@@ -202,14 +202,14 @@ inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8
 
 __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int n = tid / 32;
     const int l = tid - 32 * n;
     const int is = 8 * n + l / 16;
 
     const uint8_t q = x[i].qs[32 * n + l];
-    __global float *y = yy + i * QK_K + 128 * n;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * n;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -223,7 +223,7 @@ __kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __globa
 __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
 {
     int r = get_local_id(0) / 4;
-    int i = get_group_id(0);
+    int i = get_group_id(0) + get_global_offset(0);
     int tid = r / 2;
     int is0 = r % 2;
     int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
@@ -241,7 +241,7 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa
     float d_all = vload_half(0, &x[i].d);
     float dl = d_all * (us - 32);
 
-    __global float *y = yy + i * QK_K + 128 * n + 32 * j;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * n + 32 * j;
     const __global uint8_t *q = x[i].qs + 32 * n;
     const __global uint8_t *hm = x[i].hmask;
 
@@ -251,14 +251,14 @@ __kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __globa
 
 __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int il = tid / 8;
     const int ir = tid % 8;
     const int is = 2 * il;
     const int n = 4;
 
-    __global float *y = yy + i * QK_K + 64 * il + n * ir;
+    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + n * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -281,13 +281,13 @@ __kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __globa
 
 __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int il = tid / 16;
     const int ir = tid % 16;
     const int is = 2 * il;
 
-    __global float *y = yy + i * QK_K + 64 * il + 2 * ir;
+    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + 2 * ir;
 
     const float dall = vload_half(0, &x[i].d);
     const float dmin = vload_half(0, &x[i].dmin);
@@ -313,13 +313,13 @@ __kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __globa
 
 __kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
 {
-    const int i = get_group_id(0);
+    const int i = get_group_id(0) + get_global_offset(0);
     const int tid = get_local_id(0);
     const int ip = tid / 32;
     const int il = tid - 32 * ip;
     const int is = 8 * ip + il / 16;
 
-    __global float *y = yy + i * QK_K + 128 * ip + il;
+    __global float *y = yy + get_group_id(0) * QK_K + 128 * ip + il;
 
     const float d = vload_half(0, &x[i].d);
 
@@ -730,7 +730,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
     const uint qk = QUANT_K;
     const uint qr = QUANT_R;
 
-    const int ib = i/qk; // block index
+    const int ib = i/qk + get_global_offset(0); // block index
     const int iqs = (i%qk)/qr; // quant index
     const int iybs = i - i%qk; // y block start index
     const int y_offset = qr == 1 ? 1 : qk/2;
@@ -847,7 +847,7 @@ std::array<std::string, 2> mul_str_values = {
     "mul_f32", "float"
 };
 
-std::string& replace(std::string& s, const std::string& from, const std::string& to) {
+static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
     size_t pos = 0;
     while ((pos = s.find(from, pos)) != std::string::npos) {
          s.replace(pos, from.length(), to);
@@ -856,7 +856,7 @@ std::string& replace(std::string& s, const std::string& from, const std::string&
     return s;
 }
 
-std::string generate_kernels() {
+static std::string generate_kernels() {
     std::stringstream src;
     src << program_source << '\n';
     src << k_quants_source << '\n';
@@ -1334,7 +1334,7 @@ void ggml_cl_free_data(const struct ggml_tensor* tensor) {
         return;
     }
 
-    cl_mem mem = (cl_mem)tensor->data;
+    cl_mem mem = (cl_mem)tensor->extra;
     clReleaseMemObject(mem);
 }
 
@@ -1349,30 +1349,42 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
     const enum ggml_type type = src->type;
     const size_t ts = ggml_type_size(type);
     const size_t bs = ggml_blck_size(type);
+    const uint64_t row_size = ts*ne0/bs;
 
-    const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
-    if (nb0 == ts && nb1 == ts*ne0/bs) {
-        err = clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*nb1, x, 0, NULL, ev);
-        return err;
+    const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
+    if (nb0 == ts && nb1 == row_size) {
+        return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
     }
     if (nb0 == ts) {
         const size_t buffer_origin[3] = { offset, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts*ne0/bs, ne1, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts*ne0/bs, 0, nb1, 0, x, 0, NULL, ev);
-        return err;
+        const size_t region[3] = { row_size, ne1, 1 };
+        return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
     }
+    std::vector<cl_event> events;
+    if (ev && ne1>1) events.reserve(ne1-1);
     for (uint64_t i1 = 0; i1 < ne1; i1++) {
         // pretend the row is a matrix with cols=1
-        const size_t buffer_origin[3] = { offset, i1, 0 };
+        const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
         const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts/bs, ne0, 1 };
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, 0, 0, nb0, 0, ((const char *)x) + i1*nb0, 0, NULL, ev);
+        const size_t region[3] = { ts, ne0/bs, 1 };
+        // if an event is requested, make the last write wait for all previous writes to complete
+        if (ev && i1) {
+            events.push_back(*ev);
+        }
+        cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
+        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
         if (err != CL_SUCCESS) {
-            break;
+            for (auto event : events) {
+                clReleaseEvent(event);
+            }
+            return err;
         }
     }
-    return err;
+    for (auto event : events) {
+        CL_CHECK(clReleaseEvent(event));
+    }
+    return CL_SUCCESS;
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -1393,7 +1405,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
     size_t d_size;
 
     cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
-    cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted.
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
     cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
 
 
@@ -1476,10 +1488,15 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
 
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
     const float alpha = 1.0f;
     const float beta = 0.0f;
     const int x_ne = ne01 * ne00;
@@ -1491,20 +1508,32 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
+        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
+    size_t x_offset = 0;
+    int64_t pi02 = -1;
+    int64_t pi03 = -1;
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        int64_t i03 = i13 / r3;
+
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            int64_t i02 = i12 / r2;
+
             // copy data to device
-            if (src0->backend != GGML_BACKEND_GPU) {
+            if (src0->backend == GGML_BACKEND_GPU) {
+                x_offset = (i03 * ne02 + i02) * x_ne;
+            } else if (i02 != pi02 || i03 != pi03) {
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
+                pi02 = i02;
+                pi03 = i03;
             }
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
 
             CL_CHECK(clFinish(queue));
 
@@ -1514,7 +1543,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                                                        clblast::Transpose::kYes, clblast::Transpose::kNo,
                                                        ne01, ne11, ne10,
                                                        alpha,
-                                                       d_X, 0, ne00,
+                                                       d_X, x_offset, ne00,
                                                        d_Y, 0, ne10,
                                                        beta,
                                                        d_D, 0, ne01,
@@ -1525,7 +1554,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
             }
 
             // copy dst to host
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
             CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
         }
     }
@@ -1547,6 +1576,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb10 = src1->nb[0];
     const int nb11 = src1->nb[1];
@@ -1556,6 +1587,9 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
 
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
     const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
     const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
     const int x_ne = ne01 * ne00;
@@ -1567,7 +1601,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t d_size;
     cl_mem d_X;
     if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->data;
+        d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
     }
@@ -1577,32 +1611,44 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     bool src1_cont_rows = nb10 == sizeof(float);
     bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
 
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
+    size_t x_offset = 0;
+    int64_t pi02 = -1;
+    int64_t pi03 = -1;
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        int64_t i03 = i13 / r3;
+
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            int64_t i02 = i12 / r2;
+
             // copy src0 to device
-            if (src0->backend != GGML_BACKEND_GPU) {
+            if (src0->backend == GGML_BACKEND_GPU) {
+                x_offset = (i03 * ne02 + i02) * x_ne;
+            } else if (i02 != pi02 || i03 != pi03) {
                 CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
+                pi02 = i02;
+                pi03 = i03;
             }
 
             // convert src1 to fp16
             // TODO: use multiple threads
-            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
-            char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
+            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i13 * ne12 + i12);
+            char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
             if (src1_cont_rows) {
                 if (src1_cont_cols) {
                     ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
                 }
                 else {
-                    for (int64_t i01 = 0; i01 < ne11; i01++) {
-                        ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
+                    for (int64_t i11 = 0; i11 < ne11; i11++) {
+                        ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
                     }
                 }
             }
             else {
-                for (int64_t i01 = 0; i01 < ne11; i01++) {
-                    for (int64_t i00 = 0; i00 < ne10; i00++) {
+                for (int64_t i11 = 0; i11 < ne11; i11++) {
+                    for (int64_t i10 = 0; i10 < ne10; i10++) {
                         // very slow due to no inlining
-                        tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
+                        tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
                     }
                 }
             }
@@ -1618,7 +1664,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                                                        clblast::Transpose::kYes, clblast::Transpose::kNo,
                                                        ne01, ne11, ne10,
                                                        alpha,
-                                                       d_X, 0, ne00,
+                                                       d_X, x_offset, ne00,
                                                        d_Y, 0, ne10,
                                                        beta,
                                                        d_D, 0, ne01,
@@ -1631,7 +1677,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
             // copy dst to host, then convert to float
             CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
 
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
 
             ggml_fp16_to_fp32_row(tmp, d, d_ne);
         }
@@ -1652,18 +1698,24 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
     const int64_t ne10 = src1->ne[0];
     const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
     const ggml_type type = src0->type;
     const bool mul_mat_vec = ne11 == 1;
 
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
     const float alpha = 1.0f;
     const float beta = 0.0f;
     const int x_ne = ne01 * ne00;
     const int y_ne = ne11 * ne10;
     const int d_ne = ne11 * ne01;
-    const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
+    const int x_bps = x_ne / ggml_blck_size(type); // blocks per 2D slice
+    const size_t q_sz = ggml_type_size(type) * x_bps;
 
     size_t x_size;
     size_t y_size;
@@ -1690,21 +1742,32 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     size_t ev_idx = 0;
     std::vector<cl_event> events;
 
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
+    int64_t pi02 = -1;
+    int64_t pi03 = -1;
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        int64_t i03 = i13 / r3;
+
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            int64_t i02 = i12 / r2;
+
             // copy src0 to device if necessary
             if (src0->backend == GGML_BACKEND_CPU) {
-                events.emplace_back();
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
+                if (i02 != pi02 || i03 != pi03) {
+                    events.emplace_back();
+                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
+                    pi02 = i02;
+                    pi03 = i03;
+                }
             } else if (src0->backend == GGML_BACKEND_GPU) {
-                d_Q = (cl_mem) src0->data;
+                d_Q = (cl_mem) src0->extra;
             } else {
                 GGML_ASSERT(false);
             }
             if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
                 // copy src1 to device
                 events.emplace_back();
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
+                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
 
                 // compute
                 const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
@@ -1720,12 +1783,13 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
             } else { // general dequantization kernel + CLBlast matrix matrix multiplication
                 // convert src0 to fp32 on device
                 const size_t global = x_ne / global_denom;
+                const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                 CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
+                CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, offset > 0 ? &offset : NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
 
                 // copy src1 to device
-                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
+                CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
 
                 events.emplace_back();
 
@@ -1749,7 +1813,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
             }
 
             // copy dst to host
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
             CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
             for (auto *event : events) {
                 clReleaseEvent(event);
@@ -1788,7 +1852,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     return false;
 }
 
-bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
     // If device doesn't support FP16
     if (!fp16_support) {
         return false;
@@ -1844,22 +1908,24 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     const int64_t ne3 = tensor->ne[3];
 
     const ggml_type type = tensor->type;
-    const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
+    const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
+    const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
 
     size_t q_size;
     cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
 
     tensor->data = data;
     // copy tensor to device
+    size_t offset = 0;
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
-            int i = i3*ne2 + i2;
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, i*ne0*ne1, tensor, i3, i2, NULL));
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, offset, tensor, i3, i2, NULL));
+            offset += s_sz;
         }
     }
 
     CL_CHECK(clFinish(queue));
 
-    tensor->data = dst;
+    tensor->extra = dst;
     GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
diff --git a/ggml.c b/ggml.c
index dffb97731..911a63988 100644
--- a/ggml.c
+++ b/ggml.c
@@ -1,4 +1,3 @@
-#define _GNU_SOURCE // Defines CLOCK_MONOTONIC on Linux
 #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
 
 #include "ggml.h"
@@ -47,6 +46,10 @@
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
 #pragma warning(disable: 4244 4267)
+
+// disable POSIX deprecation warnigns
+// these functions are never going away, anyway
+#pragma warning(disable: 4996)
 #endif
 
 #if defined(_WIN32)
@@ -86,7 +89,9 @@ static int pthread_create(pthread_t * out, void * unused, thread_ret_t(*func)(vo
 
 static int pthread_join(pthread_t thread, void * unused) {
     (void) unused;
-    return (int) WaitForSingleObject(thread, INFINITE);
+    int ret = (int) WaitForSingleObject(thread, INFINITE);
+    CloseHandle(thread);
+    return ret;
 }
 
 static int sched_yield (void) {
@@ -103,6 +108,9 @@ typedef void * thread_ret_t;
 #include <sys/stat.h>
 #include <unistd.h>
 
+#endif
+#ifdef GGML_USE_CPU_HBM
+#include <hbwmalloc.h>
 #endif
 
 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
@@ -123,9 +131,12 @@ typedef void * thread_ret_t;
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
 #define GGML_SILU_FP16
+// #define GGML_CROSS_ENTROPY_EXP_FP16
+// #define GGML_FLASH_ATTN_EXP_FP16
 
 #define GGML_SOFT_MAX_UNROLL 4
 #define GGML_VEC_DOT_UNROLL  2
+#define GGML_VEC_MAD_UNROLL  32
 
 //
 // logging
@@ -157,12 +168,6 @@ typedef void * thread_ret_t;
 //#define GGML_SOFT_MAX_ACCELERATE
 #endif
 
-#if UINTPTR_MAX == 0xFFFFFFFF
-    #define GGML_MEM_ALIGN 4
-#else
-    #define GGML_MEM_ALIGN 16
-#endif
-
 //
 // logging
 //
@@ -192,13 +197,19 @@ typedef void * thread_ret_t;
 //
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
-#define GGML_ALIGNED_MALLOC(size)  _aligned_malloc(size, GGML_MEM_ALIGN)
-#define GGML_ALIGNED_FREE(ptr)     _aligned_free(ptr)
+#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
+#define GGML_ALIGNED_FREE(ptr)    _aligned_free(ptr)
 #else
 inline static void * ggml_aligned_malloc(size_t size) {
+    if (size == 0) {
+        GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
+        return NULL;
+    }
     void * aligned_memory = NULL;
-#ifdef GGML_USE_METAL
-    int result = posix_memalign(&aligned_memory, getpagesize(), size);
+#ifdef GGML_USE_CPU_HBM
+    int result = hbw_posix_memalign(&aligned_memory, 16, size);
+#elif GGML_USE_METAL
+    int result = posix_memalign(&aligned_memory, sysconf(_SC_PAGESIZE), size);
 #else
     int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
 #endif
@@ -218,8 +229,12 @@ inline static void * ggml_aligned_malloc(size_t size) {
     }
     return aligned_memory;
 }
-#define GGML_ALIGNED_MALLOC(size)  ggml_aligned_malloc(size)
-#define GGML_ALIGNED_FREE(ptr)     free(ptr)
+#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
+#ifdef GGML_USE_CPU_HBM
+#define GGML_ALIGNED_FREE(ptr)    if(NULL != ptr) hbw_free(ptr)
+#else
+#define GGML_ALIGNED_FREE(ptr)    free(ptr)
+#endif
 #endif
 
 #define UNUSED GGML_UNUSED
@@ -230,18 +245,18 @@ inline static void * ggml_aligned_malloc(size_t size) {
 //
 
 #define GGML_TENSOR_UNARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb); \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne); \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
 #define GGML_TENSOR_BINARY_OP_LOCALS \
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); \
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb); \
-    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne); \
-    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb); \
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne); \
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
@@ -271,7 +286,7 @@ typedef double ggml_float;
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
-#ifdef __ARM_NEON
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
 
 // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
 //
@@ -298,12 +313,18 @@ typedef double ggml_float;
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>
 #else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
 #if !defined(__riscv)
 #include <immintrin.h>
 #endif
 #endif
 #endif
 #endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
 
 #ifdef __F16C__
 
@@ -817,46 +838,6 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 
 #if !defined(__aarch64__)
 
-inline static uint16_t vaddvq_u8(uint8x16_t v) {
-    return
-        (uint16_t)vgetq_lane_u8(v, 0)  + (uint16_t)vgetq_lane_u8(v, 1)  +
-        (uint16_t)vgetq_lane_u8(v, 2)  + (uint16_t)vgetq_lane_u8(v, 3)  +
-        (uint16_t)vgetq_lane_u8(v, 4)  + (uint16_t)vgetq_lane_u8(v, 5)  +
-        (uint16_t)vgetq_lane_u8(v, 6)  + (uint16_t)vgetq_lane_u8(v, 7)  +
-        (uint16_t)vgetq_lane_u8(v, 8)  + (uint16_t)vgetq_lane_u8(v, 9)  +
-        (uint16_t)vgetq_lane_u8(v, 10) + (uint16_t)vgetq_lane_u8(v, 11) +
-        (uint16_t)vgetq_lane_u8(v, 12) + (uint16_t)vgetq_lane_u8(v, 13) +
-        (uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15);
-}
-
-inline static int16_t vaddvq_s8(int8x16_t v) {
-    return
-        (int16_t)vgetq_lane_s8(v, 0)  + (int16_t)vgetq_lane_s8(v, 1)  +
-        (int16_t)vgetq_lane_s8(v, 2)  + (int16_t)vgetq_lane_s8(v, 3)  +
-        (int16_t)vgetq_lane_s8(v, 4)  + (int16_t)vgetq_lane_s8(v, 5)  +
-        (int16_t)vgetq_lane_s8(v, 6)  + (int16_t)vgetq_lane_s8(v, 7)  +
-        (int16_t)vgetq_lane_s8(v, 8)  + (int16_t)vgetq_lane_s8(v, 9)  +
-        (int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) +
-        (int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) +
-        (int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15);
-}
-
-inline static int32_t vaddvq_s16(int16x8_t v) {
-    return
-        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
-        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
-        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
-        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
-}
-
-inline static uint32_t vaddvq_u16(uint16x8_t v) {
-    return
-        (uint32_t)vgetq_lane_u16(v, 0) + (uint32_t)vgetq_lane_u16(v, 1) +
-        (uint32_t)vgetq_lane_u16(v, 2) + (uint32_t)vgetq_lane_u16(v, 3) +
-        (uint32_t)vgetq_lane_u16(v, 4) + (uint32_t)vgetq_lane_u16(v, 5) +
-        (uint32_t)vgetq_lane_u16(v, 6) + (uint32_t)vgetq_lane_u16(v, 7);
-}
-
 inline static int32_t vaddvq_s32(int32x4_t v) {
     return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
 }
@@ -865,12 +846,6 @@ inline static float vaddvq_f32(float32x4_t v) {
     return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
 }
 
-inline static float vminvq_f32(float32x4_t v) {
-    return
-        MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
 inline static float vmaxvq_f32(float32x4_t v) {
     return
         MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
@@ -1057,8 +1032,8 @@ static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * r
             y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
 
             // get the 5-th bit and store it in qh at the right position
-            qh |= ((xi0 & 0x10) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10) >> 4) << (j + qk/2);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
         }
 
         memcpy(&y[i].qh, &qh, sizeof(qh));
@@ -1105,8 +1080,8 @@ static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * r
             y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
 
             // get the 5-th bit and store it in qh at the right position
-            qh |= ((xi0 & 0x10) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10) >> 4) << (j + qk/2);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
         }
 
         memcpy(&y[i].qh, &qh, sizeof(y[i].qh));
@@ -1297,6 +1272,33 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
 #endif
     }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_0);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+    }
 #else
     // scalar
     quantize_row_q8_0_reference(x, y, k);
@@ -1515,6 +1517,41 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int
         _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
 #endif
     }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_1);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d  = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+
+        // compute sum for y[i].s
+        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
+        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
+
+        // set y[i].s
+        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
+        y[i].s = sum*d;
+    }
 #else
     // scalar
     quantize_row_q8_1_reference(x, y, k);
@@ -1891,7 +1928,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
     #define GGML_F16x8_ADD          vaddq_f16
     #define GGML_F16x8_MUL          vmulq_f16
     #define GGML_F16x8_REDUCE(res, x)                             \
-    {                                                             \
+    do {                                                          \
         int offset = GGML_F16_ARR >> 1;                           \
         for (int i = 0; i < offset; ++i) {                        \
             x[i] = vaddq_f16(x[i], x[offset+i]);                  \
@@ -1907,7 +1944,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
         const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
         const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
         res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
-    }
+    } while (0)
 
     #define GGML_F16_VEC                GGML_F16x8
     #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
@@ -1968,7 +2005,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
 #define GGML_F32x8_ADD     _mm256_add_ps
 #define GGML_F32x8_MUL     _mm256_mul_ps
 #define GGML_F32x8_REDUCE(res, x)                                 \
-{                                                                 \
+do {                                                              \
     int offset = GGML_F32_ARR >> 1;                               \
     for (int i = 0; i < offset; ++i) {                            \
         x[i] = _mm256_add_ps(x[i], x[offset+i]);                  \
@@ -1985,7 +2022,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
                                  _mm256_extractf128_ps(x[0], 1)); \
     const __m128 t1 = _mm_hadd_ps(t0, t0);                        \
     res = _mm_cvtss_f32(_mm_hadd_ps(t1, t1));                     \
-}
+} while (0)
 // TODO: is this optimal ?
 
 #define GGML_F32_VEC        GGML_F32x8
@@ -2436,7 +2473,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     const int nb = n / qk;
 
     assert(n % qk == 0);
-    assert(nb % 2 == 0);
 
     const block_q4_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
@@ -2445,6 +2481,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 0; i < nb; i += 2) {
         const block_q4_0 * restrict x0 = &x[i + 0];
         const block_q4_0 * restrict x1 = &x[i + 1];
@@ -2623,6 +2660,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     }
 
     // Main loop
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 2; i < nb; i+=2) {
         _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0);
         _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
@@ -2680,6 +2718,43 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#elif defined(__riscv_v_intrinsic)
+    float sumf = 0.0;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        // subtract offset
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
+    }
+
+    *s = sumf;
 #else
     // scalar
     float sumf = 0.0;
@@ -2706,7 +2781,6 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
     const int nb = n / qk;
 
     assert(n % qk == 0);
-    assert(nb % 2 == 0);
 
     const block_q4_1 * restrict x = vx;
     const block_q8_1 * restrict y = vy;
@@ -2718,6 +2792,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
 
     float summs = 0;
 
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 0; i < nb; i += 2) {
         const block_q4_1 * restrict x0 = &x[i + 0];
         const block_q4_1 * restrict x1 = &x[i + 1];
@@ -2806,6 +2881,39 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
     }
 
     *s = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    float sumf = 0.0;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+    }
+
+    *s = sumf;
 #else
     // scalar
     float sumf = 0.0;
@@ -2832,7 +2940,6 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     const int nb = n / qk;
 
     assert(n % qk == 0);
-    assert(nb % 2 == 0);
     assert(qk == QK5_0);
 
     const block_q5_0 * restrict x = vx;
@@ -2848,6 +2955,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     uint64_t tmp0[4];
     uint64_t tmp1[4];
 
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 0; i < nb; i += 2) {
         const block_q5_0 * restrict x0 = &x[i];
         const block_q5_0 * restrict x1 = &x[i + 1];
@@ -3040,6 +3148,71 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = hsum_float_8(acc);
+#elif defined(__riscv_v_intrinsic)
+    float sumf = 0.0;
+
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // These tempory registers are for masking and shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
+
+    vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
+    vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (int i = 0; i < nb; i++) {
+        memcpy(&qh, x[i].qh, sizeof(uint32_t));
+
+        // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+
+        // ((qh & (1u << (j + 16))) >> (j + 12));
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
+        vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
+    }
+
+    *s = sumf;
 #else
     // scalar
     float sumf = 0.0;
@@ -3072,7 +3245,6 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
     const int nb = n / qk;
 
     assert(n % qk == 0);
-    assert(nb % 2 == 0);
     assert(qk == QK5_1);
 
     const block_q5_1 * restrict x = vx;
@@ -3091,6 +3263,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
     uint64_t tmp0[4];
     uint64_t tmp1[4];
 
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 0; i < nb; i += 2) {
         const block_q5_1 * restrict x0 = &x[i];
         const block_q5_1 * restrict x1 = &x[i + 1];
@@ -3296,6 +3469,68 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
     }
 
     *s = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    float sumf = 0.0;
+
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // temporary registers for shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (int i = 0; i < nb; i++) {
+        memcpy(&qh, x[i].qh, sizeof(uint32_t));
+
+        // load qh
+        vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
+
+        // ((qh >> (j +  0)) << 4) & 0x10;
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
+
+        // ((qh >> (j + 12))     ) & 0x10;
+        vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+    }
+
+    *s = sumf;
 #else
     // scalar
     float sumf = 0.0;
@@ -3328,7 +3563,6 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
     const int nb = n / qk;
 
     assert(n % qk == 0);
-    assert(nb % 2 == 0);
 
     const block_q8_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
@@ -3337,6 +3571,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb
     for (int i = 0; i < nb; i += 2) {
         const block_q8_0 * restrict x0 = &x[i + 0];
         const block_q8_0 * restrict x1 = &x[i + 1];
@@ -3407,6 +3642,26 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = hsum_float_8(acc);
+#elif defined(__riscv_v_intrinsic)
+    float sumf = 0.0;
+    size_t vl = __riscv_vsetvl_e8m1(qk);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl);
+        vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl);
+
+        vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl);
+
+        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
+    }
+
+    *s = sumf;
 #else
     // scalar
     float sumf = 0.0;
@@ -3511,6 +3766,58 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float
 #endif
 }
 
+// xs and vs are byte strides of x and v
+inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) {
+
+    const float * restrict x[GGML_VEC_MAD_UNROLL];
+    const float * restrict v[GGML_VEC_MAD_UNROLL];
+
+    for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) {
+        x[i] = (const float *) ((const char *) xv + i*xs);
+        v[i] = (const float *) ((const char *) vv + i*vs);
+    }
+
+#if defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F32_STEP - 1));
+
+    GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
+
+    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+        vx[k] = GGML_F32_VEC_SET1(v[k][0]);
+    }
+
+    GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
+    GGML_F32_VEC ay[GGML_F32_ARR];
+
+    for (int i = 0; i < np; i += GGML_F32_STEP) {
+        for (int j = 0; j < GGML_F32_ARR; j++) {
+            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+            for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+                ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
+            }
+
+            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+        }
+    }
+
+    // leftovers
+    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+        for (int i = np; i < n; ++i) {
+            y[i] += x[k][i]*v[k][0];
+        }
+    }
+#else
+    // scalar
+    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+        for (int i = 0; i < n; ++i) {
+            y[i] += x[k][i]*v[k][0];
+        }
+    }
+#endif
+}
+
 //inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) { for (int i = 0; i < n; ++i) y[i] *= v;          }
 inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
@@ -3554,9 +3861,9 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) {
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 
-static const float GELU_COEF_A    = 0.044715f;
-static const float GELU_QUICK_COEF    = -1.702f;
-static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+static const float GELU_COEF_A     = 0.044715f;
+static const float GELU_QUICK_COEF = -1.702f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
 
 inline static float ggml_gelu_f32(float x) {
     return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
@@ -3774,12 +4081,16 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ALIBI",
     "CLAMP",
     "CONV_1D",
+    "CONV_TRANSPOSE_1D",
     "CONV_2D",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
     "UPSCALE",
 
+    "CONV_1D_STAGE_0",
+    "CONV_1D_STAGE_1",
+
     "FLASH_ATTN",
     "FLASH_FF",
     "FLASH_ATTN_BACK",
@@ -3805,7 +4116,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3856,12 +4167,16 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "alibi(x)",
     "clamp(x)",
     "conv_1d(x)",
+    "conv_transpose_1d(x)",
     "conv_2d(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
     "upscale(x)",
 
+    "conv_1d_stage_0(x)",
+    "conv_1d_stage_1(x)",
+
     "flash_attn(x)",
     "flash_ff(x)",
     "flash_attn_back(x)",
@@ -3887,7 +4202,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68");
+static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -3916,7 +4231,10 @@ static void ggml_setup_op_has_task_pass(void) {
         p[GGML_OP_DIAG_MASK_INF          ] = true;
         p[GGML_OP_DIAG_MASK_ZERO         ] = true;
         p[GGML_OP_CONV_1D                ] = true;
+        p[GGML_OP_CONV_1D_STAGE_0        ] = true;
+        p[GGML_OP_CONV_1D_STAGE_1        ] = true;
         p[GGML_OP_CONV_2D                ] = true;
+        p[GGML_OP_CONV_TRANSPOSE_1D      ] = true;
         p[GGML_OP_CONV_TRANSPOSE_2D      ] = true;
         p[GGML_OP_FLASH_ATTN_BACK        ] = true;
         p[GGML_OP_CROSS_ENTROPY_LOSS     ] = true;
@@ -4107,16 +4425,22 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
 }
 
 size_t ggml_nbytes(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+    size_t nbytes;
+    size_t blck_size = ggml_blck_size(tensor->type);
+    if (blck_size == 1) {
+        nbytes = ggml_type_size(tensor->type);
+        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+        }
+    }
+    else {
+        nbytes = tensor->ne[0]*tensor->nb[0]/blck_size;
+        for (int i = 1; i < GGML_MAX_DIMS; ++i) {
+            nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
+        }
+    }
 
-    // this should handle cases where the tensor is not contiguous in memory
-    // probaby just:
-    //
-    //     return tensor->ne[3]*tensor->nb[3]
-    //
-    // is enough, but just in case, adding the second part
-
-    return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type));
+    return nbytes;
 }
 
 size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
@@ -4190,10 +4514,9 @@ static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct
 static inline bool ggml_can_out_prod(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
-    return
-        (t0->ne[1] == t1->ne[1])  &&
-        (t0->ne[2] == t1->ne[2])  &&
-        (t0->ne[3] == t1->ne[3]);
+    return (t0->ne[1] == t1->ne[1])   &&
+           (t1->ne[2]%t0->ne[2] == 0) && // verify t0 is broadcastable
+           (t1->ne[3]%t0->ne[3] == 0);
 }
 
 enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
@@ -4393,6 +4716,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         return NULL;
     }
 
+    // allow to call ggml_init with 0 size
+    if (params.mem_size == 0) {
+        params.mem_size = GGML_MEM_ALIGN;
+    }
+
     const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);
 
     *ctx = (struct ggml_context) {
@@ -4570,36 +4898,51 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne,
-        void                * data) {
+        struct ggml_tensor  * view_src,
+        size_t                view_offs) {
 
     assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
-    size_t data_size = 0;
+    // find the base tensor and absolute offset
+    if (view_src != NULL && view_src->view_src != NULL) {
+        view_offs += view_src->view_offs;
+        view_src   = view_src->view_src;
+    }
 
-    if (data == NULL && !ctx->no_alloc) {
-        data_size += ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
-        for (int i = 1; i < n_dims; i++) {
-            data_size *= ne[i];
+    size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type));
+    for (int i = 1; i < n_dims; i++) {
+        data_size *= ne[i];
+    }
+
+    GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src));
+
+    void * data = view_src != NULL ? view_src->data : NULL;
+    if (data != NULL) {
+        data = (char *) data + view_offs;
+    }
+
+    size_t obj_alloc_size = 0;
+
+    if (view_src == NULL && !ctx->no_alloc) {
+        if (ctx->scratch.data != NULL) {
+            // allocate tensor data in the scratch buffer
+            if (ctx->scratch.offs + data_size > ctx->scratch.size) {
+                GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
+                        __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
+                assert(false);
+                return NULL;
+            }
+
+            data = (char * const) ctx->scratch.data + ctx->scratch.offs;
+
+            ctx->scratch.offs += data_size;
+        } else {
+            // allocate tensor data in the context's memory pool
+            obj_alloc_size = data_size;
         }
     }
 
-    if (ctx->scratch.data != NULL && data == NULL) {
-        // allocate tensor data in the scratch buffer
-        if (ctx->scratch.offs + data_size > ctx->scratch.size) {
-            GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
-                    __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
-            assert(false);
-            return NULL;
-        }
-
-        data = (char * const) ctx->scratch.data + ctx->scratch.offs;
-
-        ctx->scratch.offs += data_size;
-
-        data_size = 0;
-    }
-
-    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size);
+    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
 
     // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here
 
@@ -4619,7 +4962,9 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
-        /*.data         =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data,
+        /*.view_src     =*/ view_src,
+        /*.view_offs    =*/ view_offs,
+        /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
         /*.padding      =*/ { 0 },
@@ -4643,28 +4988,12 @@ static struct ggml_tensor * ggml_new_tensor_impl(
     return result;
 }
 
-static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
-    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
-    assert(params_size <= GGML_MAX_OP_PARAMS);
-    memcpy(tensor->op_params, params, params_size);
-}
-
-static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    return ((const int32_t *)(tensor->op_params))[i];
-}
-
-static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
-    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    ((int32_t *)(tensor->op_params))[i] = value;
-}
-
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type      type,
         int                   n_dims,
         const int64_t       * ne) {
-    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
+    return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL, 0);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
@@ -4729,7 +5058,23 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
 }
 
 struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
-    return ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, NULL);
+    return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
+}
+
+static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
+    GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
+    assert(params_size <= GGML_MAX_OP_PARAMS);
+    memcpy(tensor->op_params, params, params_size);
+}
+
+static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    return ((const int32_t *)(tensor->op_params))[i];
+}
+
+static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
+    ((int32_t *)(tensor->op_params))[i] = value;
 }
 
 struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
@@ -4841,43 +5186,78 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
     return tensor;
 }
 
+void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3) {
+    const int64_t ne2 = tensor->ne[2];
+    const int64_t ne1 = tensor->ne[1];
+    const int64_t ne0 = tensor->ne[0];
+
+    const int64_t i3_ = (i/(ne2*ne1*ne0));
+    const int64_t i2_ = (i - i3_*ne2*ne1*ne0)/(ne1*ne0);
+    const int64_t i1_ = (i - i3_*ne2*ne1*ne0 - i2_*ne1*ne0)/ne0;
+    const int64_t i0_ = (i - i3_*ne2*ne1*ne0 - i2_*ne1*ne0 - i1_*ne0);
+
+    if (i0) {
+        * i0 = i0_;
+    }
+    if (i1) {
+        * i1 = i1_;
+    }
+    if (i2) {
+        * i2 = i2_;
+    }
+    if (i3) {
+        * i3 = i3_;
+    }
+}
+
 int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = { 0, 0, 0, 0 };
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        return ggml_get_i32_nd(tensor, id[0], id[1], id[2], id[3]);
+    }
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 return ((int8_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_I16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 return ((int16_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_I32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 return ((int32_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
-            } break;
+            }
         case GGML_TYPE_F32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 return ((float *)(tensor->data))[i];
-            } break;
+            }
         default:
             {
                 GGML_ASSERT(false);
-            } break;
+            }
     }
 
     return 0.0f;
 }
 
 void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = { 0, 0, 0, 0 };
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_i32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
@@ -4911,43 +5291,104 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
     }
 }
 
+int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3) {
+    void * data   = (char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3];
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            return ((int8_t *) data)[0];
+        case GGML_TYPE_I16:
+            return ((int16_t *) data)[0];
+        case GGML_TYPE_I32:
+            return ((int32_t *) data)[0];
+        case GGML_TYPE_F16:
+            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+        case GGML_TYPE_F32:
+            return ((float *) data)[0];
+        default:
+            GGML_ASSERT(false);
+    }
+
+    return 0.0f;
+}
+
+void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value) {
+    void * data   = (char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3];
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            {
+                ((int8_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_I16:
+            {
+                ((int16_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_I32:
+            {
+                ((int32_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ((float *)(data))[0] = value;
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = { 0, 0, 0, 0 };
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        return ggml_get_f32_nd(tensor, id[0], id[1], id[2], id[3]);
+    }
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 return ((int8_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_I16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 return ((int16_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_I32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 return ((int32_t *)(tensor->data))[i];
-            } break;
+            }
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
-            } break;
+            }
         case GGML_TYPE_F32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 return ((float *)(tensor->data))[i];
-            } break;
+            }
         default:
             {
                 GGML_ASSERT(false);
-            } break;
+            }
     }
 
     return 0.0f;
 }
 
 void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = { 0, 0, 0, 0 };
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
@@ -4981,6 +5422,56 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
     }
 }
 
+float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3) {
+    void * data   = (char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3];
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            return ((int8_t *) data)[0];
+        case GGML_TYPE_I16:
+            return ((int16_t *) data)[0];
+        case GGML_TYPE_I32:
+            return ((int32_t *) data)[0];
+        case GGML_TYPE_F16:
+            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+        case GGML_TYPE_F32:
+            return ((float *) data)[0];
+        default:
+            GGML_ASSERT(false);
+    }
+
+    return 0.0f;
+}
+
+void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value) {
+    void * data   = (char *) tensor->data + i0*tensor->nb[0] + i1*tensor->nb[1] + i2*tensor->nb[2] + i3*tensor->nb[3];
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            {
+                ((int8_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_I16:
+            {
+                ((int16_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_I32:
+            {
+                ((int32_t *)(data))[0] = value;
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ((float *)(data))[0] = value;
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 void * ggml_get_data(const struct ggml_tensor * tensor) {
     return tensor->data;
 }
@@ -5015,14 +5506,13 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
 
 struct ggml_tensor * ggml_view_tensor(
         struct ggml_context * ctx,
-        const struct ggml_tensor * src) {
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data);
+        struct ggml_tensor  * src) {
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
     ggml_format_name(result, "%s (view)", src->name);
 
-    result->nb[0] = src->nb[0];
-    result->nb[1] = src->nb[1];
-    result->nb[2] = src->nb[2];
-    result->nb[3] = src->nb[3];
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = src->nb[i];
+    }
 
     return result;
 }
@@ -5124,6 +5614,44 @@ struct ggml_tensor * ggml_add_inplace(
     return ggml_add_impl(ctx, a, b, true);
 }
 
+// ggml_add_cast
+
+static struct ggml_tensor * ggml_add_cast_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor * a,
+        struct ggml_tensor * b,
+        enum   ggml_type     type) {
+    // TODO: support less-strict constraint
+    //       GGML_ASSERT(ggml_can_repeat(b, a));
+    GGML_ASSERT(ggml_can_repeat_rows(b, a));
+    GGML_ASSERT(ggml_is_quantized(a->type)); // currently only supported for quantized input
+
+    bool is_node = false;
+
+    if (a->grad || b->grad) {
+        // TODO: support backward pass for broadcasting
+        GGML_ASSERT(ggml_are_same_shape(a, b));
+        is_node = true;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, type, a->n_dims, a->ne);
+
+    result->op   = GGML_OP_ADD;
+    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_add_cast(
+        struct ggml_context * ctx,
+        struct ggml_tensor * a,
+        struct ggml_tensor * b,
+        enum   ggml_type     type) {
+    return ggml_add_cast_impl(ctx, a, b, type);
+}
+
 // ggml_add1
 
 static struct ggml_tensor * ggml_add1_impl(
@@ -5280,7 +5808,7 @@ static struct ggml_tensor * ggml_mul_impl(
     }
 
     if (inplace) {
-        GGML_ASSERT(is_node == false);
+        GGML_ASSERT(!is_node);
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@@ -5323,7 +5851,7 @@ static struct ggml_tensor * ggml_div_impl(
     }
 
     if (inplace) {
-        GGML_ASSERT(is_node == false);
+        GGML_ASSERT(!is_node);
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@@ -5555,16 +6083,11 @@ struct ggml_tensor * ggml_repeat(
         is_node = true;
     }
 
-    if (ggml_are_same_shape(a, b) && !is_node) {
-        return a;
-    }
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
 
     result->op   = GGML_OP_REPEAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
-    result->src[1] = b;
 
     return result;
 }
@@ -5592,14 +6115,13 @@ struct ggml_tensor * ggml_repeat_back(
     result->op   = GGML_OP_REPEAT_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
-    result->src[1] = b;
 
     return result;
 }
 
 // ggml_concat
 
-struct ggml_tensor* ggml_concat(
+struct ggml_tensor * ggml_concat(
     struct ggml_context* ctx,
     struct ggml_tensor* a,
     struct ggml_tensor* b) {
@@ -5789,6 +6311,7 @@ struct ggml_tensor * ggml_silu_back(
 static struct ggml_tensor * ggml_norm_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
+        float eps,
         bool inplace) {
     bool is_node = false;
 
@@ -5799,7 +6322,7 @@ static struct ggml_tensor * ggml_norm_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    // TODO: maybe store epsilon here?
+    ggml_set_op_params(result, &eps, sizeof(eps));
 
     result->op   = GGML_OP_NORM;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -5810,14 +6333,16 @@ static struct ggml_tensor * ggml_norm_impl(
 
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a) {
-    return ggml_norm_impl(ctx, a, false);
+        struct ggml_tensor  * a,
+        float eps) {
+    return ggml_norm_impl(ctx, a, eps, false);
 }
 
 struct ggml_tensor * ggml_norm_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a) {
-    return ggml_norm_impl(ctx, a, true);
+        struct ggml_tensor  * a,
+        float eps) {
+    return ggml_norm_impl(ctx, a, eps, true);
 }
 
 // ggml_rms_norm
@@ -5863,7 +6388,8 @@ struct ggml_tensor * ggml_rms_norm_inplace(
 struct ggml_tensor * ggml_rms_norm_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b) {
+        struct ggml_tensor  * b,
+        float  eps) {
     bool is_node = false;
 
     if (a->grad) {
@@ -5873,6 +6399,8 @@ struct ggml_tensor * ggml_rms_norm_back(
 
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
+    ggml_set_op_params(result, &eps, sizeof(eps));
+
     result->op   = GGML_OP_RMS_NORM_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
@@ -5961,8 +6489,9 @@ struct ggml_tensor * ggml_out_prod(
         is_node = true;
     }
 
-    const int64_t ne[4] = { a->ne[0], b->ne[0], a->ne[2], b->ne[3] };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MIN(a->n_dims, b->n_dims), ne);
+    // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
+    const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
 
     result->op   = GGML_OP_OUT_PROD;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6181,6 +6710,53 @@ struct ggml_tensor * ggml_cont_inplace(
     return ggml_cont_impl(ctx, a, true);
 }
 
+
+// make contiguous, with new shape
+GGML_API struct ggml_tensor * ggml_cont_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0) {
+    return ggml_cont_4d(ctx, a, ne0, 1, 1, 1);
+}
+
+GGML_API struct ggml_tensor * ggml_cont_2d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1) {
+    return ggml_cont_4d(ctx, a, ne0, ne1, 1, 1);
+}
+
+GGML_API struct ggml_tensor * ggml_cont_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2) {
+    return ggml_cont_4d(ctx, a, ne0, ne1, ne2, 1);
+}
+
+struct ggml_tensor * ggml_cont_4d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        int64_t               ne3) {
+    GGML_ASSERT(ggml_nelements(a) == (ne0*ne1*ne2*ne3));
+
+    bool is_node = false;
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
+    ggml_format_name(result, "%s (cont)", a->name);
+
+    result->op   = GGML_OP_CONT;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_reshape
 
 struct ggml_tensor * ggml_reshape(
@@ -6188,7 +6764,7 @@ struct ggml_tensor * ggml_reshape(
         struct ggml_tensor * a,
         struct ggml_tensor * b) {
     GGML_ASSERT(ggml_is_contiguous(a));
-    GGML_ASSERT(ggml_is_contiguous(b));
+    // as only the shape of b is relevant, and not its memory layout, b is allowed to be non contiguous.
     GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
 
     bool is_node = false;
@@ -6202,7 +6778,7 @@ struct ggml_tensor * ggml_reshape(
         //GGML_ASSERT(false);
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6226,7 +6802,7 @@ struct ggml_tensor * ggml_reshape_1d(
     }
 
     const int64_t ne[1] = { ne0 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6251,7 +6827,7 @@ struct ggml_tensor * ggml_reshape_2d(
     }
 
     const int64_t ne[2] = { ne0, ne1 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6277,7 +6853,7 @@ struct ggml_tensor * ggml_reshape_3d(
     }
 
     const int64_t ne[3] = { ne0, ne1, ne2 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6287,7 +6863,6 @@ struct ggml_tensor * ggml_reshape_3d(
     return result;
 }
 
-
 struct ggml_tensor * ggml_reshape_4d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6305,7 +6880,7 @@ struct ggml_tensor * ggml_reshape_4d(
     }
 
     const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
     result->op   = GGML_OP_RESHAPE;
@@ -6315,34 +6890,12 @@ struct ggml_tensor * ggml_reshape_4d(
     return result;
 }
 
-// ggml_view_1d
-
-static struct ggml_tensor * ggml_view_tensor_offset(
+static struct ggml_tensor * ggml_view_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         int                   n_dims,
         const int64_t       * ne,
         size_t                offset) {
-    // don't calculate an offset from an unallocated tensor
-    void * data = NULL;
-    if (a->data != NULL) {
-        data = (char *) a->data + offset;
-    }
-
-    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data);
-
-    ggml_format_name(result, "%s (view)", a->name);
-
-    ggml_set_op_params(result, &offset, sizeof(offset));
-
-    return result;
-}
-
-struct ggml_tensor * ggml_view_1d(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int64_t               ne0,
-        size_t                offset) {
 
     bool is_node = false;
 
@@ -6350,7 +6903,10 @@ struct ggml_tensor * ggml_view_1d(
         is_node = true;
     }
 
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset);
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset);
+    ggml_format_name(result, "%s (view)", a->name);
+
+    ggml_set_op_params(result, &offset, sizeof(offset));
 
     result->op   = GGML_OP_VIEW;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -6359,6 +6915,19 @@ struct ggml_tensor * ggml_view_1d(
     return result;
 }
 
+// ggml_view_1d
+
+struct ggml_tensor * ggml_view_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        size_t                offset) {
+
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset);
+
+    return result;
+}
+
 // ggml_view_2d
 
 struct ggml_tensor * ggml_view_2d(
@@ -6369,24 +6938,14 @@ struct ggml_tensor * ggml_view_2d(
         size_t                nb1,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[2] = { ne0, ne1 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = result->nb[1]*ne1;
     result->nb[3] = result->nb[2];
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6402,24 +6961,14 @@ struct ggml_tensor * ggml_view_3d(
         size_t                nb2,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[3] = { ne0, ne1, ne2 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = result->nb[2]*ne2;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6437,24 +6986,14 @@ struct ggml_tensor * ggml_view_4d(
         size_t                nb3,
         size_t                offset) {
 
-    bool is_node = false;
+    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
 
-    if (a->grad) {
-        is_node = true;
-    }
-
-    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 };
-
-    struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset);
+    struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset);
 
     result->nb[1] = nb1;
     result->nb[2] = nb2;
     result->nb[3] = nb3;
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
     return result;
 }
 
@@ -6598,7 +7137,6 @@ struct ggml_tensor * ggml_get_rows_back(
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
-    result->src[2] = c;
 
     return result;
 }
@@ -6641,7 +7179,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_INF;
@@ -6658,7 +7196,6 @@ struct ggml_tensor * ggml_diag_mask_inf(
     return ggml_diag_mask_inf_impl(ctx, a, n_past, false);
 }
 
-
 struct ggml_tensor * ggml_diag_mask_inf_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -6681,7 +7218,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[] = { n_past, inplace ? 1 : 0 };
+    int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_DIAG_MASK_ZERO;
@@ -6781,7 +7318,7 @@ struct ggml_tensor * ggml_soft_max_back_inplace(
 static struct ggml_tensor * ggml_rope_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
@@ -6790,7 +7327,10 @@ static struct ggml_tensor * ggml_rope_impl(
         float                 xpos_base,
         bool                  xpos_down,
         bool                  inplace) {
-    GGML_ASSERT(n_past >= 0);
+    GGML_ASSERT(ggml_is_vector(b));
+    GGML_ASSERT(b->type == GGML_TYPE_I32);
+    GGML_ASSERT(a->ne[2] == b->ne[0]);
+
     bool is_node = false;
 
     if (a->grad) {
@@ -6799,7 +7339,7 @@ static struct ggml_tensor * ggml_rope_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[8] = { n_past, n_dims, mode, n_ctx };
+    int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx };
     memcpy(params + 4, &freq_base,  sizeof(float));
     memcpy(params + 5, &freq_scale, sizeof(float));
     memcpy(params + 6, &xpos_base,  sizeof(float));
@@ -6809,6 +7349,7 @@ static struct ggml_tensor * ggml_rope_impl(
     result->op   = GGML_OP_ROPE;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
+    result->src[1] = b;
 
     return result;
 }
@@ -6816,55 +7357,55 @@ static struct ggml_tensor * ggml_rope_impl(
 struct ggml_tensor * ggml_rope(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, false);
+    return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, false);
 }
 
 struct ggml_tensor * ggml_rope_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, true);
+    return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, true);
 }
 
 struct ggml_tensor * ggml_rope_custom(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
         float                 freq_base,
         float                 freq_scale) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, false);
+    return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, false);
 }
 
 struct ggml_tensor * ggml_rope_custom_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
         float                 freq_base,
         float                 freq_scale) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, true);
+    return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, true);
 }
 
 struct ggml_tensor * ggml_rope_xpos_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         float                 base,
         bool                  down) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, 0, 0, 10000.0f, 1.0f, base, down, true);
+    return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 10000.0f, 1.0f, base, down, true);
 }
 
 // ggml_rope_back
@@ -6872,7 +7413,7 @@ struct ggml_tensor * ggml_rope_xpos_inplace(
 struct ggml_tensor * ggml_rope_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   n_past,
+        struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
@@ -6880,7 +7421,10 @@ struct ggml_tensor * ggml_rope_back(
         float                 freq_scale,
         float                 xpos_base,
         bool                  xpos_down) {
-    GGML_ASSERT(n_past >= 0);
+    GGML_ASSERT(ggml_is_vector(b));
+    GGML_ASSERT(b->type == GGML_TYPE_I32);
+    GGML_ASSERT(a->ne[2] == b->ne[0]);
+
     GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
 
     bool is_node = false;
@@ -6891,7 +7435,7 @@ struct ggml_tensor * ggml_rope_back(
 
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
-    int32_t params[8] = { n_past, n_dims, mode, n_ctx };
+    int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx };
     memcpy(params + 4, &freq_base,  sizeof(float));
     memcpy(params + 5, &freq_scale, sizeof(float));
     memcpy(params + 6, &xpos_base,  sizeof(float));
@@ -6901,6 +7445,7 @@ struct ggml_tensor * ggml_rope_back(
     result->op   = GGML_OP_ROPE_BACK;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
+    result->src[1] = b;
 
     return result;
 }
@@ -6969,14 +7514,17 @@ static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p,
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }
 
-GGML_API struct ggml_tensor * ggml_conv_1d(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
-        int                   s0,
-        int                   p0,
-        int                   d0) {
-    GGML_ASSERT(ggml_is_matrix(b));
+// im2col: [N, IC, IL] => [N, OL, IC*K]
+// a: [OC，IC, K]
+// b: [N, IC, IL]
+// result: [N, OL, IC*K]
+static struct ggml_tensor * ggml_conv_1d_stage_0(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b,
+    int                   s0,
+    int                   p0,
+    int                   d0) {
     GGML_ASSERT(a->ne[1] == b->ne[1]);
     bool is_node = false;
 
@@ -6985,16 +7533,20 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
         is_node = true;
     }
 
+    const int64_t OL = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+
     const int64_t ne[4] = {
-        ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
-        a->ne[2], 1, 1,
+        a->ne[1] * a->ne[0],
+        OL,
+        b->ne[2],
+        1,
     };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne);
 
     int32_t params[] = { s0, p0, d0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_CONV_1D;
+    result->op = GGML_OP_CONV_1D_STAGE_0;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
@@ -7002,6 +7554,87 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
     return result;
 }
 
+// ggml_conv_1d_stage_1
+
+// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
+// a: [OC, IC, K]
+// b: [N, OL, IC * K]
+// result: [N, OC, OL]
+static struct ggml_tensor * ggml_conv_1d_stage_1(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b) {
+
+    bool is_node = false;
+
+    if (a->grad || b->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    const int64_t ne[4] = {
+        b->ne[1],
+        a->ne[2],
+        b->ne[2],
+        1,
+    };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    result->op = GGML_OP_CONV_1D_STAGE_1;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+// ggml_conv_1d
+
+GGML_API struct ggml_tensor * ggml_conv_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   p0,
+        int                   d0) {
+    struct ggml_tensor * result = ggml_conv_1d_stage_0(ctx, a, b, s0, p0, d0);
+    result = ggml_conv_1d_stage_1(ctx, a, result);
+    return result;
+}
+
+// GGML_API struct ggml_tensor * ggml_conv_1d(
+//         struct ggml_context * ctx,
+//         struct ggml_tensor  * a,
+//         struct ggml_tensor  * b,
+//         int                   s0,
+//         int                   p0,
+//         int                   d0) {
+//     GGML_ASSERT(ggml_is_matrix(b));
+//     GGML_ASSERT(a->ne[1] == b->ne[1]);
+//     bool is_node = false;
+
+//     if (a->grad || b->grad) {
+//         GGML_ASSERT(false); // TODO: implement backward
+//         is_node = true;
+//     }
+
+//     const int64_t ne[4] = {
+//         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
+//         a->ne[2], 1, 1,
+//     };
+//     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+
+//     int32_t params[] = { s0, p0, d0 };
+//     ggml_set_op_params(result, params, sizeof(params));
+
+//     result->op = GGML_OP_CONV_1D;
+//     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+//     result->src[0] = a;
+//     result->src[1] = b;
+
+//     return result;
+// }
+
 // ggml_conv_1d_ph
 
 struct ggml_tensor* ggml_conv_1d_ph(
@@ -7013,6 +7646,50 @@ struct ggml_tensor* ggml_conv_1d_ph(
     return ggml_conv_1d(ctx, a, b, s, a->ne[0] / 2, d);
 }
 
+// ggml_conv_transpose_1d
+
+static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
+    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
+}
+
+GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   p0,
+        int                   d0) {
+    GGML_ASSERT(ggml_is_matrix(b));
+    GGML_ASSERT(a->ne[2] == b->ne[1]);
+    GGML_ASSERT(a->ne[3] == 1);
+
+    GGML_ASSERT(p0 == 0);
+    GGML_ASSERT(d0 == 1);
+
+    bool is_node = false;
+
+    if (a->grad || b->grad) {
+        GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    const int64_t ne[4] = {
+        ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
+        a->ne[1], b->ne[2], 1,
+    };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    int32_t params[] = { s0, p0, d0 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op = GGML_OP_CONV_TRANSPOSE_1D;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 // ggml_conv_2d
 
 struct ggml_tensor * ggml_conv_2d(
@@ -7098,11 +7775,13 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0(
     };
 
     struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    ggml_set_op_params_i32(result, 0, stride);
+
     result->op = GGML_OP_CONV_TRANSPOSE_2D;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
-    result->src[2] = ggml_new_i32(ctx, stride);
 
     return result;
 }
@@ -7295,27 +7974,30 @@ struct ggml_tensor * ggml_flash_attn_back(
 
     // d shape [D,N,ne2,ne3]
     // q shape [D,N,ne2,ne3]
-    // k shape [D,M,ne2,ne3]
-    // v shape [M,D,ne2,ne3]
+    // k shape [D,M,kvne2,ne3]
+    // v shape [M,D,kvne2,ne3]
 
-    const int64_t   D = q->ne[0];
-    const int64_t   N = q->ne[1];
-    const int64_t   M = k->ne[1];
-    const int64_t ne2 = q->ne[2];
-    const int64_t ne3 = q->ne[3];
+    const int64_t     D = q->ne[0];
+    const int64_t     N = q->ne[1];
+    const int64_t     M = k->ne[1];
+    const int64_t   ne2 = q->ne[2];
+    const int64_t   ne3 = q->ne[3];
+    const int64_t kvne2 = k->ne[2];
 
     GGML_ASSERT(k->ne[0] == D);
     GGML_ASSERT(v->ne[0] == M);
     GGML_ASSERT(v->ne[1] == D);
     GGML_ASSERT(d->ne[0] == D);
     GGML_ASSERT(d->ne[1] == N);
-    GGML_ASSERT(k->ne[2] == ne2);
+    GGML_ASSERT(k->ne[2] == kvne2);
     GGML_ASSERT(k->ne[3] == ne3);
-    GGML_ASSERT(v->ne[2] == ne2);
+    GGML_ASSERT(v->ne[2] == kvne2);
     GGML_ASSERT(v->ne[3] == ne3);
     GGML_ASSERT(d->ne[2] == ne2);
     GGML_ASSERT(d->ne[3] == ne3);
 
+    GGML_ASSERT(ne2 % kvne2 == 0);
+
     bool is_node = false;
 
     if (q->grad || k->grad || v->grad) {
@@ -7325,14 +8007,23 @@ struct ggml_tensor * ggml_flash_attn_back(
     }
 
     // store gradients of q, k and v as continuous tensors concatenated in result.
-    // q shape[D,N,ne2,ne3] ; k shape [D,M,ne2,ne3] ; v shape [M,D,ne2,ne3]
-    // gradq->data = result->data
-    // gradk->data = result->data + nb0*D*N*ne2*ne3
-    // gradv->data = result->data + nb0*D*N*ne2*ne3 + nb0*D*M*ne2*ne3
     // note: v and gradv are actually transposed, i.e. v->ne[0] != D.
-    int64_t ne[4] = {D,M+N+M,ne2,ne3};
+    const int64_t elem_q = ggml_nelements(q);
+    const int64_t elem_k = ggml_nelements(k);
+    const int64_t elem_v = ggml_nelements(v);
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    enum ggml_type result_type = GGML_TYPE_F32;
+    GGML_ASSERT(ggml_blck_size(result_type) == 1);
+    const size_t tsize = ggml_type_size(result_type);
+
+    const size_t offs_q = 0;
+    const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
+    const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
+    const size_t end    = offs_v + GGML_PAD(elem_v * tsize, GGML_MEM_ALIGN);
+
+    const size_t nelements = (end + tsize - 1)/tsize;
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nelements);
 
     int32_t masked_i = masked ? 1 : 0;
     ggml_set_op_params(result, &masked_i, sizeof(masked_i));
@@ -8025,7 +8716,7 @@ static void ggml_compute_forward_dup_f16(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
@@ -8296,7 +8987,7 @@ static void ggml_compute_forward_dup_f32(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
@@ -8577,7 +9268,7 @@ static void ggml_compute_forward_add_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
@@ -8609,8 +9300,6 @@ static void ggml_compute_forward_add_f32(
 #else
             ggml_vec_add_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
 #endif
-                // }
-            // }
         }
     } else {
         // src1 is not contiguous
@@ -8652,7 +9341,7 @@ static void ggml_compute_forward_add_f16_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -8706,7 +9395,7 @@ static void ggml_compute_forward_add_f16_f16(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F16);
@@ -8757,14 +9446,15 @@ static void ggml_compute_forward_add_q_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
 
     const enum ggml_type type = src0->type;
+    const enum ggml_type dtype = dst->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
-    ggml_from_float_t const quantize_row_q = type_traits[type].from_float;
+    ggml_from_float_t const quantize_row_q = type_traits[dtype].from_float;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -8776,7 +9466,6 @@ static void ggml_compute_forward_add_q_f32(
     GGML_ASSERT(nb2 <= nb3);
 
     GGML_ASSERT(ggml_is_quantized(src0->type));
-    GGML_ASSERT(dst->type == src0->type);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
     // rows per thread
@@ -8814,7 +9503,11 @@ static void ggml_compute_forward_add_q_f32(
         // add src1
         ggml_vec_acc_f32(ne00, wdata, src1_row);
         // quantize row to dst
-        quantize_row_q(wdata, dst_row, ne00);
+        if (quantize_row_q != NULL) {
+            quantize_row_q(wdata, dst_row, ne00);
+        } else {
+            memcpy(dst_row, wdata, ne0*nb0);
+        }
     }
 }
 
@@ -8879,7 +9572,7 @@ static void ggml_compute_forward_add1_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
@@ -8934,7 +9627,7 @@ static void ggml_compute_forward_add1_f16_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -8984,7 +9677,7 @@ static void ggml_compute_forward_add1_f16_f16(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F16);
@@ -9034,7 +9727,7 @@ static void ggml_compute_forward_add1_q_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
@@ -9162,8 +9855,8 @@ static void ggml_compute_forward_acc_f32(
     const int nr = ggml_nrows(src1);
     const int nc = src1->ne[0];
 
-    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
 
     // src0 and dst as viewed during acc
     const size_t nb0 = ggml_element_size(src0);
@@ -9252,7 +9945,7 @@ static void ggml_compute_forward_sub_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
@@ -9342,7 +10035,7 @@ static void ggml_compute_forward_mul_f32(
 
     const int64_t nr = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
@@ -9433,7 +10126,7 @@ static void ggml_compute_forward_div_f32(
 
     const int nr  = ggml_nrows(src0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
@@ -9447,6 +10140,8 @@ static void ggml_compute_forward_div_f32(
 
 
 #ifdef GGML_USE_ACCELERATE
+            UNUSED(ggml_vec_div_f32);
+
             vDSP_vdiv(
                     (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
                     (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
@@ -9640,8 +10335,8 @@ static void ggml_compute_forward_sum_f32(
     assert(ggml_is_scalar(dst));
     assert(src0->nb[0] == sizeof(float));
 
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
 
     ggml_float sum     = 0;
     ggml_float row_sum = 0;
@@ -9672,8 +10367,8 @@ static void ggml_compute_forward_sum_f16(
 
     assert(src0->nb[0] == sizeof(ggml_fp16_t));
 
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
 
     float sum = 0;
     float row_sum = 0;
@@ -9726,7 +10421,7 @@ static void ggml_compute_forward_sum_rows_f32(
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(dst->nb[0] == sizeof(float));
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     GGML_ASSERT(ne0 == 1);
     GGML_ASSERT(ne1 == ne01);
@@ -9776,7 +10471,7 @@ static void ggml_compute_forward_mean_f32(
 
     assert(src0->nb[0] == sizeof(float));
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     assert(ne0 == 1);
     assert(ne1 == ne01);
@@ -9876,7 +10571,7 @@ static void ggml_compute_forward_repeat_f32(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
     const int nr0 = (int)(ne0/ne00);
@@ -9908,11 +10603,61 @@ static void ggml_compute_forward_repeat_f32(
     }
 }
 
+static void ggml_compute_forward_repeat_f16(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    GGML_ASSERT(params->ith == 0);
+    GGML_ASSERT(ggml_can_repeat(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    // guaranteed to be an integer due to the check in ggml_can_repeat
+    const int nr0 = (int)(ne0/ne00);
+    const int nr1 = (int)(ne1/ne01);
+    const int nr2 = (int)(ne2/ne02);
+    const int nr3 = (int)(ne3/ne03);
+
+    // TODO: support for transposed / permuted tensors
+    GGML_ASSERT(nb0  == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+
+    // TODO: maybe this is not optimal?
+    for                         (int i3 = 0; i3 < nr3;  i3++) {
+        for                     (int k3 = 0; k3 < ne03; k3++) {
+            for                 (int i2 = 0; i2 < nr2;  i2++) {
+                for             (int k2 = 0; k2 < ne02; k2++) {
+                    for         (int i1 = 0; i1 < nr1;  i1++) {
+                        for     (int k1 = 0; k1 < ne01; k1++) {
+                            for (int i0 = 0; i0 < nr0;  i0++) {
+                                ggml_fp16_t * y = (ggml_fp16_t *) ((char *)  dst->data + (i3*ne03 + k3)*nb3  + (i2*ne02 + k2)*nb2  + (i1*ne01 + k1)*nb1  + (i0*ne00)*nb0);
+                                ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + (          k3)*nb03 + (          k2)*nb02 + (          k1)*nb01);
+                                // ggml_vec_cpy_f16(ne00, y, x)
+                                for (int i = 0; i < ne00; ++i) {
+                                    y[i]  = x[i];
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
 static void ggml_compute_forward_repeat(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
     switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_repeat_f16(params, src0, dst);
+            } break;
         case GGML_TYPE_F32:
             {
                 ggml_compute_forward_repeat_f32(params, src0, dst);
@@ -9937,7 +10682,7 @@ static void ggml_compute_forward_repeat_back_f32(
         return;
     }
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
     const int nr0 = (int)(ne00/ne0);
@@ -10015,7 +10760,7 @@ static void ggml_compute_forward_concat_f32(
 
     const int ith = params->ith;
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     // TODO: support for transposed / permuted tensors
     GGML_ASSERT(nb0  == sizeof(float));
@@ -10617,9 +11362,10 @@ static void ggml_compute_forward_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-    const float eps = 1e-5f; // TODO: make this a parameter
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
 
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -10685,7 +11431,7 @@ static void ggml_compute_forward_rms_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
@@ -10750,9 +11496,10 @@ static void ggml_compute_forward_rms_norm_back_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    const float eps = 1e-6f; // TODO: make this a parameter
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
 
     // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -10924,7 +11671,7 @@ static void ggml_compute_forward_group_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const float eps = 1e-6f; // TODO: make this a parameter
 
@@ -11035,7 +11782,7 @@ static void ggml_compute_forward_mul_mat(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -11072,11 +11819,6 @@ static void ggml_compute_forward_mul_mat(
 
 #if defined(GGML_USE_CLBLAST)
     if (ggml_cl_can_mul_mat(src0, src1, dst)) {
-        // TODO: handle case when src0 is broadcast-able into src1 across 2nd,3rd dimension
-        //       ref: https://github.com/ggerganov/ggml/pull/224
-        GGML_ASSERT(ne02 == ne12);
-        GGML_ASSERT(ne03 == ne13);
-
         if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
             ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
         }
@@ -11250,10 +11992,10 @@ static void ggml_compute_forward_out_prod_f32(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
+    // int64_t t0 = ggml_perf_time_us();
+    // UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -11292,6 +12034,146 @@ static void ggml_compute_forward_out_prod_f32(
         return;
     }
 
+    // dst[:,:,:,:] = 0
+    // for i2,i3:
+    //   for i1:
+    //     for i01:
+    //       for i0:
+    //         dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3]
+
+    // parallelize by last three dimensions
+
+    // total rows in dst
+    const int64_t nr = ne1*ne2*ne3;
+
+    // rows per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    // block-tiling attempt
+    const int64_t blck_0 = MAX(GGML_VEC_MAD_UNROLL, 32);
+    const int64_t blck_1 = 16;
+
+    for (int64_t bir = ir0; bir < ir1; bir += blck_1) {
+        const int64_t bir1 = MIN(bir + blck_1, ir1);
+        for (int64_t bi01 = 0; bi01 < ne01; bi01 += blck_0) {
+            const int64_t bne01 = MIN(bi01 + blck_0, ne01);
+            for (int64_t ir = bir; ir < bir1; ++ir) {
+                // dst indices
+                const int64_t i3 = ir/(ne2*ne1);
+                const int64_t i2 = (ir - i3*ne2*ne1)/ne1;
+                const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+                const int64_t i02 = i2;
+                const int64_t i03 = i3;
+
+                //const int64_t i10 = i1;
+                const int64_t i12 = i2;
+                const int64_t i13 = i3;
+
+#if GGML_VEC_MAD_UNROLL > 2
+                const int64_t bne01_unroll = bne01 - (bne01 % GGML_VEC_MAD_UNROLL);
+                for (int64_t i01 = bi01; i01 < bne01_unroll; i01 += GGML_VEC_MAD_UNROLL) {
+                    const int64_t i11 = i01;
+
+                    float * s0 = (float *) ((char *) src0->data + (          i01*nb01 + i02*nb02 + i03*nb03));
+                    float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
+                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
+
+                    ggml_vec_mad_f32_unroll(ne0, nb01, nb11, d, s0, s1);
+                }
+                for (int64_t i01 = bne01_unroll; i01 < bne01; ++i01) {
+                    const int64_t i11 = i01;
+
+                    float * s0 = (float *) ((char *) src0->data + (          i01*nb01 + i02*nb02 + i03*nb03));
+                    float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
+                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
+
+                    ggml_vec_mad_f32(ne0, d, s0, *s1);
+                }
+#else
+                for (int64_t i01 = bi01; i01 < bne01; ++i01) {
+                    const int64_t i11 = i01;
+
+                    float * s0 = (float *) ((char *) src0->data + (          i01*nb01 + i02*nb02 + i03*nb03));
+                    float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
+                    float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
+
+                    ggml_vec_mad_f32(ne0, d, s0, *s1);
+                }
+#endif
+            }
+        }
+    }
+
+
+    //int64_t t1 = ggml_perf_time_us();
+    //static int64_t acc = 0;
+    //acc += t1 - t0;
+    //if (t1 - t0 > 10) {
+    //    printf("\n");
+    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
+    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
+    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
+    //    printf("nb10 = %5d, nb11 = %5d, nb12 = %5d, nb13 = %5d\n", nb10, nb11, nb12, nb13);
+
+    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
+    //}
+}
+
+static void ggml_compute_forward_out_prod_q_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    // int64_t t0 = ggml_perf_time_us();
+    // UNUSED(t0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const enum ggml_type type = src0->type;
+    ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
+
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne03 == ne13);
+    GGML_ASSERT(ne2  == ne12);
+    GGML_ASSERT(ne3  == ne13);
+
+    // we don't support permuted src0 dim0
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+
+    // dst dim0 cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(ne0 == ne00);
+    GGML_ASSERT(ne1 == ne10);
+    GGML_ASSERT(ne2 == ne02);
+    GGML_ASSERT(ne3 == ne03);
+
+    // nb01 >= nb00 - src0 is not transposed
+    //   compute by src0 rows
+
+    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
+    // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+
+    if (params->type == GGML_TASK_INIT) {
+        ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
     // parallelize by last three dimensions
 
     // total rows in dst
@@ -11311,6 +12193,8 @@ static void ggml_compute_forward_out_prod_f32(
     //       for i0:
     //         dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3]
 
+    float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith;
+
     for (int64_t ir = ir0; ir < ir1; ++ir) {
         // dst indices
         const int64_t i3 = ir/(ne2*ne1);
@@ -11331,10 +12215,8 @@ static void ggml_compute_forward_out_prod_f32(
             float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
             float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
 
-            ggml_vec_mad_f32(ne0, d, s0, *s1);
-            // for (int64_t i0 = 0; i0 < ne0; ++i0) {
-            //     d[i0] += s0[i0] * s1[i1];
-            // }
+            dequantize_row_q(s0, wdata, ne0);
+            ggml_vec_mad_f32(ne0, d, wdata, *s1);
         }
     }
 
@@ -11363,10 +12245,13 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
-        case GGML_TYPE_Q8_1:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
             {
-                GGML_ASSERT(false); // todo
-                // ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
+                ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F16:
             {
@@ -11484,8 +12369,8 @@ static void ggml_compute_forward_set_f32(
     const int nr = ggml_nrows(src1);
     const int nc = src1->ne[0];
 
-    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
 
     // src0 and dst as viewed during set
     const size_t nb0 = ggml_element_size(src0);
@@ -11754,14 +12639,15 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
               struct ggml_tensor * dst) {
     GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_are_same_shape(opt0, dst));
-    GGML_ASSERT(ggml_is_contiguous(opt0));
     GGML_ASSERT(ggml_is_contiguous(dst));
 
-    ggml_compute_forward_dup_same_cont(params, opt0, dst);
+    // ggml_compute_forward_dup_same_cont(params, opt0, dst);
+
+    if (params->type == GGML_TASK_INIT) {
+        memset(dst->data, 0, ggml_nbytes(dst));
+    }
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -11787,11 +12673,8 @@ static void ggml_compute_forward_get_rows_back_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
               struct ggml_tensor * dst) {
     GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_are_same_shape(opt0, dst));
-    GGML_ASSERT(ggml_is_contiguous(opt0));
     GGML_ASSERT(ggml_is_contiguous(dst));
 
     // ggml_compute_forward_dup_same_cont(params, opt0, dst);
@@ -11825,16 +12708,15 @@ static void ggml_compute_forward_get_rows_back(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, opt0, dst);
+                ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_get_rows_back_f32(params, src0, src1, opt0, dst);
+                ggml_compute_forward_get_rows_back_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -11875,7 +12757,7 @@ static void ggml_compute_forward_diag_f32(
 
     // TODO: handle transposed/permuted matrices
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     GGML_ASSERT(ne00 == ne0);
     GGML_ASSERT(ne00 == ne1);
@@ -11930,8 +12812,8 @@ static void ggml_compute_forward_diag_mask_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int  n_past  =       ((int32_t *) dst->op_params)[0];
-    const bool inplace = (bool)((int32_t *) dst->op_params)[1];
+    const int  n_past  = ((int32_t *) dst->op_params)[0];
+    const bool inplace = src0->data == dst->data;
 
     GGML_ASSERT(n_past >= 0);
 
@@ -12142,6 +13024,7 @@ static void ggml_compute_forward_soft_max_back_f32(
         // dx = J * dy
         // dxk = sum_i(Jki * dyi)
         // dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk
+        // dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk
         // dxk = sum_i(-yk*yi * dyi) + yk*dyk
         // dxk = -yk * sum_i(yi * dyi) + yk*dyk
         // dxk = -yk * dot(y, dy) + yk*dyk
@@ -12199,7 +13082,7 @@ static void ggml_compute_forward_alibi_f32(
         return;
     }
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
@@ -12220,7 +13103,6 @@ static void ggml_compute_forward_alibi_f32(
     //const int nb3 = src0->nb[3];
 
     GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(ne1 + n_past == ne0);
     GGML_ASSERT(n_head == ne2);
 
     // add alibi to src0 (KQ_scaled)
@@ -12262,13 +13144,11 @@ static void ggml_compute_forward_alibi_f16(
         return;
     }
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    assert(n_past >= 0);
-
     const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
     const int ne1 = src0->ne[1]; // seq_len_without_past
     const int ne2 = src0->ne[2]; // n_head -> this is k
@@ -12283,7 +13163,7 @@ static void ggml_compute_forward_alibi_f16(
     //const int nb3 = src0->nb[3];
 
     GGML_ASSERT(nb0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(ne1 + n_past == ne0); (void) n_past;
+    //GGML_ASSERT(ne1 + n_past == ne0); (void) n_past;
     GGML_ASSERT(n_head == ne2);
 
     // add alibi to src0 (KQ_scaled)
@@ -12429,8 +13309,8 @@ static void ggml_compute_forward_clamp(
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -12440,9 +13320,9 @@ static void ggml_compute_forward_rope_f32(
 
     // these two only relevant for xPos RoPE:
     float xpos_base;
-    bool xpos_down;
+    bool  xpos_down;
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_dims = ((int32_t *) dst->op_params)[1];
     const int mode   = ((int32_t *) dst->op_params)[2];
     const int n_ctx  = ((int32_t *) dst->op_params)[3];
@@ -12451,9 +13331,7 @@ static void ggml_compute_forward_rope_f32(
     memcpy(&xpos_base,  (int32_t *) dst->op_params + 6, sizeof(float));
     memcpy(&xpos_down,  (int32_t *) dst->op_params + 7, sizeof(bool));
 
-    assert(n_past >= 0);
-
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
     //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
@@ -12483,9 +13361,11 @@ static void ggml_compute_forward_rope_f32(
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    const int32_t * pos = (const int32_t *) src1->data;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t p = pos[i2];
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -12522,7 +13402,7 @@ static void ggml_compute_forward_rope_f32(
                         const float cos_theta = cosf(theta);
                         const float sin_theta = sinf(theta);
                         // zeta scaling for xPos only:
-                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), (n_past + i2) / xpos_base) : 1.0f;
+                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
                         if (xpos_down) zeta = 1.0f / zeta;
 
                         theta *= theta_scale;
@@ -12537,7 +13417,7 @@ static void ggml_compute_forward_rope_f32(
                         dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta;
                     }
                 } else {
-                    // TODO: this is probably wrong, but I can't figure it out ..
+                    // TODO: this might be wrong for ne0 != n_dims - need double check
                     // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
                     for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
                         for (int64_t ic = 0; ic < n_dims; ic += 2) {
@@ -12567,8 +13447,8 @@ static void ggml_compute_forward_rope_f32(
 static void ggml_compute_forward_rope_f16(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
@@ -12576,16 +13456,14 @@ static void ggml_compute_forward_rope_f16(
     float freq_base;
     float freq_scale;
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_dims = ((int32_t *) dst->op_params)[1];
     const int mode   = ((int32_t *) dst->op_params)[2];
     const int n_ctx  = ((int32_t *) dst->op_params)[3];
     memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
     memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
-    assert(n_past >= 0);
-
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
     //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
@@ -12615,9 +13493,11 @@ static void ggml_compute_forward_rope_f16(
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    const int32_t * pos = (const int32_t *) src1->data;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t p = pos[i2];
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -12666,7 +13546,7 @@ static void ggml_compute_forward_rope_f16(
                         dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 } else {
-                    // TODO: this is probably wrong, but I can't figure it out ..
+                    // TODO: this might be wrong for ne0 != n_dims - need double check
                     // ref:  https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28
                     for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
                         for (int64_t ic = 0; ic < n_dims; ic += 2) {
@@ -12696,15 +13576,16 @@ static void ggml_compute_forward_rope_f16(
 static void ggml_compute_forward_rope(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, src0, dst);
+                ggml_compute_forward_rope_f16(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, src0, dst);
+                ggml_compute_forward_rope_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -12718,6 +13599,7 @@ static void ggml_compute_forward_rope(
 static void ggml_compute_forward_rope_back_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -12735,7 +13617,7 @@ static void ggml_compute_forward_rope_back_f32(
     float xpos_base;
     bool xpos_down;
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_dims = ((int32_t *) dst->op_params)[1];
     const int mode   = ((int32_t *) dst->op_params)[2];
     const int n_ctx  = ((int32_t *) dst->op_params)[3]; UNUSED(n_ctx);
@@ -12744,9 +13626,7 @@ static void ggml_compute_forward_rope_back_f32(
     memcpy(&xpos_base,  (int32_t *) dst->op_params + 6, sizeof(float));
     memcpy(&xpos_down,  (int32_t *) dst->op_params + 7, sizeof(bool));
 
-    assert(n_past >= 0);
-
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
     //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
@@ -12772,9 +13652,11 @@ static void ggml_compute_forward_rope_back_f32(
 
     const bool is_neox = mode & 2;
 
+    const int32_t * pos = (const int32_t *) src1->data;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t p = pos[i2];
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -12786,7 +13668,7 @@ static void ggml_compute_forward_rope_back_f32(
                         const float cos_theta = cosf(theta);
                         const float sin_theta = sinf(theta);
                         // zeta scaling for xPos only:
-                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), (n_past + i2) / xpos_base) : 1.0f;
+                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
                         if (xpos_down) zeta = 1.0f / zeta;
 
                         theta *= theta_scale;
@@ -12829,6 +13711,7 @@ static void ggml_compute_forward_rope_back_f32(
 static void ggml_compute_forward_rope_back_f16(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
@@ -12839,13 +13722,11 @@ static void ggml_compute_forward_rope_back_f16(
     // dx = rope_back(dy, src1)
     // src0 is dy, src1 contains options
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_dims = ((int32_t *) dst->op_params)[1];
     const int mode   = ((int32_t *) dst->op_params)[2];
 
-    assert(n_past >= 0);
-
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
     //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
@@ -12871,9 +13752,11 @@ static void ggml_compute_forward_rope_back_f16(
 
     const bool is_neox = mode & 2;
 
+    const int32_t * pos = (const int32_t *) src1->data;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            const int64_t p = pos[i2];
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -12925,15 +13808,16 @@ static void ggml_compute_forward_rope_back_f16(
 static void ggml_compute_forward_rope_back(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_back_f16(params, src0, dst);
+                ggml_compute_forward_rope_back_f16(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_back_f32(params, src0, dst);
+                ggml_compute_forward_rope_back_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -12944,7 +13828,7 @@ static void ggml_compute_forward_rope_back(
 
 // ggml_compute_forward_conv_1d
 
-static void ggml_compute_forward_conv_1d_s1_ph_f16_f32(
+static void ggml_compute_forward_conv_1d_f16_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
@@ -12956,52 +13840,43 @@ static void ggml_compute_forward_conv_1d_s1_ph_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
 
     const int nk = ne00;
-    const int nh = nk/2;
 
-    const int ew0 = ggml_up32(ne01);
+    // size of the convolution row - the kernel size unrolled across all input channels
+    const int ew0 = nk*ne01;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
 
-    GGML_ASSERT(ne00 % 2 == 1); // TODO: support even kernel sizes
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
         memset(params->wdata, 0, params->wsize);
 
-        // prepare kernel data (src0)
-        {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
+        ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
-                    ggml_fp16_t * dst_data = wdata + i02*ew0*ne00;
-                    for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i00*ew0 + i01] = src[i00];
+        for (int64_t i11 = 0; i11 < ne11; i11++) {
+            const float * const src = (float *)((char *) src1->data + i11*nb11);
+            ggml_fp16_t * dst_data = wdata;
+
+            for (int64_t i0 = 0; i0 < ne0; i0++) {
+                for (int64_t ik = 0; ik < nk; ik++) {
+                    const int idx0 = i0*s0 + ik*d0 - p0;
+
+                    if(!(idx0 < 0 || idx0 >= ne10)) {
+                        dst_data[i0*ew0 + i11*nk + ik] = GGML_FP32_TO_FP16(src[idx0]);
                     }
                 }
             }
         }
 
-        // prepare source data (src1)
-        {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + ne02*ew0*ne00;
-
-            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                const float * const src = (float *)((char *) src1->data + i11*nb11);
-                ggml_fp16_t * dst_data = wdata;
-                for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[(i10 + nh)*ew0 + i11] = GGML_FP32_TO_FP16(src[i10]);
-                }
-            }
-        }
-
         return;
     }
 
@@ -13010,7 +13885,7 @@ static void ggml_compute_forward_conv_1d_s1_ph_f16_f32(
     }
 
     // total rows in dst
-    const int nr = ne02;
+    const int nr = ne2;
 
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
@@ -13019,23 +13894,22 @@ static void ggml_compute_forward_conv_1d_s1_ph_f16_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int64_t i0 = 0; i0 < ne10; ++i0) {
-            dst_data[i0] = 0;
-            for (int k = -nh; k <= nh; k++) {
-                float v = 0.0f;
-                ggml_vec_dot_f16(ew0, &v,
-                        (ggml_fp16_t *) params->wdata +   i1*ew0*ne00 +      (nh + k)*ew0,
-                        (ggml_fp16_t *) params->wdata + ne02*ew0*ne00 + (i0 + nh + k)*ew0);
+    ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-                dst_data[i0] += v;
+    for (int i2 = 0; i2 < ne2; i2++) {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1);
+
+            for (int i0 = 0; i0 < ne0; i0++) {
+                ggml_vec_dot_f16(ew0, dst_data + i0,
+                        (ggml_fp16_t *) ((char *) src0->data + i1*nb02),
+                        (ggml_fp16_t *)                wdata + i2*nb2 + i0*ew0);
             }
         }
     }
 }
 
-static void ggml_compute_forward_conv_1d_s1_ph_f32(
+static void ggml_compute_forward_conv_1d_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
@@ -13047,52 +13921,42 @@ static void ggml_compute_forward_conv_1d_s1_ph_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
 
     const int nk = ne00;
-    const int nh = nk/2;
 
-    const int ew0 = ggml_up32(ne01);
+    const int ew0 = nk*ne01;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
 
-    GGML_ASSERT(ne00 % 2 == 1); // TODO: support even kernel sizes
     GGML_ASSERT(nb00 == sizeof(float));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
         memset(params->wdata, 0, params->wsize);
 
-        // prepare kernel data (src0)
-        {
-            float * const wdata = (float *) params->wdata + 0;
+        float * const wdata = (float *) params->wdata + 0;
 
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
-                    float * dst_data = wdata + i02*ew0*ne00;
-                    for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i00*ew0 + i01] = src[i00];
+        for (int64_t i11 = 0; i11 < ne11; i11++) {
+            const float * const src = (float *)((char *) src1->data + i11*nb11);
+            float * dst_data = wdata;
+
+            for (int64_t i0 = 0; i0 < ne0; i0++) {
+                for (int64_t ik = 0; ik < nk; ik++) {
+                    const int idx0 = i0*s0 + ik*d0 - p0;
+
+                    if(!(idx0 < 0 || idx0 >= ne10)) {
+                        dst_data[i0*ew0 + i11*nk + ik] = src[idx0];
                     }
                 }
             }
         }
 
-        // prepare source data (src1)
-        {
-            float * const wdata = (float *) params->wdata + ne02*ew0*ne00;
-
-            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                const float * const src = (float *)((char *) src1->data + i11*nb11);
-                float * dst_data = wdata;
-                for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[(i10 + nh)*ew0 + i11] = src[i10];
-                }
-            }
-        }
-
         return;
     }
 
@@ -13110,101 +13974,126 @@ static void ggml_compute_forward_conv_1d_s1_ph_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int64_t i0 = 0; i0 < ne10; ++i0) {
-            dst_data[i0] = 0;
-            for (int k = -nh; k <= nh; k++) {
-                float v = 0.0f;
-                ggml_vec_dot_f32(ew0, &v,
-                        (float *) params->wdata +   i1*ew0*ne00 +      (nh + k)*ew0,
-                        (float *) params->wdata + ne02*ew0*ne00 + (i0 + nh + k)*ew0);
+    float * const wdata = (float *) params->wdata + 0;
 
-                dst_data[i0] += v;
+    for (int i2 = 0; i2 < ne2; i2++) {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1);
+
+            for (int i0 = 0; i0 < ne0; i0++) {
+                ggml_vec_dot_f32(ew0, dst_data + i0,
+                        (float *) ((char *) src0->data + i1*nb02),
+                        (float *)                wdata + i2*nb2 + i0*ew0);
             }
         }
     }
 }
 
-static void ggml_compute_forward_conv_1d_s1_ph(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_conv_1d_s1_ph_f16_f32(params, src0, src1, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_conv_1d_s1_ph_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
+static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k,
+                             ggml_fp16_t * A,
+                             ggml_fp16_t * B,
+                             float * C,
+                             const int ith, const int nth) {
+    // does not seem to make a difference
+    int64_t m0, m1, n0, n1;
+    // patches per thread
+    if (m > n) {
+        n0 = 0;
+        n1 = n;
+
+        // total patches in dst
+        const int np = m;
+
+        // patches per thread
+        const int dp = (np + nth - 1)/nth;
+
+        // patch range for this thread
+        m0 = dp*ith;
+        m1 = MIN(m0 + dp, np);
+    } else {
+        m0 = 0;
+        m1 = m;
+
+        // total patches in dst
+        const int np = n;
+
+        // patches per thread
+        const int dp = (np + nth - 1)/nth;
+
+        // patch range for this thread
+        n0 = dp*ith;
+        n1 = MIN(n0 + dp, np);
+    }
+
+    // block-tiling attempt
+    int64_t blck_n = 16;
+    int64_t blck_m = 16;
+
+    // int64_t CACHE_SIZE = 2 * 1024 * 1024; // 2MB
+    // int64_t blck_size = CACHE_SIZE / (sizeof(float) + 2 * sizeof(ggml_fp16_t) * K);
+    // if (blck_size > 0) {
+    //     blck_0 = 4;
+    //     blck_1 = blck_size / blck_0;
+    //     if (blck_1 < 0) {
+    //         blck_1 = 1;
+    //     }
+    //     // blck_0 = (int64_t)sqrt(blck_size);
+    //     // blck_1 = blck_0;
+    // }
+    // // printf("%zd %zd %zd %zd\n", blck_size, K, blck_0, blck_1);
+
+    for (int j = n0; j < n1; j+=blck_n) {
+        for (int i = m0; i < m1; i+=blck_m) {
+            // printf("i j k => %d %d %d\n", i, j, K);
+            for (int ii = i; ii < i + blck_m && ii < m1; ii++) {
+                for (int jj = j; jj < j + blck_n && jj < n1; jj++) {
+                    ggml_vec_dot_f16(k,
+                                    C + ii*n + jj,
+                                    A + ii * k,
+                                    B + jj * k);
+                }
+            }
+        }
     }
 }
 
-static void ggml_compute_forward_conv_1d_s2_ph_f16_f32(
+// src0: kernel [OC, IC, K]
+// src1: signal [N, IC, IL]
+// dst:  result [N, OL, IC*K]
+static void ggml_compute_forward_conv_1d_stage_0_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16);
 
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
     GGML_TENSOR_BINARY_OP_LOCALS;
 
+    const int64_t N  = ne12;
+    const int64_t IC = ne11;
+    const int64_t IL = ne10;
+
+    const int64_t K = ne00;
+
+    const int64_t OL = ne1;
+
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int nk = ne00;
-    const int nh = nk/2;
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
 
-    const int ew0 = ggml_up32(ne01);
-
-    GGML_ASSERT(ne00 % 2 == 1); // TODO: support even kernel sizes
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
-        memset(params->wdata, 0, params->wsize);
-
-        // prepare kernel data (src0)
-        {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
-                    ggml_fp16_t * dst_data = wdata + i02*ew0*ne00;
-                    for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i00*ew0 + i01] = src[i00];
-                    }
-                }
-            }
-        }
-
-        // prepare source data (src1)
-        {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + ne02*ew0*ne00;
-
-            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                const float * const src = (float *)((char *) src1->data + i11*nb11);
-                ggml_fp16_t * dst_data = wdata;
-                for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[(i10 + nh)*ew0 + i11] = GGML_FP32_TO_FP16(src[i10]);
-                }
-            }
-        }
-
+        memset(dst->data, 0, ggml_nbytes(dst));
         return;
     }
 
@@ -13212,90 +14101,48 @@ static void ggml_compute_forward_conv_1d_s2_ph_f16_f32(
         return;
     }
 
-    // total rows in dst
-    const int nr = ne02;
+    // im2col: [N, IC, IL] => [N, OL, IC*K]
+    {
+        ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data;
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+        for (int64_t in = 0; in < N; in++) {
+            for (int64_t iol = 0; iol < OL; iol++) {
+                for (int64_t iic = ith; iic < IC; iic+=nth) {
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+                    // micro kernel
+                    ggml_fp16_t * dst_data = wdata + (in*OL + iol)*(IC*K); // [IC, K]
+                    const float * const src_data = (float *)((char *) src1->data + in*nb12 + iic*nb11); // [IL]
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int64_t i0 = 0; i0 < ne10; i0 += 2) {
-            dst_data[i0/2] = 0;
-            for (int k = -nh; k <= nh; k++) {
-                float v = 0.0f;
-                ggml_vec_dot_f16(ew0, &v,
-                        (ggml_fp16_t *) params->wdata +   i1*ew0*ne00 +      (nh + k)*ew0,
-                        (ggml_fp16_t *) params->wdata + ne02*ew0*ne00 + (i0 + nh + k)*ew0);
+                    for (int64_t ik = 0; ik < K; ik++) {
+                        const int64_t iil = iol*s0 + ik*d0 - p0;
 
-                dst_data[i0/2] += v;
+                        if (!(iil < 0 || iil >= IL)) {
+                            dst_data[iic*K + ik] = GGML_FP32_TO_FP16(src_data[iil]);
+                        }
+                    }
+                }
             }
         }
     }
 }
 
-static void ggml_compute_forward_conv_1d_s2_ph_f32(
+// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
+// src0: [OC, IC, K]
+// src1: [N, OL, IC * K]
+// result: [N, OC, OL]
+static void ggml_compute_forward_conv_1d_stage_1_f16(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nk = ne00;
-    const int nh = nk/2;
-
-    const int ew0 = ggml_up32(ne01);
-
-    GGML_ASSERT(ne00 % 2 == 1); // TODO: support even kernel sizes
-    GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(nb10 == sizeof(float));
-
     if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
-        memset(params->wdata, 0, params->wsize);
-
-        // prepare kernel data (src0)
-        {
-            float * const wdata = (float *) params->wdata + 0;
-
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
-                    float * dst_data = wdata + i02*ew0*ne00;
-                    for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i00*ew0 + i01] = src[i00];
-                    }
-                }
-            }
-        }
-
-        // prepare source data (src1)
-        {
-            float * const wdata = (float *) params->wdata + ne02*ew0*ne00;
-
-            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                const float * const src = (float *)((char *) src1->data + i11*nb11);
-                float * dst_data = wdata;
-                for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[(i10 + nh)*ew0 + i11] = src[i10];
-                }
-            }
-        }
-
         return;
     }
 
@@ -13303,72 +14150,294 @@ static void ggml_compute_forward_conv_1d_s2_ph_f32(
         return;
     }
 
-    // total rows in dst
-    const int nr = ne02;
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb0  == sizeof(float));
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+    const int N = ne12;
+    const int OL = ne11;
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int64_t i0 = 0; i0 < ne10; i0 += 2) {
-            dst_data[i0/2] = 0;
-            for (int k = -nh; k <= nh; k++) {
-                float v = 0.0f;
-                ggml_vec_dot_f32(ew0, &v,
-                        (float *) params->wdata +   i1*ew0*ne00 +      (nh + k)*ew0,
-                        (float *) params->wdata + ne02*ew0*ne00 + (i0 + nh + k)*ew0);
+    const int OC = ne02;
+    const int IC = ne01;
+    const int K  = ne00;
 
-                dst_data[i0/2] += v;
-            }
-        }
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    int64_t m = OC;
+    int64_t n = OL;
+    int64_t k = IC * K;
+
+    // [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
+    for (int i = 0; i < N; i++) {
+        ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k]
+        ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k]
+        float * C = (float *)dst->data + i * m * n; // [m, n]
+
+        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
     }
 }
 
-static void ggml_compute_forward_conv_1d_s2_ph(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_conv_1d_s2_ph_f16_f32(params, src0, src1, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_conv_1d_s2_ph_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// ggml_compute_forward_conv_1d
-
 static void ggml_compute_forward_conv_1d(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+    switch(src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_1d_f16_f32(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_conv_1d_f32(params, src0, src1, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+static void ggml_compute_forward_conv_1d_stage_0(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    switch(src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_1d_stage_0_f32(params, src0, src1, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+static void ggml_compute_forward_conv_1d_stage_1(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    switch(src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_1d_stage_1_f16(params, src0, src1, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+// ggml_compute_forward_conv_transpose_1d
+
+static void ggml_compute_forward_conv_transpose_1d_f16_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int64_t t0 = ggml_perf_time_us();
+    UNUSED(t0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nk = ne00*ne01*ne02;
+
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    if (params->type == GGML_TASK_INIT) {
+        memset(params->wdata, 0, params->wsize);
+
+        // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
+        {
+            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
+
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
+                    ggml_fp16_t * dst_data = wdata + i01*ne00*ne02;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        dst_data[i00*ne02 + i02] = src[i00];
+                    }
+                }
+            }
+        }
+
+        // permute source data (src1) from (L x Cin) to (Cin x L)
+        {
+            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk;
+            ggml_fp16_t * dst_data = wdata;
+
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
+                const float * const src = (float *)((char *) src1->data + i11*nb11);
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
+                    dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]);
+                }
+            }
+        }
+
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
     const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
-    GGML_ASSERT(d0 == 1); // dilation not supported
-    GGML_ASSERT(p0 == src0->ne[0]/2); // only half padding supported
-    if (s0 == 1) {
-        ggml_compute_forward_conv_1d_s1_ph(params, src0, src1, dst);
-    } else if (s0 == 2) {
-        ggml_compute_forward_conv_1d_s2_ph(params, src0, src1, dst);
-    } else {
-        GGML_ASSERT(false); // only stride 1 and 2 supported
-    };
+
+    // total rows in dst
+    const int nr = ne1;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    ggml_fp16_t * const wdata     = (ggml_fp16_t *) params->wdata + 0;
+    ggml_fp16_t * const wdata_src = wdata + nk;
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * dst_data = (float *)((char *) dst->data + i1*nb1);
+        ggml_fp16_t * wdata_kernel = wdata + i1*ne02*ne00;
+        for (int i10 = 0; i10 < ne10; i10++) {
+            const int i1n = i10*ne11;
+            for (int i00 = 0; i00 < ne00; i00++) {
+                float v = 0;
+                ggml_vec_dot_f16(ne02, &v,
+                        (ggml_fp16_t *)    wdata_src + i1n,
+                        (ggml_fp16_t *) wdata_kernel + i00*ne02);
+                dst_data[i10*s0 + i00] += v;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_transpose_1d_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int64_t t0 = ggml_perf_time_us();
+    UNUSED(t0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nk = ne00*ne01*ne02;
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    if (params->type == GGML_TASK_INIT) {
+        memset(params->wdata, 0, params->wsize);
+
+        // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
+        {
+            float * const wdata = (float *) params->wdata + 0;
+
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
+                    float * dst_data = wdata + i01*ne00*ne02;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        dst_data[i01*ne00*ne02 + i00*ne02 + i02] = src[i00];
+                    }
+                }
+            }
+        }
+
+        // prepare source data (src1)
+        {
+            float * const wdata = (float *) params->wdata + nk;
+            float * dst_data = wdata;
+
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
+                const float * const src = (float *)((char *) src1->data + i11*nb11);
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
+                    dst_data[i10*ne11 + i11] = src[i10];
+                }
+            }
+        }
+
+        return;
+    }
+
+    if (params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+
+    // total rows in dst
+    const int nr = ne1;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    float * const wdata     = (float *) params->wdata + 0;
+    float * const wdata_src = wdata + nk;
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * dst_data = (float *)((char *) dst->data + i1*nb1);
+        float * wdata_kernel = wdata + i1*ne02*ne00;
+        for (int i10 = 0; i10 < ne10; i10++) {
+            const int i1n = i10*ne11;
+            for (int i00 = 0; i00 < ne00; i00++) {
+                float v = 0;
+                ggml_vec_dot_f32(ne02, &v,
+                        wdata_src + i1n,
+                        wdata_kernel + i00*ne02);
+                dst_data[i10*s0 + i00] += v;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_transpose_1d(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_transpose_1d_f16_f32(params, src0, src1, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_conv_transpose_1d_f32(params, src0, src1, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
 }
 
 // ggml_compute_forward_conv_2d
@@ -13385,7 +14454,7 @@ static void ggml_compute_forward_conv_2d_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -13413,20 +14482,22 @@ static void ggml_compute_forward_conv_2d_f16_f32(
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int i12 = 0; i12 < ne12; i12++) {
-                const float * const src = (float *)((char *) src1->data + i12*nb12);
-                ggml_fp16_t * dst_data = wdata;
+            for (int i13 = 0; i13 < ne13; i13++) {
+                for (int i12 = 0; i12 < ne12; i12++) {
+                    const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12);
+                    ggml_fp16_t * dst_data = wdata + i13*(ne1*ne0*ew0);
 
-                for (int i1 = 0; i1 < ne1; i1++) {
-                    for (int i0 = 0; i0 < ne0; i0++) {
-                        for (int ik1 = 0; ik1 < nk1; ik1++) {
-                            for (int ik0 = 0; ik0 < nk0; ik0++) {
-                                const int idx0 = i0*s0 + ik0*d0 - p0;
-                                const int idx1 = i1*s1 + ik1*d1 - p1;
+                    for (int i1 = 0; i1 < ne1; i1++) {
+                        for (int i0 = 0; i0 < ne0; i0++) {
+                            for (int ik1 = 0; ik1 < nk1; ik1++) {
+                                for (int ik0 = 0; ik0 < nk0; ik0++) {
+                                    const int idx0 = i0*s0 + ik0*d0 - p0;
+                                    const int idx1 = i1*s1 + ik1*d1 - p1;
 
-                                if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) {
-                                    dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
-                                        GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]);
+                                    if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) {
+                                        dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
+                                            GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]);
+                                    }
                                 }
                             }
                         }
@@ -13497,7 +14568,6 @@ static void ggml_compute_forward_conv_transpose_2d(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
               struct ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -13506,7 +14576,7 @@ static void ggml_compute_forward_conv_transpose_2d(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_BINARY_OP_LOCALS;
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -13557,7 +14627,7 @@ static void ggml_compute_forward_conv_transpose_2d(
         return;
     }
 
-    const int32_t stride = ((const int32_t*)(opt0->data))[0];
+    const int32_t stride = ggml_get_op_params_i32(dst, 0);
 
     // total patches in dst
     const int np = ne2;
@@ -13570,7 +14640,7 @@ static void ggml_compute_forward_conv_transpose_2d(
     const int ip1 = MIN(ip0 + dp, np);
 
     ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-    ggml_fp16_t * const wdata_src = (ggml_fp16_t *) params->wdata + nk;
+    ggml_fp16_t * const wdata_src = wdata + nk;
 
     for (int i2 = ip0; i2 < ip1; i2++) { // Cout
         float * dst_data = (float *)((char *) dst->data + i2*nb2);
@@ -13582,9 +14652,8 @@ static void ggml_compute_forward_conv_transpose_2d(
                     for (int i00 = 0; i00 < ne00; i00++) {
                         float v = 0;
                         ggml_vec_dot_f16(ne03, &v,
-                                (ggml_fp16_t *) wdata_src + i1n,
-                                (ggml_fp16_t *) wdata_kernel + i01*ne00*ne03 + i00*ne03);
-
+                                wdata_src + i1n,
+                                wdata_kernel + i01*ne00*ne03 + i00*ne03);
                         dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
@@ -13766,7 +14835,7 @@ static void ggml_compute_forward_upscale_f32(
 
     const int ith = params->ith;
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const int scale_factor = dst->op_params[0];
 
@@ -13818,14 +14887,14 @@ static void ggml_compute_forward_flash_attn_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb);
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb);
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -13895,10 +14964,11 @@ static void ggml_compute_forward_flash_attn_f32(
             S[i] = -INFINITY;
         }
 
-        for (int64_t ic = 0; ic < nek1; ++ic) {
+        const int64_t masked_begin = masked ? (P + iq1 + 1) : M;
+        for (int64_t ic = 0; ic < masked_begin; ++ic) {
             // k indices
             const int ik3 = iq3;
-            const int ik2 = iq2;
+            const int ik2 = iq2 % nek2;
             const int ik1 = ic;
 
             // S indices
@@ -13911,20 +14981,18 @@ static void ggml_compute_forward_flash_attn_f32(
         }
 
         // scale
-        ggml_vec_scale_f32(nek1, S, scale);
+        ggml_vec_scale_f32(masked_begin, S, scale);
 
-        if (masked) {
-            for (int64_t i = P; i < M; i++) {
-                if (i > P + iq1) {
-                    S[i] = -INFINITY;
-                }
-            }
+        for (int64_t i = masked_begin; i < M; i++) {
+            S[i] = -INFINITY;
         }
 
         // softmax
+        // exclude known -INF S[..] values from max and loop
+        // dont forget to set their SW values to zero
         {
             float max = -INFINITY;
-            ggml_vec_max_f32(M, &max, S);
+            ggml_vec_max_f32(masked_begin, &max, S);
 
             ggml_float sum = 0.0;
             {
@@ -13934,19 +15002,28 @@ static void ggml_compute_forward_flash_attn_f32(
                 vvexpf(S, S, &Mup);
                 ggml_vec_sum_f32(Mup, &sum, S);
 #else
-                uint16_t   scvt[GGML_SOFT_MAX_UNROLL];
+                uint16_t   scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
                 ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
 
                 for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
+                    if (i >= masked_begin) {
+                        break;
+                    }
                     float * SS = S + i;
 
                     for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
-                        if (SS[j] == -INFINITY) {
+                        if (i + j >= masked_begin) {
+                            break;
+                        } else if (SS[j] == -INFINITY) {
                             SS[j] = 0.0f;
                         } else {
+#ifndef GGML_FLASH_ATTN_EXP_FP16
+                            const float val = expf(SS[j] - max);
+#else
                             ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
                             memcpy(&scvt[j], &s, sizeof(uint16_t));
                             const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
+#endif
                             sump[j] += (ggml_float)val;
                             SS[j] = val;
                         }
@@ -13962,10 +15039,10 @@ static void ggml_compute_forward_flash_attn_f32(
             assert(sum > 0.0);
 
             sum = 1.0/sum;
-            ggml_vec_scale_f32(M, S, sum);
+            ggml_vec_scale_f32(masked_begin, S, sum);
 
 #ifndef NDEBUG
-            for (int i = 0; i < M; ++i) {
+            for (int i = 0; i < masked_begin; ++i) {
                 assert(!isnan(S[i]));
                 assert(!isinf(S[i]));
             }
@@ -13978,9 +15055,13 @@ static void ggml_compute_forward_flash_attn_f32(
             const int i2 = iq2;
             const int i3 = iq3;
 
-            ggml_vec_dot_f32(nek1,
-                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
-                    (float *) ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
+            // v indices
+            const int iv2 = iq2 % nev2;
+            const int iv3 = iq3;
+
+            ggml_vec_dot_f32(masked_begin,
+                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
+                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
                     S);
         }
     }
@@ -13996,14 +15077,14 @@ static void ggml_compute_forward_flash_attn_f16(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb);
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb);
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -14077,7 +15158,7 @@ static void ggml_compute_forward_flash_attn_f16(
             for (int64_t ic = 0; ic < nek1; ++ic) {
                 // k indices
                 const int ik3 = iq3;
-                const int ik2 = iq2;
+                const int ik2 = iq2 % nek2;
                 const int ik1 = ic;
 
                 // S indices
@@ -14092,7 +15173,7 @@ static void ggml_compute_forward_flash_attn_f16(
             for (int64_t ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
                 // k indices
                 const int ik3 = iq3;
-                const int ik2 = iq2;
+                const int ik2 = iq2 % nek2;
                 const int ik1 = ic;
 
                 // S indices
@@ -14117,6 +15198,8 @@ static void ggml_compute_forward_flash_attn_f16(
         }
 
         // softmax
+        // todo: exclude known -INF S[..] values from max and loop, assuming their results to be zero.
+        // dont forget to set their S values to zero
         {
             float max = -INFINITY;
             ggml_vec_max_f32(M, &max, S);
@@ -14173,6 +15256,7 @@ static void ggml_compute_forward_flash_attn_f16(
             S16[i] = GGML_FP32_TO_FP16(S[i]);
         }
 
+        // todo: exclude known zero S[..] values from dot (reducing nev0 and increasing begin of v and S16).
         if (GGML_VEC_DOT_UNROLL == 1 || (nev1 % GGML_VEC_DOT_UNROLL != 0)) {
             for (int64_t ic = 0; ic < nev1; ++ic) {
                 // dst indices
@@ -14180,9 +15264,13 @@ static void ggml_compute_forward_flash_attn_f16(
                 const int i2 = iq2;
                 const int i3 = iq3;
 
-                ggml_vec_dot_f16(nek1,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
-                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
+                // v indices
+                const int iv2 = iq2 % nev2;
+                const int iv3 = iq3;
+
+                ggml_vec_dot_f16(nev0,
+                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
+                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
                         S16);
             }
         } else {
@@ -14192,9 +15280,13 @@ static void ggml_compute_forward_flash_attn_f16(
                 const int i2 = iq2;
                 const int i3 = iq3;
 
-                ggml_vec_dot_f16_unroll(nek1, nbv1,
-                        (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
-                        ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
+                // v indices
+                const int iv2 = iq2 % nev2;
+                const int iv3 = iq3;
+
+                ggml_vec_dot_f16_unroll(nev0, nbv1,
+                        (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
+                        ((char *)             v->data + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
                         S16);
             }
         }
@@ -14237,18 +15329,18 @@ static void ggml_compute_forward_flash_ff_f16(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_LOCALS(int64_t, nea,  a,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nba,  a,   nb);
-    GGML_TENSOR_LOCALS(int64_t, neb0, b0,  ne);
-    GGML_TENSOR_LOCALS(size_t,  nbb0, b0,  nb);
-    GGML_TENSOR_LOCALS(int64_t, neb1, b1,  ne);
-    GGML_TENSOR_LOCALS(size_t,  nbb1, b1,  nb);
-    GGML_TENSOR_LOCALS(int64_t, nec0, c0,  ne);
-    GGML_TENSOR_LOCALS(size_t,  nbc0, c0,  nb);
-    GGML_TENSOR_LOCALS(int64_t, nec1, c1,  ne);
-    GGML_TENSOR_LOCALS(size_t,  nbc1, c1,  nb);
-    GGML_TENSOR_LOCALS(int64_t, ne,   dst, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb,   dst, nb);
+    GGML_TENSOR_LOCALS(int64_t, nea,  a,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nba,  a,   nb)
+    GGML_TENSOR_LOCALS(int64_t, neb0, b0,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nbb0, b0,  nb)
+    GGML_TENSOR_LOCALS(int64_t, neb1, b1,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nbb1, b1,  nb)
+    GGML_TENSOR_LOCALS(int64_t, nec0, c0,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nbc0, c0,  nb)
+    GGML_TENSOR_LOCALS(int64_t, nec1, c1,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nbc1, c1,  nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,   dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,   dst, nb)
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -14396,16 +15488,16 @@ static void ggml_compute_forward_flash_attn_back_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb);
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb);
-    GGML_TENSOR_LOCALS(int64_t, ned, d,   ne);
-    GGML_TENSOR_LOCALS(size_t,  nbd, d,   nb);
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne);
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb);
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ned, d,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbd, d,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -14453,10 +15545,37 @@ static void ggml_compute_forward_flash_attn_back_f32(
         return;
     }
 
-    // parallelize by q rows using ggml_vec_dot_f32
+    const int64_t elem_q = ggml_nelements(q);
+    const int64_t elem_k = ggml_nelements(k);
 
-    // total rows in q
-    const int nr = neq2*neq3;
+    enum ggml_type result_type = dst->type;
+    GGML_ASSERT(ggml_blck_size(result_type) == 1);
+    const size_t tsize = ggml_type_size(result_type);
+
+    const size_t offs_q = 0;
+    const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
+    const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
+
+    void * grad_q = (char *) dst->data;
+    void * grad_k = (char *) dst->data + offs_k;
+    void * grad_v = (char *) dst->data + offs_v;
+
+    const size_t nbgq1 = nb0*neq0;
+    const size_t nbgq2 = nb0*neq0*neq1;
+    const size_t nbgq3 = nb0*neq0*neq1*neq2;
+
+    const size_t nbgk1 = nb0*nek0;
+    const size_t nbgk2 = nb0*nek0*nek1;
+    const size_t nbgk3 = nb0*nek0*nek1*neq2;
+
+    const size_t nbgv1 = nb0*nev0;
+    const size_t nbgv2 = nb0*nev0*nev1;
+    const size_t nbgv3 = nb0*nev0*nev1*neq2;
+
+    // parallelize by k rows using ggml_vec_dot_f32
+
+    // total rows in k
+    const int nr = nek2*nek3;
 
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
@@ -14469,264 +15588,243 @@ static void ggml_compute_forward_flash_attn_back_f32(
 
     //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
 
+    // how often k2 (and v2) is repeated in q2
+    int nrep = neq2/nek2;
+
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
-        const int iq3 = ir/(neq2);
-        const int iq2 = ir - iq3*neq2;
-        for ( int iq1 = 0; iq1 < neq1; ++iq1) {
+        const int ik3 = ir/(nek2);
+        const int ik2 = ir - ik3*nek2;
 
+        const int iq3 = ik3;
+        const int id3 = ik3;
+        const int iv3 = ik3;
+        const int iv2 = ik2;
 
-            // not sure about CACHE_LINE_SIZE_F32..
-            // - maybe it must not be multiplied by 2 and excluded from .. in SM 1*(..) offset?
-            float * S  = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 0*(mxDM+CACHE_LINE_SIZE_F32);
-            float * SM = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 1*(mxDM+CACHE_LINE_SIZE_F32);
+        for (int irep = 0; irep < nrep; ++irep) {
+            const int iq2 = ik2 + irep*nek2;
+            const int id2 = iq2;
 
-            for (int i = M; i < Mup; ++i) {
-                S[i] = -INFINITY;
-            }
+            // (ik2 + irep*nek2) % nek2 == ik2
+            for (int iq1 = 0; iq1 < neq1; ++iq1) {
+                const int id1 = iq1;
 
-            for (int64_t ic = 0; ic < nek1; ++ic) {
-                // k indices
-                const int ik3 = iq3;
-                const int ik2 = iq2;
-                const int ik1 = ic;
+                // not sure about CACHE_LINE_SIZE_F32..
+                // - maybe it must not be multiplied by 2 and excluded from .. in SM 1*(..) offset?
+                float * S  = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 0*(mxDM+CACHE_LINE_SIZE_F32);
+                float * SM = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 1*(mxDM+CACHE_LINE_SIZE_F32);
 
-                // S indices
-                const int i1 = ik1;
-
-                ggml_vec_dot_f32(neq0,
-                        S + i1,
-                        (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                        (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-            }
-
-            // scale
-            ggml_vec_scale_f32(nek1, S, scale);
-
-            if (masked) {
-                for (int64_t i = P; i < M; i++) {
-                    if (i > P + iq1) {
-                        S[i] = -INFINITY;
-                    }
+                for (int i = M; i < Mup; ++i) {
+                    S[i] = -INFINITY;
                 }
-            }
 
-            // softmax
-            {
-                float max = -INFINITY;
-                ggml_vec_max_f32(M, &max, S);
+                const int64_t masked_begin = masked ? (P + iq1 + 1) : M;
+                for (int64_t ic = 0; ic < masked_begin; ++ic) {
+                    // k indices
+                    const int ik1 = ic;
 
-                ggml_float sum = 0.0;
+                    // S indices
+                    const int i1 = ik1;
+
+                    ggml_vec_dot_f32(neq0,
+                            S + i1,
+                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                }
+
+                // scale
+                ggml_vec_scale_f32(masked_begin, S, scale);
+
+                for (int64_t i = masked_begin; i < M; i++) {
+                    S[i] = -INFINITY;
+                }
+
+                // softmax
+                // exclude known -INF S[..] values from max and loop
+                // dont forget to set their SM values to zero
                 {
+                    float max = -INFINITY;
+                    ggml_vec_max_f32(masked_begin, &max, S);
+
+                    ggml_float sum = 0.0;
+                    {
 #ifdef GGML_SOFT_MAX_ACCELERATE
-                    max = -max;
-                    vDSP_vsadd(SM, 1, &max, SM, 1, Mup);
-                    vvexpf(SM, SM, &Mup);
-                    ggml_vec_sum_f32(Mup, &sum, SM);
+                        max = -max;
+                        vDSP_vsadd(SM, 1, &max, SM, 1, Mup);
+                        vvexpf(SM, SM, &Mup);
+                        ggml_vec_sum_f32(Mup, &sum, SM);
 #else
-                    uint16_t   scvt[GGML_SOFT_MAX_UNROLL];
-                    ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
+                        uint16_t   scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
+                        ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
 
-                    for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
-                        float * SR =  S + i;
-                        float * SW = SM + i;
+                        for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
+                            if (i >= masked_begin) {
+                                break;
+                            }
+                            float * SR =  S + i;
+                            float * SW = SM + i;
 
-                        for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
-                            if (SR[j] == -INFINITY) {
-                                SW[j] = 0.0f;
-                            } else {
-                                ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max);
-                                memcpy(&scvt[j], &s, sizeof(uint16_t));
-                                const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
-                                sump[j] += (ggml_float)val;
-                                SW[j] = val;
+                            for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
+                                if (i + j >= masked_begin) {
+                                    break;
+                                } else if (SR[j] == -INFINITY) {
+                                    SW[j] = 0.0f;
+                                } else {
+#ifndef GGML_FLASH_ATTN_EXP_FP16
+                                    const float val = expf(SR[j] - max);
+#else
+                                    ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max);
+                                    memcpy(&scvt[j], &s, sizeof(uint16_t));
+                                    const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
+#endif
+                                    sump[j] += (ggml_float)val;
+                                    SW[j] = val;
+                                }
                             }
                         }
-                    }
 
-                    for (int i = 0; i < GGML_SOFT_MAX_UNROLL; i++) {
-                        sum += sump[i];
-                    }
+                        for (int i = 0; i < GGML_SOFT_MAX_UNROLL; i++) {
+                            sum += sump[i];
+                        }
 #endif
-                }
-
-                assert(sum > 0.0);
-
-                sum = 1.0/sum;
-                ggml_vec_scale_f32(M, SM, sum);
-
-            }
-
-            // step-by-step explanation
-            {
-                // forward-process                   shape      grads from backward process
-                // parallel_for iq2,iq3:
-                //  k[:D,:M,:,:]                     [D,M,:,:]  grad[k][:D,:M,iq2,iq3]  += grad[kcur]
-                //  q[:D,:N,:,:]                     [D,N,:,:]  grad[q][:D,iq1,iq2,iq3] += grad[qcur]
-                //  v[:M,:D,:,:]                     [M,D,:,:]  grad[v][:M,:D,iq2,iq3]  += grad[vcur]
-                //  for iq1:
-                //   kcur   = k[:D,:M,iq2,iq3]       [D,M,1,1]  grad[kcur] = grad[S1].T @ qcur
-                //   qcur   = q[:D,iq1,iq2,iq3]      [D,1,1,1]  grad[qcur] = grad[S1]   @ kcur
-                //   vcur   = v[:M,:D,iq2,iq3]       [M,D,1,1]  grad[vcur] = grad[S5].T @ S4
-                //   S0     = -Inf                   [D,1,1,1]
-                //  ~S1[i]  = dot(kcur[:D,i], qcur)
-                //   S1     = qcur @ kcur.T          [M,1,1,1]  grad[S1]   = grad[S2] * scale
-                //   S2     = S1 * scale             [M,1,1,1]  grad[S2]   = diag_mask_zero(grad[S3], P)
-                //   S3     = diag_mask_inf(S2, P)   [M,1,1,1]  grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
-                //   S4     = softmax(S3)            [M,1,1,1]  grad[S4]   = grad[S5] @ vcur
-                //  ~S5[i]  = dot(vcur[:,i], S4)
-                //   S5     = S4 @ vcur.T            [D,1,1,1]  grad[S5]   = d[:D,iq1,iq2,iq3]
-                //  ~dst[i,iq1,iq2,iq3]  = S5[i]              ^
-                //   dst[:D,iq1,iq2,iq3] = S5                 | grad[dst[:D,iq1,iq2,iq3]] = d[:D,iq1,iq2,iq3]
-                // dst                               backward-/ grad[dst]                 = d
-                //
-                // output gradients with their dependencies:
-                //
-                // grad[kcur] = grad[S1].T @ qcur
-                // grad[S1]   = diag_mask_zero(grad[S3], P) * scale
-                // grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
-                // grad[S4]   = grad[S5] @ vcur
-                // grad[S4]   = d[:D,iq1,iq2,iq3] @ vcur
-                // grad[qcur] = grad[S1]   @ kcur
-                // grad[vcur] = grad[S5].T @ S4
-                // grad[vcur] = d[:D,iq1,iq2,iq3].T @ S4
-                //
-                // in post-order:
-                //
-                // S1         = qcur @ kcur.T
-                // S2         = S1 * scale
-                // S3         = diag_mask_inf(S2, P)
-                // S4         = softmax(S3)
-                // grad[S4]   = d[:D,iq1,iq2,iq3] @ vcur
-                // grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
-                // grad[S1]   = diag_mask_zero(grad[S3], P) * scale
-                // grad[qcur] = grad[S1]   @ kcur
-                // grad[kcur] = grad[S1].T @ qcur
-                // grad[vcur] = d[:D,iq1,iq2,iq3].T @ S4
-                //
-                // using less variables (SM=S4):
-                //
-                // S             = diag_mask_inf(qcur @ kcur.T * scale, P)
-                // SM            = softmax(S)
-                // S             = d[:D,iq1,iq2,iq3] @ vcur
-                // dot_SM_gradSM = dot(SM, S)
-                // S             = SM * (S - dot(SM, S))
-                // S             = diag_mask_zero(S, P) * scale
-                //
-                // grad[q][:D,iq1,iq2,iq3] += S   @ kcur
-                // grad[k][:D,:M,iq2,iq3]  += S.T @ qcur
-                // grad[v][:M,:D,iq2,iq3]  += d[:D,iq1,iq2,iq3].T @ SM
-            }
-
-            // S = gradSM = d[:D,iq1,iq2,iq3] @ vcur
-            // S = d[:D,iq1,iq2,iq3] @ vcur
-            // S[:M] += vcur[:M,ic] * d[ic,iq1,iq2,iq3]
-            ggml_vec_set_f32(M, S, 0);
-            for (int64_t ic = 0; ic < D; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
-
-                ggml_vec_mad_f32(M,
-                        S,
-                         (float *) ((char *) v->data + (          ic*nbv1 + i2*nbv2 + i3*nbv3)),
-                        *(float *) ((char *) d->data + (ic*nbd0 + i1*nbd1 + i2*nbd2 + i3*nbd3)));
-            }
-
-            // S = SM * (S - dot(SM, S))
-            float dot_SM_gradSM = 0;
-            ggml_vec_dot_f32 (M, &dot_SM_gradSM, SM, S);
-            ggml_vec_acc1_f32(M, S, -dot_SM_gradSM);
-            ggml_vec_mul_f32 (M, S, S, SM);
-
-            // S = diag_mask_zero(S, P) * scale
-            if (masked) {
-                // for (int64_t i = P + iq1 + 1; i < M; i++) {
-                //     S[i] = 0;
-                // }
-                for (int64_t i = P; i < M; i++) {
-                    if (i > P + iq1) {
-                        S[i] = 0;
                     }
+
+                    assert(sum > 0.0);
+
+                    sum = 1.0/sum;
+                    ggml_vec_scale_f32(masked_begin, SM, sum);
+
                 }
-            }
-            ggml_vec_scale_f32(M, S, scale);
 
-            void * grad_q = (char *) dst->data;
-            void * grad_k = (char *) dst->data + nb0*D*N*neq2*neq3;
-            void * grad_v = (char *) dst->data + nb0*D*N*neq2*neq3 + nb0*D*M*neq2*neq3;
+                // step-by-step explanation
+                {
+                    // forward-process                    shape      grads from backward process
+                    // parallel_for ik2,ik3:
+                    //  for irep:
+                    //   iq2 = ik2 + irep*nek2
+                    //   k[:D,:M,:,:]                     [D,M,:,:]  grad[k][:D,:M,ik2,ik3]  += grad[kcur]
+                    //   q[:D,:N,:,:]                     [D,N,:,:]  grad[q][:D,iq1,iq2,iq3] += grad[qcur]
+                    //   v[:M,:D,:,:]                     [M,D,:,:]  grad[v][:M,:D,iv2,iv3]  += grad[vcur]
+                    //   for iq1:
+                    //    kcur   = k[:D,:M,ik2,ik3]       [D,M,1,1]  grad[kcur] = grad[S1].T @ qcur
+                    //    qcur   = q[:D,iq1,iq2,iq3]      [D,1,1,1]  grad[qcur] = grad[S1]   @ kcur
+                    //    vcur   = v[:M,:D,iv2,iv3]       [M,D,1,1]  grad[vcur] = grad[S5].T @ S4
+                    //    S0     = -Inf                   [D,1,1,1]
+                    //   ~S1[i]  = dot(kcur[:D,i], qcur)
+                    //    S1     = qcur @ kcur.T          [M,1,1,1]  grad[S1]   = grad[S2] * scale
+                    //    S2     = S1 * scale             [M,1,1,1]  grad[S2]   = diag_mask_zero(grad[S3], P)
+                    //    S3     = diag_mask_inf(S2, P)   [M,1,1,1]  grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
+                    //    S4     = softmax(S3)            [M,1,1,1]  grad[S4]   = grad[S5] @ vcur
+                    //   ~S5[i]  = dot(vcur[:,i], S4)
+                    //    S5     = S4 @ vcur.T            [D,1,1,1]  grad[S5]   = d[:D,id1,id2,id3]
+                    //   ~dst[i,iq1,iq2,iq3]  = S5[i]              ^
+                    //    dst[:D,iq1,iq2,iq3] = S5                 | grad[dst[:D,iq1,iq2,iq3]] = d[:D,id1,id2,id3]
+                    // dst                               backward-/ grad[dst]                 = d
+                    //
+                    // output gradients with their dependencies:
+                    //
+                    // grad[kcur] = grad[S1].T @ qcur
+                    // grad[S1]   = diag_mask_zero(grad[S3], P) * scale
+                    // grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
+                    // grad[S4]   = grad[S5] @ vcur
+                    // grad[S4]   = d[:D,id1,id2,id3] @ vcur
+                    // grad[qcur] = grad[S1]   @ kcur
+                    // grad[vcur] = grad[S5].T @ S4
+                    // grad[vcur] = d[:D,id1,id2,id3].T @ S4
+                    //
+                    // in post-order:
+                    //
+                    // S1         = qcur @ kcur.T
+                    // S2         = S1 * scale
+                    // S3         = diag_mask_inf(S2, P)
+                    // S4         = softmax(S3)
+                    // grad[S4]   = d[:D,id1,id2,id3] @ vcur
+                    // grad[S3]   = S4 * (grad[S4] - dot(S4, grad[S4]))
+                    // grad[S1]   = diag_mask_zero(grad[S3], P) * scale
+                    // grad[qcur] = grad[S1]   @ kcur
+                    // grad[kcur] = grad[S1].T @ qcur
+                    // grad[vcur] = d[:D,id1,id2,id3].T @ S4
+                    //
+                    // using less variables (SM=S4):
+                    //
+                    // S             = diag_mask_inf(qcur @ kcur.T * scale, P)
+                    // SM            = softmax(S)
+                    // S             = d[:D,iq1,iq2,iq3] @ vcur
+                    // dot_SM_gradSM = dot(SM, S)
+                    // S             = SM * (S - dot(SM, S))
+                    // S             = diag_mask_zero(S, P) * scale
+                    //
+                    // grad[q][:D,iq1,iq2,iq3] += S   @ kcur
+                    // grad[k][:D,:M,ik2,ik3]  += S.T @ qcur
+                    // grad[v][:M,:D,iv2,iv3]  += d[:D,id1,id2,id3].T @ SM
+                }
 
-            const size_t nbgq1 = nb0*neq0;
-            const size_t nbgq2 = nb0*neq0*neq1;
-            const size_t nbgq3 = nb0*neq0*neq1*neq2;
+                // S = gradSM = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3]
+                // S = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3]
+                // for ic:
+                //   S[:M] += vcur[:M,ic,iv2,iv3] * d[ic,id1,id2,id3]
+                // exclude known future zero S[..] values from operation
+                ggml_vec_set_f32(masked_begin, S, 0);
+                for (int64_t ic = 0; ic < D; ++ic) {
+                    ggml_vec_mad_f32(masked_begin,
+                            S,
+                             (float *) ((char *) v->data + (          ic*nbv1  + iv2*nbv2 + iv3*nbv3)),
+                            *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3)));
+                }
 
-            const size_t nbgk1 = nb0*nek0;
-            const size_t nbgk2 = nb0*nek0*nek1;
-            const size_t nbgk3 = nb0*nek0*nek1*neq2;
+                // S = SM * (S - dot(SM, S))
+                float dot_SM_gradSM = 0;
+                ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, SM, S);
+                ggml_vec_acc1_f32(M, S, -dot_SM_gradSM);
+                ggml_vec_mul_f32 (masked_begin, S, S, SM);
 
-            const size_t nbgv1 = nb0*nev0;
-            const size_t nbgv2 = nb0*nev0*nev1;
-            const size_t nbgv3 = nb0*nev0*nev1*neq2;
+                // S = diag_mask_zero(S, P) * scale
+                // already done by above ggml_vec_set_f32
 
-            // S    shape [M,1]
-            // SM   shape [M,1]
-            // kcur shape [D,M]
-            // qcur shape [D,1]
-            // vcur shape [M,D]
-            //
-            // grad[q][:D,iq1,iq2,iq3] += S @ kcur
-            // grad[q][:D,iq1,iq2,iq3] += shape[M,1] @ shape[D,M]
-            // grad[q][:D,iq1,iq2,iq3] += S[ic] * kcur[:D,ic]
-            //
-            //// grad[q][ic,iq1,iq2,iq3] += dot(kcur[:,ic],S.T)
-            //// grad[q][ic,iq1,iq2,iq3] += dot(k[:D,ic,iq2,iq3],S.T)
-            for (int64_t ic = 0; ic < M; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
+                // exclude known zero S[..] values from operation
+                ggml_vec_scale_f32(masked_begin, S, scale);
 
-                ggml_vec_mad_f32(D,
-                        (float *) ((char *) grad_q  + (i1*nbgq1  + i2*nbgq2  + i3*nbgq3)),
-                        (float *) ((char *) k->data + (ic*nbk1   + i2*nbk2   + i3*nbk3)),
-                        S[ic]);
-            }
+                // S    shape [M,1]
+                // SM   shape [M,1]
+                // kcur shape [D,M]
+                // qcur shape [D,1]
+                // vcur shape [M,D]
 
-            // grad[k][:D,:M,iq2,iq3] += S.T       @ qcur
-            // grad[k][:D,ic,iq2,iq3] += S.T[0,ic] * qcur[:D,0]
-            // grad[k][:D,ic,iq2,iq3] += S[ic]     * qcur[:D,0]
-            for (int64_t ic = 0; ic < M; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
+                // grad[q][:D,iq1,iq2,iq3] += S @ kcur
+                // grad[q][:D,iq1,iq2,iq3] += shape[M,1] @ shape[D,M]
+                // for ic:
+                //  grad[q][:D,iq1,iq2,iq3] += S[ic] * kcur[:D,ic,ik2,ik3]
+                // exclude known zero S[..] values from loop
+                for (int64_t ic = 0; ic < masked_begin; ++ic) {
+                    ggml_vec_mad_f32(D,
+                            (float *) ((char *) grad_q  + (iq1*nbgq1 + iq2*nbgq2  + iq3*nbgq3)),
+                            (float *) ((char *) k->data + (ic*nbk1   + ik2*nbk2   + ik3*nbk3)),
+                            S[ic]);
+                }
 
-                // ggml_vec_set_f32(D,
-                //         (float *) ((char *) grad_k  + (ic*nbgk1  + i2*nbgk2  + i3*nbgk3)),
-                //         0);
-                ggml_vec_mad_f32(D,
-                        (float *) ((char *) grad_k  + (ic*nbgk1  + i2*nbgk2  + i3*nbgk3)),
-                        (float *) ((char *) q->data + (i1*nbq1   + i2*nbq2   + i3*nbq3)),
-                        S[ic]);
-            }
+                // grad[k][:D,:M,iq2,iq3] += S.T @ qcur
+                // for ic:
+                //  grad[k][:D,ic,iq2,iq3] += S.T[0,ic] * qcur[:D,0]
+                //  grad[k][:D,ic,iq2,iq3] += S[ic]     * qcur[:D,0]
+                // exclude known zero S[..] values from loop
+                for (int64_t ic = 0; ic < masked_begin; ++ic) {
+                    ggml_vec_mad_f32(D,
+                            (float *) ((char *) grad_k  + (ic*nbgk1  + ik2*nbgk2  + ik3*nbgk3)),
+                            (float *) ((char *) q->data + (iq1*nbq1  + iq2*nbq2   + iq3*nbq3)),
+                            S[ic]);
+                }
 
-            // grad[v][:M,:D,iq2,iq3] += d[:D,iq1,iq2,iq3].T       @ SM
-            // grad[v][:M,ic,iq2,iq3] += d[:D,iq1,iq2,iq3].T[0,ic] * SM[:M]
-            // grad[v][:M,ic,iq2,iq3] += d[ic,iq1,iq2,iq3]         * SM[:M]
-            for (int64_t ic = 0; ic < D; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
-
-                // ggml_vec_set_f32(M,
-                //         (float *) ((char *) grad_v   + (          ic*nbgv1 + i2*nbgv2 + i3*nbgv3)),
-                //         0);
-                ggml_vec_mad_f32(M,
-                        (float *) ((char *) grad_v   + (          ic*nbgv1 + i2*nbgv2 + i3*nbgv3)),
-                        SM,
-                        *(float *) ((char *) d->data + (ic*nbd0 + i1*nbd1  + i2*nbd2  + i3*nbd3)));
+                // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T       @ SM
+                // for ic:
+                //  grad[v][:M,ic,iv2,iv3] += d[:D,id1,id2,id3].T[0,ic] * SM[:M]
+                //  grad[v][:M,ic,iv2,iv3] += d[ic,id1,id2,id3]         * SM[:M]
+                // exclude known zero SM[..] values from mad
+                for (int64_t ic = 0; ic < D; ++ic) {
+                    ggml_vec_mad_f32(masked_begin,
+                            (float *) ((char *) grad_v   + (          ic*nbgv1 + iv2*nbgv2 + iv3*nbgv3)),
+                            SM,
+                            *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2  + id3*nbd3)));
+                }
             }
         }
     }
@@ -14762,8 +15860,8 @@ static void ggml_compute_forward_win_part_f32(
         return;
     }
 
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
 
     const int32_t nep0 = ((const int32_t *)(dst->op_params))[0];
     const int32_t nep1 = ((const int32_t *)(dst->op_params))[1];
@@ -14824,8 +15922,8 @@ static void ggml_compute_forward_win_unpart_f32(
         return;
     }
 
-    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne);
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
 
     const int32_t w = ((const int32_t *)(dst->op_params))[0];
 
@@ -14942,7 +16040,7 @@ static void ggml_compute_forward_get_rel_pos_f16(
 
     // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L292-L322
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
     const int64_t w = ne1;
 
@@ -15275,6 +16373,8 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
     const int nc = src0->ne[0];
     const int nr = ggml_nrows(src0);
 
+    GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc));
+
     if (params->type == GGML_TASK_INIT) {
         if (ith == 0) {
             memset(sums, 0, sizeof(float) * (nth + nth * nc));
@@ -15286,7 +16386,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
         if (ith == 0) {
             float * dp = (float *) dst->data;
             ggml_vec_sum_f32(nth, dp, sums);
-            dp[0] *= -1.0f;
+            dp[0] *= -1.0f / (float) nr;
         }
         return;
     }
@@ -15303,7 +16403,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]);
         float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]);
-        float * st = (float *) params->wdata + nth + ith*nc;
+        float * st = ((float *) params->wdata) + nth + ith*nc;
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
@@ -15318,15 +16418,19 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             float max = -INFINITY;
             ggml_vec_max_f32(nc, &max, s0);
 
-            uint16_t scvt;
+            uint16_t scvt; UNUSED(scvt);
             for (int i = 0; i < nc; i++) {
                 if (s0[i] == -INFINITY) {
                     st[i] = 0.0f;
                 } else {
-                    // const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max);
+#ifndef GGML_CROSS_ENTROPY_EXP_FP16
+                    const float s = s0[i] - max;
+                    const float val = expf(s);
+#else
                     ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
                     memcpy(&scvt, &s, sizeof(scvt));
                     const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
+#endif
                     sum += (ggml_float)val;
                     st[i] = val;
                 }
@@ -15342,7 +16446,9 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
         ggml_vec_log_f32(nc, st, st);
         ggml_vec_mul_f32(nc, st, st, s1);
 
-        ggml_vec_sum_f32(nc, sums + ith, st);
+        float st_sum = 0;
+        ggml_vec_sum_f32(nc, &st_sum, st);
+        sums[ith] += st_sum;
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
@@ -15392,7 +16498,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         return;
     }
 
-    const float eps = 1e-9f;
+    const double eps = 1e-9;
 
     // TODO: handle transposed/permuted matrices
     const int64_t nc = src0->ne[0];
@@ -15411,7 +16517,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         float * ds0 = (float *)((char *) dst->data  + i1*dst->nb[1]);
         float * s0  = (float *)((char *) src0->data + i1*src0->nb[1]);
         float * s1  = (float *)((char *) src1->data + i1*src1->nb[1]);
-        float * sm  = (float *) params->wdata + ith*nc;
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
@@ -15420,54 +16525,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(s1[i]));
         }
 #endif
-        // step by step explanation:
-        {
-            //float * sums = (float *) params->wdata;
-
-            // forward pass with annotated gradients from backward pass
-            // (built by going in reverse operation order, adding to gradients of current operation args)
-            // st0 = exp(s0-max(s0))                                                       grad[st0] = grad[st1]*(1.0 - eps)/sum
-                                                          // from softmax_back:            grad[s0]  = st1_k * (grad[st1]_k - dot(st1, grad[st1]))
-            // ggml_vec_scale_f32(nc, st, sum);           // st1 = st0*/sum = softmax(s0)  grad[st1] = grad[st2]*(1.0 - eps)
-            // ggml_vec_scale_f32(nc, st, (1.0f - eps));  // st2 = st1*(1.0 - eps)         grad[st2] = grad[st3]
-            // ggml_vec_add1_f32(nc, st, st, eps);        // st3 = st2 + eps               grad[st3] = grad[st4]/st3
-            // ggml_vec_log_f32(nc, st, st);              // st4 = log(st3)                grad[st4] = grad[st5] * s1
-            // ggml_vec_mul_f32(nc, st, st, s1);          // st5 = st4 * s1                grad[st5] = grad[sums[ith]]
-            // ggml_vec_sum_f32(nc, sums + ith, st);      // sums[ith] = st5               grad[sums[ith]] = grad[cross_entropy_loss] = -grad[cel]
-
-            // substitute into grad[st1], because we can reuse softmax_back from this point on
-            // grad[st1] = -grad[cel]*s1*(1.0 - eps)/(eps + softmax(s0)*(1.0 - eps))
-            // postorder:
-            // grad[st1] := softmax(s0)
-            // grad[st1] := grad[st1]*(1.0 - eps)
-            // grad[st1] := grad[st1] + eps
-            // grad[st1] := s1 / grad[st1]
-            // grad[st1] := grad[st1]*(1.0-eps)*-grad[cel]
-
-            // src0 gradients by going through softmax_back
-            // grad[s0] = st1_k * (grad[st1]_k - dot(st1, grad[st1]))
-            //   from softmax_back:
-            //   dxk = yk * (dyk - dot(y, dy))
-            //   dot_y_dy := dot(y, dy)
-            //   dx := dy
-            //   dx := dx - dot_y_dy
-            //   dx := dx * y
-            //   postorder:
-            //   dot_st1_dst1 := dot(st1, grad[st1])
-            //   grad[s0] := grad[st1]
-            //   grad[s0] := grad[s0] - dot_st1_dst1
-            //   grad[s0] := grad[s0] * st1
-
-            // prepend postorder from grad[st1] directly using grad[s0] as memory location, as we will grad[s0] := grad[st1]
-            // sm           := softmax(s0)
-            // grad[s0]     := sm*(1.0 - eps)
-            // grad[s0]     := grad[s0] + eps
-            // grad[s0]     := s1 / grad[s0]
-            // grad[s0]     := grad[s0]*(1.0-eps)*-grad[cel]
-            // dot_st1_dst1 := dot(sm, grad[s0])
-            // grad[s0]     := grad[s0] - dot_st1_dst1
-            // grad[s0]     := grad[s0] * sm
-        }
 
         // soft_max
         ggml_float sum = 0.0;
@@ -15475,39 +16532,37 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             float max = -INFINITY;
             ggml_vec_max_f32(nc, &max, s0);
 
-            uint16_t scvt;
+            uint16_t scvt; UNUSED(scvt);
             for (int i = 0; i < nc; i++) {
                 if (s0[i] == -INFINITY) {
-                    sm[i] = 0.0f;
+                    ds0[i] = 0.0f;
                 } else {
-                    // const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max);
+#ifndef GGML_CROSS_ENTROPY_EXP_FP16
+                    const float s = s0[i] - max;
+                    const float val = expf(s);
+#else
                     ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
                     memcpy(&scvt, &s, sizeof(scvt));
                     const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
+#endif
                     sum += (ggml_float)val;
-                    sm[i] = val;
+                    ds0[i] = val;
                 }
             }
 
             assert(sum > 0.0);
-            sum = 1.0/sum;
+            sum = (1.0 - eps)/sum;
         }
 
-        float dot_st1_dst1 = 0;
-        ggml_vec_scale_f32(nc, sm, sum);
-        ggml_vec_cpy_f32  (nc, ds0, sm);
-        ggml_vec_scale_f32(nc, ds0, (1.0f - eps));
-        ggml_vec_add1_f32 (nc, ds0, ds0, eps);
-        ggml_vec_div_f32  (nc, ds0, s1, ds0);
-        ggml_vec_scale_f32(nc, ds0, -(1.0f - eps)*d[0]);
-        ggml_vec_dot_f32  (nc, &dot_st1_dst1, sm, ds0);
-        ggml_vec_acc1_f32 (nc, ds0, -dot_st1_dst1);
-        ggml_vec_mul_f32  (nc, ds0, ds0, sm);
+        // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr
+        ggml_vec_scale_f32(nc, ds0, sum);
+        ggml_vec_add1_f32(nc, ds0, ds0, eps);
+        ggml_vec_sub_f32(nc, ds0, ds0, s1);
+        ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr);
+
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
-            assert(!isnan(sm[i]));
-            assert(!isinf(sm[i]));
             assert(!isnan(ds0[i]));
             assert(!isinf(ds0[i]));
         }
@@ -15683,7 +16738,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_GET_ROWS_BACK:
             {
-                ggml_compute_forward_get_rows_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_get_rows_back(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_DIAG:
             {
@@ -15707,11 +16762,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_ROPE:
             {
-                ggml_compute_forward_rope(params, tensor->src[0], tensor);
+                ggml_compute_forward_rope(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_ROPE_BACK:
             {
-                ggml_compute_forward_rope_back(params, tensor->src[0], tensor);
+                ggml_compute_forward_rope_back(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_ALIBI:
             {
@@ -15725,13 +16780,25 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor);
             } break;
+        case GGML_OP_CONV_1D_STAGE_0:
+            {
+                ggml_compute_forward_conv_1d_stage_0(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
+        case GGML_OP_CONV_1D_STAGE_1:
+            {
+                ggml_compute_forward_conv_1d_stage_1(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor);
+            } break;
         case GGML_OP_CONV_2D:
             {
                 ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
-                ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_POOL_1D:
             {
@@ -15856,7 +16923,218 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
 
 ////////////////////////////////////////////////////////////////////////////////
 
-static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, bool inplace) {
+static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small");
+
+static size_t hash(void * p) {
+    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
+}
+
+static size_t hash_find(void * hash_table[], void * p) {
+    size_t h = hash(p);
+
+    // linear probing
+    size_t i = h;
+    while (hash_table[i] != NULL && hash_table[i] != p) {
+        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
+        if (i == h) {
+            // visited all hash table entries -> not found
+            return GGML_GRAPH_HASHTABLE_SIZE;
+        }
+    }
+    return i;
+}
+
+static bool hash_insert(void * hash_table[], void * p) {
+    size_t i = hash_find(hash_table, p);
+
+    GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
+
+    if (hash_table[i] == p) {
+        return true;
+    }
+
+    // insert
+    GGML_ASSERT(hash_table[i] == NULL);
+    hash_table[i] = p;
+    return false;
+}
+
+static bool hash_contains(void * hash_table[], void * p) {
+    size_t i = hash_find(hash_table, p);
+    return (i < GGML_GRAPH_HASHTABLE_SIZE) && (hash_table[i] == p);
+}
+
+struct hash_map {
+    void * keys[GGML_GRAPH_HASHTABLE_SIZE];
+    void * vals[GGML_GRAPH_HASHTABLE_SIZE];
+};
+
+static struct hash_map * new_hash_map(void) {
+    struct hash_map * result = malloc(sizeof(struct hash_map));
+    for (int i=0; i<GGML_GRAPH_HASHTABLE_SIZE; ++i) {
+        result->keys[i] = NULL;
+        result->vals[i] = NULL;
+    }
+    return result;
+}
+
+static void free_hash_map(struct hash_map * map) {
+    free(map);
+}
+
+// gradient checkpointing
+
+static struct ggml_tensor * ggml_recompute_graph_node(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * graph,
+        struct hash_map     * replacements,
+        struct ggml_tensor  * node) {
+
+    if (node == NULL) {
+        return NULL;
+    }
+
+    if (node->is_param) {
+        return node;
+    }
+
+    if (!hash_contains(graph->visited_hash_table, node)) {
+        return node;
+    }
+
+    int count_children = 0;
+    for (int k = 0; k < GGML_MAX_SRC; ++k) {
+        if (node->src[k]) {
+            ++count_children;
+        }
+    }
+
+    if (count_children == 0) {
+        return node;
+    }
+
+    size_t i = hash_find(replacements->keys, node);
+    GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
+    if (replacements->keys[i] == node) {
+        return (struct ggml_tensor *) replacements->vals[i];
+    }
+
+    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne);
+
+    // insert clone into replacements
+    GGML_ASSERT(replacements->keys[i] == NULL); // assert that we don't overwrite
+    replacements->keys[i] = node;
+    replacements->vals[i] = clone;
+
+    clone->op       = node->op;
+    clone->grad     = node->grad;
+    clone->is_param = node->is_param;
+    clone->extra    = node->extra;
+    for (int k = 0; k < GGML_MAX_DIMS; ++k) {
+        clone->nb[k] = node->nb[k];
+    }
+    for (int k = 0; k < GGML_MAX_SRC; ++k) {
+        clone->src[k] = ggml_recompute_graph_node(ctx, graph, replacements, node->src[k]);
+    }
+    if (node->view_src != NULL) {
+        clone->data = (node->view_src->data == NULL)
+                        ? NULL // view_src not yet allocated
+                        : (char *) node->view_src->data // view_src already allocated
+                                 + node->view_offs;
+        clone->view_src  = node->view_src;
+        clone->view_offs = node->view_offs;
+    }
+
+    GGML_ASSERT(sizeof(node->op_params) == sizeof(int32_t) * (GGML_MAX_OP_PARAMS / sizeof(int32_t)));
+    GGML_ASSERT(sizeof(node->name)      == GGML_MAX_NAME);
+    memcpy(clone->op_params, node->op_params, sizeof(node->op_params));
+    ggml_format_name(clone, "%s (clone)", ggml_get_name(node));
+
+    return clone;
+}
+
+void ggml_build_backward_gradient_checkpointing(
+        struct ggml_context   * ctx,
+        struct ggml_cgraph    * gf,
+        struct ggml_cgraph    * gb,
+        struct ggml_cgraph    * gb_tmp,
+        struct ggml_tensor  * * checkpoints,
+        int                     n_checkpoints) {
+    *gb_tmp = *gf;
+    ggml_build_backward_expand(ctx, gf, gb_tmp, true);
+
+    if (n_checkpoints <= 0) {
+        *gb = *gb_tmp;
+        return;
+    }
+
+    struct hash_map * replacements = new_hash_map();
+
+    // insert checkpoints in replacements
+    for (int i = 0; i < n_checkpoints; ++i) {
+        size_t k = hash_find(replacements->keys, checkpoints[i]);
+        GGML_ASSERT(k < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full
+        GGML_ASSERT(replacements->keys[k] == NULL); // assert that we don't overwrite
+        replacements->keys[k] = checkpoints[i];
+        replacements->vals[k] = checkpoints[i];
+    }
+
+    *gb = *gf;
+    // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes],
+    // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]),
+    // by recomputing them from checkpoints
+    for (int i = gf->n_nodes; i<gb_tmp->n_nodes; ++i) {
+        struct ggml_tensor * node = gb_tmp->nodes[i];
+        for (int k = 0; k < GGML_MAX_SRC; ++k) {
+            // insert new tensors recomputing src, reusing already made replacements,
+            // remember replacements: remember new tensors with mapping from corresponding gf nodes
+            // recurse for input tensors,
+            // unless (i.e. terminating when) input tensors are replacments (like checkpoints)
+            node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
+        }
+        // insert rewritten backward node with replacements made into resulting backward graph gb
+        ggml_build_forward_expand(gb, node);
+    }
+
+    free_hash_map(replacements);
+}
+
+// functions to change gradients considering the case that input a might be initial gradient with zero value
+
+static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) {
+    if (hash_contains(zero_table, a)) {
+        return b;
+    } else {
+        return ggml_add_impl(ctx, a, b, false);
+    }
+}
+
+static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, void * zero_table[]) {
+    if (hash_contains(zero_table, a)) {
+        struct ggml_tensor * a_zero = ggml_scale(ctx, a, ggml_new_f32(ctx, 0));
+        return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
+    } else {
+        return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
+    }
+}
+
+static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) {
+    if (hash_contains(zero_table, a)) {
+        return ggml_repeat(ctx, b, a);
+    } else {
+        return ggml_add1_impl(ctx, a, b, false);
+    }
+}
+
+static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) {
+    if (hash_contains(zero_table, a)) {
+        return ggml_neg(ctx, b);
+    } else {
+        return ggml_sub_impl(ctx, a, b, false);
+    }
+}
+
+static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, void * zero_table[]) {
     struct ggml_tensor * src0 = tensor->src[0];
     struct ggml_tensor * src1 = tensor->src[1];
 
@@ -15864,34 +17142,34 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_DUP:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
             } break;
         case GGML_OP_ADD:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
                 if (src1->grad) {
-                    src1->grad = ggml_add_impl(ctx, src1->grad, tensor->grad, inplace);
+                    src1->grad = ggml_add_or_set(ctx, src1->grad, tensor->grad, zero_table);
                 }
             } break;
         case GGML_OP_ADD1:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
                 if (src1->grad) {
-                    src1->grad = ggml_add_impl(ctx,
+                    src1->grad = ggml_add_or_set(ctx,
                         src1->grad,
                         ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean
-                        inplace);
+                        zero_table);
                 }
             } break;
         case GGML_OP_ACC:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
                 if (src1->grad) {
                     const size_t nb1     = ((int32_t *) tensor->op_params)[0];
@@ -15908,117 +17186,117 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         nb1, nb2, nb3, offset);
 
                     src1->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                             src1->grad,
                             ggml_reshape(ctx,
                                 ggml_cont(ctx, tensor_grad_view),
                                 src1->grad),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_SUB:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
                 if (src1->grad) {
-                    src1->grad = ggml_sub_impl(ctx, src1->grad, tensor->grad, inplace);
+                    src1->grad = ggml_sub_or_set(ctx, src1->grad, tensor->grad, zero_table);
                 }
             } break;
         case GGML_OP_MUL:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_mul(ctx, src1, tensor->grad),
-                                inplace);
+                                zero_table);
                 }
                 if (src1->grad) {
                     src1->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src1->grad,
                                 ggml_mul(ctx, src0, tensor->grad),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_DIV:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_div(ctx, tensor->grad, src1),
-                                inplace);
+                                zero_table);
                 }
                 if (src1->grad) {
                     src1->grad =
-                        ggml_sub_impl(ctx,
+                        ggml_sub_or_set(ctx,
                                 src1->grad,
                                 ggml_mul(ctx,
                                     tensor->grad,
                                     ggml_div(ctx, tensor, src1)),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_SQR:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_scale(ctx,
                                     ggml_mul(ctx, src0, tensor->grad),
                                     ggml_new_f32(ctx, 2.0f)),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_SQRT:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_scale(ctx,
                                     ggml_div(ctx,
                                         tensor->grad,
                                         tensor),
                                     ggml_new_f32(ctx, 0.5f)),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_LOG:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_div(ctx,
                                     tensor->grad,
                                     src0),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_SUM:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add1_impl(ctx,
+                        ggml_add1_or_set(ctx,
                                 src0->grad,
                                 tensor->grad,
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_SUM_ROWS:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_repeat(ctx,
                                     tensor->grad,
                                     src0->grad),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_MEAN:
@@ -16030,20 +17308,20 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 // necessary for llama
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_repeat_back(ctx, tensor->grad, src0->grad),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_REPEAT_BACK:
             {
                 if (src0->grad) {
                     // TODO: test this
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_repeat(ctx, tensor->grad, src0->grad),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_CONCAT:
@@ -16062,10 +17340,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 // necessary for llama
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx,
+                    float eps;
+                    memcpy(&eps, tensor->op_params, sizeof(float));
+
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
-                            ggml_rms_norm_back(ctx, src0, tensor->grad),
-                            inplace);
+                            ggml_rms_norm_back(ctx, src0, tensor->grad, eps),
+                            zero_table);
                 }
             } break;
         case GGML_OP_RMS_NORM_BACK:
@@ -16089,37 +17370,49 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
                 // ds1 = t.T.dot(dt)
 
-                // tensor.shape [m,p]
-                // src0.shape   [n,m]
-                // src1.shape   [n,p]
+                // tensor.shape [m,p,qq,rr]
+                // src0.shape   [n,m,q1,r1]
+                // src1.shape   [n,p,qq,rr]
 
                 // necessary for llama
                 if (src0->grad) {
+                    struct ggml_tensor * s1_tg =
+                        ggml_out_prod(ctx, // [n,m,qq,rr]
+                            src1,          // [n,p,qq,rr]
+                            tensor->grad); // [m,p,qq,rr]
+                    const int64_t qq = s1_tg->ne[2];
+                    const int64_t rr = s1_tg->ne[3];
+                    const int64_t q1 = src0->ne[2];
+                    const int64_t r1 = src0->ne[3];
+                    const bool ne2_broadcasted = qq > q1;
+                    const bool ne3_broadcasted = rr > r1;
+                    if (ne2_broadcasted || ne3_broadcasted) {
+                        // sum broadcast repetitions of s1_tg into shape of src0
+                        s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
+                    }
                     src0->grad =
-                        ggml_add_impl(ctx,
-                                src0->grad,
-                                ggml_out_prod(ctx, // [n,m]
-                                    src1,          // [n,p]
-                                    tensor->grad), // [m,p]
-                                inplace);
+                        ggml_add_or_set(ctx,
+                                src0->grad, // [n,m,q1,r1]
+                                s1_tg,      // [n,m,q1,r1]
+                                zero_table);
                 }
                 if (src1->grad) {
                     src1->grad =
-                        ggml_add_impl(ctx,
-                                src1->grad,
-                                // ggml_mul_mat(ctx,                   // [n,p]
-                                //     ggml_cont(ctx,                  // [m,n]
-                                //         ggml_transpose(ctx, src0)), // [m,n]
-                                //     tensor->grad),                  // [m,p]
+                        ggml_add_or_set(ctx,
+                                src1->grad,                            // [n,p,qq,rr]
+                                // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
+                                //     ggml_cont(ctx,                  // [m,n,q1,r1]
+                                //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
+                                //     tensor->grad),                  // [m,p,qq,rr]
 
                                 // // when src0 is bigger than tensor->grad (this is mostly the case in llama),
                                 // // avoid transpose of src0, rather transpose smaller tensor->grad
                                 // // and then use ggml_out_prod
-                                ggml_out_prod(ctx,                  // [n,p]
-                                    src0,                           // [n,m]
-                                    ggml_transpose(ctx,             // [p,m]
-                                        tensor->grad)),             // [m,p]
-                                inplace);
+                                ggml_out_prod(ctx,                  // [n,p,qq,rr]
+                                    src0,                           // [n,m,q1,r1]
+                                    ggml_transpose(ctx,             // [p,m,qq,rr]
+                                        tensor->grad)),             // [m,p,qq,rr]
+                                zero_table);
                 }
             } break;
         case GGML_OP_OUT_PROD:
@@ -16131,17 +17424,17 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_scale_impl(ctx, tensor->grad, src1, false),
-                            inplace);
+                            zero_table);
                 }
                 if (src1->grad) {
                     src1->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                             src1->grad,
                             ggml_sum(ctx, ggml_mul_impl(ctx, tensor->grad, src0, false)),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_SET:
@@ -16168,23 +17461,23 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 }
 
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                         src0->grad,
                         ggml_acc_impl(ctx,
                             tensor->grad,
                             ggml_neg(ctx, tensor_grad_view),
                             nb1, nb2, nb3, offset, false),
-                        inplace);
+                        zero_table);
                 }
 
                 if (src1->grad) {
                     src1->grad =
-                        ggml_add_impl(ctx,
+                        ggml_add_or_set(ctx,
                             src1->grad,
                             ggml_reshape(ctx,
                                 ggml_cont(ctx, tensor_grad_view),
                                 src1->grad),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_CPY:
@@ -16195,7 +17488,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // tensor = src0 * 1 + src1 * 0
                 if (src0->grad) {
                     // dsrc0 = dtensor * 1
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
                 if (src1->grad) {
                     // dsrc1 = dtensor * 0 -> noop
@@ -16207,7 +17500,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad) {
                     GGML_ASSERT(ggml_is_contiguous(src0->grad));
                     GGML_ASSERT(ggml_is_contiguous(tensor->grad));
-                    src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
                 }
             } break;
         case GGML_OP_RESHAPE:
@@ -16215,9 +17508,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
-                            ggml_reshape(ctx, tensor->grad, src0->grad),
-                        inplace);
+                        ggml_add_or_set(ctx, src0->grad,
+                            ggml_reshape(ctx,
+                                ggml_is_contiguous(tensor->grad)
+                                    ? tensor->grad
+                                    : ggml_cont(ctx, tensor->grad),
+                                src0->grad),
+                        zero_table);
                 }
             } break;
         case GGML_OP_VIEW:
@@ -16246,7 +17543,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         nb3 = (nb3 / n0) * ng;
                     }
 
-                    src0->grad = ggml_acc_impl(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, inplace);
+                    src0->grad = ggml_acc_or_set(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, zero_table);
                 }
             } break;
         case GGML_OP_PERMUTE:
@@ -16264,14 +17561,14 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     axes_backward[axis2] = 2;
                     axes_backward[axis3] = 3;
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
                             ggml_permute(ctx,
                                 tensor->grad,
                                 axes_backward[0],
                                 axes_backward[1],
                                 axes_backward[2],
                                 axes_backward[3]),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_TRANSPOSE:
@@ -16279,9 +17576,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
                             ggml_transpose(ctx, tensor->grad),
-                        inplace);
+                        zero_table);
                 }
             } break;
         case GGML_OP_GET_ROWS:
@@ -16289,9 +17586,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama (only for tokenizer)
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
+                            // last ggml_get_rows_back argument src0->grad is only
+                            // necessary to setup correct output shape
                             ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad),
-                        inplace);
+                        zero_table);
                 }
                 if (src1->grad) {
                     // noop
@@ -16311,9 +17610,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad) {
                     const int n_past = ((int32_t *) tensor->op_params)[0];
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
                             ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        inplace);
+                        zero_table);
                 }
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
@@ -16322,9 +17621,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad) {
                     const int n_past = ((int32_t *) tensor->op_params)[0];
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
                             ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        inplace);
+                        zero_table);
                 }
             } break;
         case GGML_OP_SOFT_MAX:
@@ -16332,9 +17631,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_impl(ctx, src0->grad,
+                        ggml_add_or_set(ctx, src0->grad,
                             ggml_soft_max_back(ctx, tensor->grad, tensor),
-                        inplace);
+                        zero_table);
                 }
 
             } break;
@@ -16346,7 +17645,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 // necessary for llama
                 if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
+                    //const int n_past = ((int32_t *) tensor->op_params)[0];
                     const int n_dims = ((int32_t *) tensor->op_params)[1];
                     const int mode   = ((int32_t *) tensor->op_params)[2];
                     const int n_ctx  = ((int32_t *) tensor->op_params)[3];
@@ -16359,11 +17658,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     memcpy(&xpos_base,  (int32_t *) tensor->op_params + 6, sizeof(float));
                     memcpy(&xpos_down,  (int32_t *) tensor->op_params + 7, sizeof(bool));
 
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_rope_back(ctx,
                                 tensor->grad,
-                                n_past,
+                                src1,
                                 n_dims,
                                 mode,
                                 n_ctx,
@@ -16371,13 +17670,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 freq_scale,
                                 xpos_base,
                                 xpos_down),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_ROPE_BACK:
             {
                 if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
+                    //const int n_past = ((int32_t *) tensor->op_params)[0];
                     const int n_dims = ((int32_t *) tensor->op_params)[1];
                     const int mode   = ((int32_t *) tensor->op_params)[2];
                     const int n_ctx  = ((int32_t *) tensor->op_params)[3];
@@ -16390,11 +17689,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     memcpy(&xpos_base,  (int32_t *) tensor->op_params + 6, sizeof(float));
                     memcpy(&xpos_down,  (int32_t *) tensor->op_params + 7, sizeof(bool));
 
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_rope_impl(ctx,
                                 tensor->grad,
-                                n_past,
+                                src1,
                                 n_dims,
                                 mode,
                                 n_ctx,
@@ -16403,7 +17702,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 xpos_base,
                                 xpos_down,
                                 false),
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_ALIBI:
@@ -16418,10 +17717,22 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_CONV_1D_STAGE_0:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
+        case GGML_OP_CONV_1D_STAGE_1:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_CONV_2D:
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -16454,145 +17765,42 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             masked);
                 }
 
-                if (src0->grad) {
-                    struct ggml_tensor * grad_q = NULL;
-                    const size_t nb0    = flash_grad->nb[0];
-                    const size_t offset = 0;
-                    switch(src0->n_dims) {
-                        case 2:
-                            {
-                                grad_q = ggml_view_2d(ctx,
-                                    flash_grad,
-                                    src0->ne[0],
-                                    src0->ne[1],
-                                    nb0*src0->ne[0],
-                                    offset);
-                            } break;
-                        case 3:
-                            {
-                                grad_q = ggml_view_3d(ctx,
-                                    flash_grad,
-                                    src0->ne[0],
-                                    src0->ne[1],
-                                    src0->ne[2],
-                                    nb0*src0->ne[0],
-                                    nb0*src0->ne[0]*src0->ne[1],
-                                    offset);
-                            } break;
-                        case 4:
-                            {
-                                grad_q = ggml_view_4d(ctx,
-                                    flash_grad,
-                                    src0->ne[0],
-                                    src0->ne[1],
-                                    src0->ne[2],
-                                    src0->ne[3],
-                                    nb0*src0->ne[0],
-                                    nb0*src0->ne[0]*src0->ne[1],
-                                    nb0*src0->ne[0]*src0->ne[1]*src0->ne[2],
-                                    offset);
-                            } break;
-                    }
+                struct ggml_tensor * src2 = tensor->src[2];
+                const int64_t elem_q = ggml_nelements(src0);
+                const int64_t elem_k = ggml_nelements(src1);
+                const int64_t elem_v = ggml_nelements(src2);
 
-                    src0->grad = ggml_add_impl(ctx,
+                enum ggml_type result_type = flash_grad->type;
+                GGML_ASSERT(ggml_blck_size(result_type) == 1);
+                const size_t tsize = ggml_type_size(result_type);
+
+                const size_t offs_q = 0;
+                const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
+                const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
+
+                if (src0->grad) {
+                    struct ggml_tensor * view_q = ggml_view_1d(ctx, flash_grad, elem_q, offs_q);
+                    struct ggml_tensor * grad_q = ggml_reshape(ctx, view_q, src0);
+                    src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             grad_q,
-                            inplace);
+                            zero_table);
                 }
-
                 if (src1->grad) {
-                    struct ggml_tensor * grad_k = NULL;
-                    const size_t nb0    = flash_grad->nb[0];
-                    const size_t offset = nb0*src0->ne[0]*src0->ne[1]*src0->ne[2]*src0->ne[3];
-                    switch(src1->n_dims) {
-                        case 2:
-                            {
-                                grad_k = ggml_view_2d(ctx,
-                                    flash_grad,
-                                    src1->ne[0],
-                                    src1->ne[1],
-                                    nb0*src1->ne[0],
-                                    offset);
-                            } break;
-                        case 3:
-                            {
-                                grad_k = ggml_view_3d(ctx,
-                                    flash_grad,
-                                    src1->ne[0],
-                                    src1->ne[1],
-                                    src1->ne[2],
-                                    nb0*src1->ne[0],
-                                    nb0*src1->ne[0]*src1->ne[1],
-                                    offset);
-                            } break;
-                        case 4:
-                            {
-                                grad_k = ggml_view_4d(ctx,
-                                    flash_grad,
-                                    src1->ne[0],
-                                    src1->ne[1],
-                                    src1->ne[2],
-                                    src1->ne[3],
-                                    nb0*src1->ne[0],
-                                    nb0*src1->ne[0]*src1->ne[1],
-                                    nb0*src1->ne[0]*src1->ne[1]*src1->ne[2],
-                                    offset);
-                            } break;
-                    }
-
-                    src1->grad = ggml_add_impl(ctx,
+                    struct ggml_tensor * view_k = ggml_view_1d(ctx, flash_grad, elem_k, offs_k);
+                    struct ggml_tensor * grad_k = ggml_reshape(ctx, view_k, src1);
+                    src1->grad = ggml_add_or_set(ctx,
                             src1->grad,
                             grad_k,
-                            inplace);
+                            zero_table);
                 }
-
-                struct ggml_tensor * opt0 = tensor->src[2];
-
-                if (opt0->grad) {
-                    struct ggml_tensor * grad_v = NULL;
-                    const size_t nb0    = flash_grad->nb[0];
-                    const size_t offset = nb0*src0->ne[0]*src0->ne[1]*src0->ne[2]*src0->ne[3]
-                                        + nb0*src1->ne[0]*src1->ne[1]*src1->ne[2]*src1->ne[3];
-                    switch(opt0->n_dims) {
-                        case 2:
-                            {
-                                grad_v = ggml_view_2d(ctx,
-                                    flash_grad,
-                                    opt0->ne[0],
-                                    opt0->ne[1],
-                                    nb0*opt0->ne[0],
-                                    offset);
-                            } break;
-                        case 3:
-                            {
-                                grad_v = ggml_view_3d(ctx,
-                                    flash_grad,
-                                    opt0->ne[0],
-                                    opt0->ne[1],
-                                    opt0->ne[2],
-                                    nb0*opt0->ne[0],
-                                    nb0*opt0->ne[0]*opt0->ne[1],
-                                    offset);
-                            } break;
-                        case 4:
-                            {
-                                grad_v = ggml_view_4d(ctx,
-                                    flash_grad,
-                                    opt0->ne[0],
-                                    opt0->ne[1],
-                                    opt0->ne[2],
-                                    opt0->ne[3],
-                                    nb0*opt0->ne[0],
-                                    nb0*opt0->ne[0]*opt0->ne[1],
-                                    nb0*opt0->ne[0]*opt0->ne[1]*opt0->ne[2],
-                                    offset);
-                            } break;
-                    }
-
-                    opt0->grad = ggml_add_impl(ctx,
-                            opt0->grad,
+                if (src2->grad) {
+                    struct ggml_tensor * view_v = ggml_view_1d(ctx, flash_grad, elem_v, offs_v);
+                    struct ggml_tensor * grad_v = ggml_reshape(ctx, view_v, src2);
+                    src2->grad = ggml_add_or_set(ctx,
+                            src2->grad,
                             grad_v,
-                            inplace);
+                            zero_table);
                 }
             } break;
         case GGML_OP_FLASH_FF:
@@ -16612,12 +17820,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         {
                             if (src0->grad) {
                                 src0->grad =
-                                    ggml_add_impl(ctx,
+                                    ggml_add_or_set(ctx,
                                             src0->grad,
                                             ggml_mul(ctx,
                                                 ggml_sgn(ctx, src0),
                                                 tensor->grad),
-                                            inplace);
+                                            zero_table);
                             }
                         } break;
                     case GGML_UNARY_OP_SGN:
@@ -16629,7 +17837,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     case GGML_UNARY_OP_NEG:
                         {
                             if (src0->grad) {
-                                src0->grad = ggml_sub_impl(ctx, src0->grad, tensor->grad, inplace);
+                                src0->grad = ggml_sub_or_set(ctx, src0->grad, tensor->grad, zero_table);
                             }
                         } break;
                     case GGML_UNARY_OP_STEP:
@@ -16649,12 +17857,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     case GGML_UNARY_OP_RELU:
                         {
                             if (src0->grad) {
-                                src0->grad = ggml_add_impl(ctx,
+                                src0->grad = ggml_add_or_set(ctx,
                                         src0->grad,
                                         ggml_mul(ctx,
                                             ggml_step(ctx, src0),
                                             tensor->grad),
-                                        inplace);
+                                        zero_table);
                             }
                         } break;
                     case GGML_UNARY_OP_GELU:
@@ -16669,10 +17877,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         {
                             // necessary for llama
                             if (src0->grad) {
-                                src0->grad = ggml_add_impl(ctx,
+                                src0->grad = ggml_add_or_set(ctx,
                                         src0->grad,
                                         ggml_silu_back(ctx, src0, tensor->grad),
-                                        inplace);
+                                        zero_table);
                             }
                         } break;
                     default:
@@ -16695,13 +17903,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_impl(ctx,
+                    src0->grad = ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_cross_entropy_loss_back(ctx,
                                     src0,
                                     src1,
                                     tensor->grad),
-                                inplace);
+                                zero_table);
                 }
             } break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
@@ -16717,34 +17925,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 GGML_ASSERT(false);
             } break;
     }
-}
 
-static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small");
-
-static size_t hash(void * p) {
-    return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE;
-}
-
-static bool hash_insert(void * hash_table[], void * p) {
-    size_t h = hash(p);
-
-    // linear probing
-    size_t i = h;
-    while (hash_table[i] != NULL && hash_table[i] != p) {
-        i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE;
-        if (i == h) {
-            // hash table is full
-            GGML_ASSERT(false);
+    for (int i = 0; i < GGML_MAX_SRC; ++i) {
+        if (tensor->src[i] && tensor->src[i]->grad) {
+            GGML_ASSERT(ggml_are_same_shape(tensor->src[i], tensor->src[i]->grad));
         }
     }
-
-    if (hash_table[i] == p) {
-        return true;
-    }
-
-    // insert
-    hash_table[i] = p;
-    return false;
 }
 
 static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
@@ -16762,8 +17948,12 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
     }
 
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
-        if (node->src[i]) {
-            ggml_visit_parents(cgraph, node->src[i]);
+        const int k =
+            (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
+            (cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) :
+            /* unknown order, just fall back to using i*/ i;
+        if (node->src[k]) {
+            ggml_visit_parents(cgraph, node->src[k]);
         }
     }
 
@@ -16822,6 +18012,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) {
         /*.grads        =*/ { NULL },
         /*.leafs        =*/ { NULL },
         /*.hash_table   =*/ { NULL },
+        /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
@@ -16832,9 +18023,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) {
     return result;
 }
 
-struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) {
-    struct ggml_cgraph result = *gf;
-
+void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) {
     GGML_ASSERT(gf->n_nodes > 0);
 
     // if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph
@@ -16849,24 +18038,40 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
         }
     }
 
-    for (int i = gf->n_nodes - 1; i >= 0; i--) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        // because we detached the grad nodes from the original graph, we can afford inplace operations
-        if (node->grad) {
-            ggml_compute_backward(ctx, node, keep);
+    // remember original gradients which start with zero values
+    void ** zero_table = malloc(sizeof(void *) * GGML_GRAPH_HASHTABLE_SIZE);
+    memset(zero_table, 0, sizeof(void*) * GGML_GRAPH_HASHTABLE_SIZE);
+    for (int i = 0; i < gf->n_nodes; i++) {
+        if (gf->grads[i]) {
+            hash_insert(zero_table, gf->grads[i]);
         }
     }
 
     for (int i = gf->n_nodes - 1; i >= 0; i--) {
         struct ggml_tensor * node = gf->nodes[i];
 
+        // inplace operations to add gradients are not created by ggml_compute_backward
+        // use allocator to automatically make inplace operations
+        if (node->grad) {
+            ggml_compute_backward(ctx, node, zero_table);
+        }
+    }
+
+    for (int i = 0; i < gf->n_nodes; i++) {
+        struct ggml_tensor * node = gf->nodes[i];
+
         if (node->is_param) {
             GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
-            ggml_build_forward_expand(&result, node->grad);
+            ggml_build_forward_expand(gb, node->grad);
         }
     }
 
+    free(zero_table);
+}
+
+struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) {
+    struct ggml_cgraph result = *gf;
+    ggml_build_backward_expand(ctx, gf, &result, keep);
     return result;
 }
 
@@ -16881,6 +18086,7 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
         /*.grads        =*/ { NULL },
         /*.leafs        =*/ { NULL },
         /*.hash_table   =*/ { NULL },
+        /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
@@ -17132,10 +18338,18 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         } else {
             // wait for other threads to finish
             const int last = node_n;
-            do {
-                //sched_yield();
+            while (true) {
+                // TODO: this sched_yield can have significant impact on the performance - either positive or negative
+                //       depending on the workload and the operating system.
+                //       since it is not clear what is the best approach, it should potentially become user-configurable
+                //       ref: https://github.com/ggerganov/ggml/issues/291
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+                sched_yield();
+#endif
+
                 node_n = atomic_load(&state->shared->node_n);
-            } while (node_n == last);
+                if (node_n != last) break;
+            };
         }
 
         // check if we should stop
@@ -17263,7 +18477,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_CONCAT:
             case GGML_OP_MUL_MAT:
-            case GGML_OP_OUT_PROD:
                 {
                     n_tasks = n_threads;
 
@@ -17305,6 +18518,18 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         cur = 0;
                     }
 
+                    work_size = MAX(work_size, cur);
+                } break;
+            case GGML_OP_OUT_PROD:
+                {
+                    n_tasks = n_threads;
+
+                    size_t cur = 0;
+
+                    if (ggml_is_quantized(node->src[0]->type)) {
+                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
+                    }
+
                     work_size = MAX(work_size, cur);
                 } break;
             case GGML_OP_SCALE:
@@ -17349,21 +18574,68 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     GGML_ASSERT(node->src[1]->ne[2] == 1);
                     GGML_ASSERT(node->src[1]->ne[3] == 1);
 
+                    const int64_t ne00 = node->src[0]->ne[0];
+                    const int64_t ne01 = node->src[0]->ne[1];
+                    const int64_t ne02 = node->src[0]->ne[2];
+
+                    const int64_t ne10 = node->src[1]->ne[0];
+                    const int64_t ne11 = node->src[1]->ne[1];
+
+                    const int64_t ne0 = node->ne[0];
+                    const int64_t ne1 = node->ne[1];
+                    const int64_t nk  = ne00;
+                    const int64_t ew0 = nk * ne01;
+
+                    UNUSED(ne02);
+                    UNUSED(ne10);
+                    UNUSED(ne11);
+
                     size_t cur = 0;
-                    const int nk = node->src[0]->ne[0];
 
                     if (node->src[0]->type == GGML_TYPE_F16 &&
-                            node->src[1]->type == GGML_TYPE_F32) {
-                        cur = sizeof(ggml_fp16_t)*(
-                                nk*ggml_up32(node->src[0]->ne[1])*node->src[0]->ne[2] +
-                                ( 2*(nk/2) + node->src[1]->ne[0])*node->src[1]->ne[1]
-                                );
+                        node->src[1]->type == GGML_TYPE_F32) {
+                        cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0);
                     } else if (node->src[0]->type == GGML_TYPE_F32 &&
-                            node->src[1]->type == GGML_TYPE_F32) {
-                        cur = sizeof(float)*(
-                                nk*ggml_up32(node->src[0]->ne[1])*node->src[0]->ne[2] +
-                                ( 2*(nk/2) + node->src[1]->ne[0])*node->src[1]->ne[1]
-                                );
+                               node->src[1]->type == GGML_TYPE_F32) {
+                        cur = sizeof(float)*(ne0*ne1*ew0);
+                    } else {
+                        GGML_ASSERT(false);
+                    }
+
+                    work_size = MAX(work_size, cur);
+                } break;
+            case GGML_OP_CONV_1D_STAGE_0:
+                {
+                    n_tasks = n_threads;
+                } break;
+            case GGML_OP_CONV_1D_STAGE_1:
+                {
+                    n_tasks = n_threads;
+                } break;
+            case GGML_OP_CONV_TRANSPOSE_1D:
+                {
+                    n_tasks = n_threads;
+
+                    GGML_ASSERT(node->src[0]->ne[3] == 1);
+                    GGML_ASSERT(node->src[1]->ne[2] == 1);
+                    GGML_ASSERT(node->src[1]->ne[3] == 1);
+
+                    const int64_t ne00 = node->src[0]->ne[0];  // K
+                    const int64_t ne01 = node->src[0]->ne[1];  // Cout
+                    const int64_t ne02 = node->src[0]->ne[2];  // Cin
+
+                    const int64_t ne10 = node->src[1]->ne[0];  // L
+                    const int64_t ne11 = node->src[1]->ne[1];  // Cin
+
+                    size_t cur = 0;
+                    if (node->src[0]->type == GGML_TYPE_F16 &&
+                        node->src[1]->type == GGML_TYPE_F32) {
+                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
+                        cur += sizeof(ggml_fp16_t)*ne10*ne11;
+                    } else if (node->src[0]->type == GGML_TYPE_F32 &&
+                               node->src[1]->type == GGML_TYPE_F32) {
+                        cur += sizeof(float)*ne00*ne01*ne02;
+                        cur += sizeof(float)*ne10*ne11;
                     } else {
                         GGML_ASSERT(false);
                     }
@@ -17542,10 +18814,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
             case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
                 {
                     n_tasks = n_threads;
-
-                    size_t cur = ggml_type_size(node->type)*node->src[0]->ne[0]*n_tasks;
-
-                    work_size = MAX(work_size, cur);
                 } break;
             case GGML_OP_NONE:
                 {
@@ -18190,10 +19458,11 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * node = cgraph->leafs[i];
 
-        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n",
+        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n",
                 i,
                 node->ne[0], node->ne[1],
-                ggml_op_name(node->op));
+                ggml_op_name(node->op),
+                ggml_get_name(node));
     }
 
     for (int i = 0; i < GGML_OP_COUNT; i++) {
@@ -18401,7 +19670,7 @@ static void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float *
 }
 
 static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g) {
-    int i = 0;
+    int64_t i = 0;
     for (int p = 0; p < np; ++p) {
         const int64_t ne = ggml_nelements(ps[p]) ;
         // TODO: add function to get all elements at once
@@ -18411,6 +19680,17 @@ static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g
     }
 }
 
+static void ggml_opt_acc_grad(int np, struct ggml_tensor * const ps[], float * g, float scale) {
+    int64_t i = 0;
+    for (int p = 0; p < np; ++p) {
+        const int64_t ne = ggml_nelements(ps[p]) ;
+        // TODO: add function to get all elements at once
+        for (int64_t j = 0; j < ne; ++j) {
+            g[i++] += ggml_get_f32_1d(ps[p]->grad, j) * scale;
+        }
+    }
+}
+
 //
 // ADAM
 //
@@ -18423,14 +19703,16 @@ static enum ggml_opt_result ggml_opt_adam(
         struct ggml_opt_params params,
         struct ggml_tensor * f,
         struct ggml_cgraph * gf,
-        struct ggml_cgraph * gb) {
+        struct ggml_cgraph * gb,
+        ggml_opt_callback callback,
+        void * callback_data) {
     GGML_ASSERT(ggml_is_scalar(f));
 
     // these will store the parameters we want to optimize
     struct ggml_tensor * ps[GGML_MAX_PARAMS];
 
     int np = 0;
-    int nx = 0;
+    int64_t nx = 0;
     for (int i = 0; i < gf->n_nodes; ++i) {
         if (gf->nodes[i]->is_param) {
             GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
@@ -18449,38 +19731,56 @@ static enum ggml_opt_result ggml_opt_adam(
     }
 
     // constants
-    const float sched = params.adam.sched;
-    const float decay = params.adam.decay * sched;
-    const float alpha = params.adam.alpha * sched;
+    float sched = params.adam.sched;
+    const float alpha = params.adam.alpha;
+    const float decay = params.adam.decay * alpha;
     const float beta1 = params.adam.beta1;
     const float beta2 = params.adam.beta2;
     const float eps   = params.adam.eps;
+    const float gclip = params.adam.gclip;
+    const int decay_min_ndim = params.adam.decay_min_ndim;
+    const int n_accum = MAX(1, params.n_gradient_accumulation);
+    const float accum_norm = 1.0f / (float) n_accum;
 
-    float * x  = opt->adam.x->data;  // view of the parameters
-    float * g1 = opt->adam.g1->data; // gradient
-    float * g2 = opt->adam.g2->data; // gradient squared
+    float * g  = opt->adam.g->data;  // gradients
     float * m  = opt->adam.m->data;  // first moment
     float * v  = opt->adam.v->data;  // second moment
-    float * mh = opt->adam.mh->data; // first moment hat
-    float * vh = opt->adam.vh->data; // second moment hat
 
     float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values
 
-    // update view
-    ggml_opt_get_params(np, ps, x);
+    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
+    cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
+
+    bool cancel = false;
 
     // compute the function value
-    ggml_graph_reset  (gf);
-    ggml_set_f32      (f->grad, 1.0f);
+    float fx = 0;
+    ggml_set_zero(opt->adam.g);
+    for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+        if (callback) {
+            callback(callback_data, accum_step, &sched, &cancel);
+            if (cancel) {
+                return GGML_OPT_CANCEL;
+            }
+        }
+        // ggml_graph_reset  (gf);
+        ggml_set_f32      (f->grad, 1.0f);
+        ggml_graph_compute(gb, &cplan);
+        ggml_opt_acc_grad(np, ps, g, accum_norm);
+        fx += ggml_get_f32_1d(f, 0);
+    }
+    fx *= accum_norm;
 
-    ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
-
-    opt->adam.fx_prev = ggml_get_f32_1d(f, 0);
+    opt->adam.fx_prev = fx;
     opt->adam.fx_best = opt->adam.fx_prev;
     if (pf) {
         pf[opt->iter % params.past] = opt->adam.fx_prev;
     }
 
+    opt->loss_before = opt->adam.fx_prev;
+    opt->loss_after  = opt->adam.fx_prev;
+
     // initialize
     if (opt->just_initialized) {
         opt->adam.n_no_improvement = 0;
@@ -18513,50 +19813,58 @@ static enum ggml_opt_result ggml_opt_adam(
         UNUSED(t_start_cpu);
 
         {
-            // update the gradient
-            ggml_opt_get_grad(np, ps, g1);
-
-            // m_t = beta1*m_t-1 + (1 - beta1)*g_t
-            ggml_vec_scale_f32(nx, m, beta1);
-            ggml_vec_mad_f32  (nx, m, g1, 1.0f - beta1);
-
-            // g2 = g1^2
-            ggml_vec_sqr_f32  (nx, g2, g1);
-
-            // v_t = beta2*v_t-1 + (1 - beta2)*g_t^2
-            ggml_vec_scale_f32(nx, v, beta2);
-            ggml_vec_mad_f32  (nx, v, g2, 1.0f - beta2);
-
-            // m^hat = m_t / (1 - beta1^t)
-            // v^hat = v_t / (1 - beta2^t)
-            // x_t = x_t-1 - sched*(alpha*m^hat/(sqrt(v^hat) + eps) + decay*x_t-1)
-            // x_t = x_t-1 - sched*alpha*m^hat/(sqrt(v^hat) + eps) - sched*decay*x_t-1
-            // x_t = x_t-1*(1-sched*decay) - sched*alpha*m^hat/(sqrt(v^hat) + eps)
-            // x_t = x_t-1*(1-sched*decay) + sched*decay*(-alpha/decay)*m^hat/(sqrt(v^hat) + eps)
-            // x_t = mix(x_t-1, (-alpha/decay)*m^hat/(sqrt(v^hat) + eps), sched*decay)
-            ggml_vec_cpy_f32  (nx, mh, m);
-            ggml_vec_cpy_f32  (nx, vh, v);
-
-            ggml_vec_scale_f32(nx, mh, alpha/(1.0f - powf(beta1, opt->iter)));
-            ggml_vec_scale_f32(nx, vh,  1.0f/(1.0f - powf(beta2, opt->iter)));
-
-            ggml_vec_sqrt_f32 (nx, vh, vh);
-            ggml_vec_acc1_f32 (nx, vh, eps);
-
-            ggml_vec_div_f32  (nx, mh, mh, vh);
-            ggml_vec_scale_f32(nx, x,  1.0f - decay);
-            ggml_vec_sub_f32  (nx, x,  x,  mh);
-
-            // update the parameters
-            ggml_opt_set_params(np, ps, x);
+            float gnorm = 1.0f;
+            if (gclip > 0.0f) {
+                // gradient clipping
+                ggml_float sum = 0.0;
+                for (int64_t i = 0; i < nx; ++i) {
+                    sum += (ggml_float)(g[i]*g[i]);
+                }
+                ggml_float norm = sqrt(sum);
+                if (norm > (ggml_float) gclip) {
+                    gnorm = (float) ((ggml_float) gclip / norm);
+                }
+            }
+            const float beta1h = alpha*sched/(1.0f - powf(beta1, opt->iter));
+            const float beta2h =        1.0f/(1.0f - powf(beta2, opt->iter));
+            int64_t i = 0;
+            for (int p = 0; p < np; ++p) {
+                const int64_t ne = ggml_nelements(ps[p]);
+                const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
+                for (int64_t j = 0; j < ne; ++j) {
+                    float x  = ggml_get_f32_1d(ps[p], j);
+                    float g_ = g[i]*gnorm;
+                    m[i] = m[i]*beta1 +    g_*(1.0f - beta1);
+                    v[i] = v[i]*beta2 + g_*g_*(1.0f - beta2);
+                    float mh = m[i]*beta1h;
+                    float vh = v[i]*beta2h;
+                    vh = sqrtf(vh) + eps;
+                    x  = x*(1.0f - p_decay) - mh/vh;
+                    ggml_set_f32_1d(ps[p], j, x);
+                    ++i;
+                }
+            }
         }
 
-        ggml_graph_reset  (gf);
-        ggml_set_f32      (f->grad, 1.0f);
+        fx = 0;
+        ggml_set_zero(opt->adam.g);
+        for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+            if (callback) {
+                callback(callback_data, accum_step, &sched, &cancel);
+                if (cancel) {
+                    return GGML_OPT_CANCEL;;
+                }
+            }
+            // ggml_graph_reset  (gf);
+            ggml_set_f32      (f->grad, 1.0f);
+            ggml_graph_compute(gb, &cplan);
+            ggml_opt_acc_grad(np, ps, g, accum_norm);
+            fx += ggml_get_f32_1d(f, 0);
+        }
+        fx *= accum_norm;
 
-        ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
+        opt->loss_after = fx;
 
-        const float fx = ggml_get_f32_1d(f, 0);
 
         // check convergence
         if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) {
@@ -18625,7 +19933,6 @@ struct ggml_lbfgs_iteration_data {
 };
 
 static enum ggml_opt_result linesearch_backtracking(
-        struct ggml_context * ctx,
         const struct ggml_opt_params * params,
         int nx,
         float * x,
@@ -18635,10 +19942,13 @@ static enum ggml_opt_result linesearch_backtracking(
         float * step,
         const float * xp,
         struct ggml_tensor * f,
-        struct ggml_cgraph * gf,
         struct ggml_cgraph * gb,
+        struct ggml_cplan  * cplan,
         const int np,
-        struct ggml_tensor * ps[]) {
+        struct ggml_tensor * ps[],
+        bool * cancel,
+        ggml_opt_callback callback,
+        void * callback_data) {
     int count = 0;
 
     float width  = 0.0f;
@@ -18650,6 +19960,9 @@ static enum ggml_opt_result linesearch_backtracking(
     const float dec = 0.5f;
     const float inc = 2.1f;
 
+    const int n_accum = MAX(1, params->n_gradient_accumulation);
+    const float accum_norm = 1.0f / (float) n_accum;
+
     if (*step <= 0.f) {
         return GGML_LINESEARCH_INVALID_PARAMETERS;
     }
@@ -18674,14 +19987,25 @@ static enum ggml_opt_result linesearch_backtracking(
         {
             ggml_opt_set_params(np, ps, x);
 
-            ggml_graph_reset  (gf);
-            ggml_set_f32      (f->grad, 1.0f);
+            *fx = 0;
+            memset(g, 0, sizeof(float)*nx);
+            for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+                if (callback) {
+                    // LBFG-S does not support learning rate -> ignore learning schedule
+                    float sched = 0;
+                    callback(callback_data, accum_step, &sched, cancel);
+                    if (*cancel) {
+                        return GGML_OPT_CANCEL;
+                    }
+                }
+                // ggml_graph_reset  (gf);
+                ggml_set_f32      (f->grad, 1.0f);
+                ggml_graph_compute(gb, cplan);
+                ggml_opt_acc_grad(np, ps, g, accum_norm);
+                *fx += ggml_get_f32_1d(f, 0);
+            }
+            *fx *= accum_norm;
 
-            ggml_graph_compute_with_ctx(ctx, gb, params->n_threads);
-
-            ggml_opt_get_grad(np, ps, g);
-
-            *fx = ggml_get_f32_1d(f, 0);
         }
 
         ++count;
@@ -18711,7 +20035,6 @@ static enum ggml_opt_result linesearch_backtracking(
                     // strong Wolfe condition (GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE)
                     return count;
                 }
-                return count;
             }
         }
 
@@ -18728,7 +20051,7 @@ static enum ggml_opt_result linesearch_backtracking(
         (*step) *= width;
     }
 
-    return GGML_LINESEARCH_FAIL;
+    GGML_UNREACHABLE();
 }
 
 static enum ggml_opt_result ggml_opt_lbfgs(
@@ -18737,7 +20060,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         struct ggml_opt_params params,
         struct ggml_tensor * f,
         struct ggml_cgraph * gf,
-        struct ggml_cgraph * gb) {
+        struct ggml_cgraph * gb,
+        ggml_opt_callback callback,
+        void * callback_data) {
     if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE ||
         params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
         if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) {
@@ -18769,6 +20094,10 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         opt->iter = iter;
     }
 
+    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
+    cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
+
     float * x  = opt->lbfgs.x->data;  // current parameters
     float * xp = opt->lbfgs.xp->data; // previous parameters
     float * g  = opt->lbfgs.g->data;  // current gradient
@@ -18777,6 +20106,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
     float * pf = params.past > 0 ? opt->lbfgs.pf->data : NULL; // past function values
 
+    const int n_accum = MAX(1, params.n_gradient_accumulation);
+    const float accum_norm = 1.0f / (float) n_accum;
+
     float fx    = 0.0f; // cost function value
     float xnorm = 0.0f; // ||x||
     float gnorm = 0.0f; // ||g||
@@ -18790,18 +20122,33 @@ static enum ggml_opt_result ggml_opt_lbfgs(
     float * lm_s     = opt->lbfgs.lms->data;
     float * lm_y     = opt->lbfgs.lmy->data;
 
+    bool cancel = false;
+
     // evaluate the function value and its gradient
     {
         ggml_opt_set_params(np, ps, x);
 
-        ggml_graph_reset  (gf);
-        ggml_set_f32      (f->grad, 1.0f);
+        fx = 0;
+        memset(g, 0, sizeof(float)*nx);
+        for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+            if (callback) {
+                // LBFG-S does not support learning rate -> ignore learning schedule
+                float sched = 0;
+                callback(callback_data, accum_step, &sched, &cancel);
+                if (cancel) {
+                    return GGML_OPT_CANCEL;
+                }
+            }
+            // ggml_graph_reset  (gf);
+            ggml_set_f32      (f->grad, 1.0f);
+            ggml_graph_compute(gb, &cplan);
+            ggml_opt_acc_grad(np, ps, g, accum_norm);
+            fx += ggml_get_f32_1d(f, 0);
+        }
+        fx *= accum_norm;
 
-        ggml_graph_compute_with_ctx(ctx, gb, params.n_threads);
-
-        ggml_opt_get_grad(np, ps, g);
-
-        fx = ggml_get_f32_1d(f, 0);
+        opt->loss_before = fx;
+        opt->loss_after  = fx;
     }
 
     // search direction = -gradient
@@ -18856,7 +20203,10 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         ggml_vec_cpy_f32(nx, xp, x);
         ggml_vec_cpy_f32(nx, gp, g);
 
-        ls = linesearch_backtracking(ctx, &params, nx, x, &fx, g, d, step, xp, f, gf, gb, np, ps);
+        ls = linesearch_backtracking(&params, nx, x, &fx, g, d, step, xp, f, gb, &cplan, np, ps, &cancel, callback, callback_data);
+        if (cancel) {
+            return GGML_OPT_CANCEL;
+        }
 
         if (ls < 0) {
             // linesearch failed - go back to the previous point and return
@@ -18866,6 +20216,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
             return ls;
         }
 
+        opt->loss_after = fx;
+
         ggml_vec_norm_f32(nx, &xnorm, x);
         ggml_vec_norm_f32(nx, &gnorm, g);
 
@@ -18923,7 +20275,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         //     ys = y^t \cdot s    -> 1 / \rho.
         //     yy = y^t \cdot y.
         //
-        ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0] *nx]);
+        ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0]*nx]);
         ggml_vec_dot_f32(nx, &yy, &lm_y[end[0]*nx], &lm_y[end[0]*nx]);
 
         lm_ys[end[0]] = ys;
@@ -18963,7 +20315,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         step[0] = 1.0;
     }
 
-    return GGML_OPT_DID_NOT_CONVERGE;
+    GGML_UNREACHABLE();
 }
 
 struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
@@ -18983,16 +20335,20 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                     .print_forward_graph  = true,
                     .print_backward_graph = true,
 
+                    .n_gradient_accumulation = 1,
+
                     .adam = {
                         .n_iter = 10000,
                         .sched  = 1.000f,
-                        .decay  = 0.001f,
+                        .decay  = 0.0f,
+                        .decay_min_ndim = 2,
                         .alpha  = 0.001f,
                         .beta1  = 0.9f,
                         .beta2  = 0.999f,
                         .eps    = 1e-8f,
                         .eps_f  = 1e-5f,
                         .eps_g  = 1e-3f,
+                        .gclip  = 0.0f,
                     },
                 };
             } break;
@@ -19009,6 +20365,8 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                     .print_forward_graph  = true,
                     .print_backward_graph = true,
 
+                    .n_gradient_accumulation = 1,
+
                     .lbfgs = {
                         .m              = 6,
                         .n_iter         = 100,
@@ -19039,44 +20397,53 @@ GGML_API void ggml_opt_init(
     opt->iter = 0;
     opt->nx = nx;
     opt->just_initialized = true;
+    if (opt->ctx == NULL) {
+        struct ggml_init_params ctx_opt_params;
+        if (opt->params.type == GGML_OPT_ADAM) {
+            ctx_opt_params.mem_size = GGML_MEM_ALIGN*3 + ggml_tensor_overhead()*3 + ggml_type_size(GGML_TYPE_F32)*nx*3;
+            if (opt->params.past > 0) {
+                ctx_opt_params.mem_size += GGML_MEM_ALIGN + ggml_tensor_overhead() + ggml_type_size(GGML_TYPE_F32)*opt->params.past;
+            }
+        } else if (opt->params.type == GGML_OPT_LBFGS) {
+            ctx_opt_params.mem_size = GGML_MEM_ALIGN*9 + ggml_tensor_overhead()*9 + ggml_type_size(GGML_TYPE_F32)*(nx*5 + opt->params.lbfgs.m*2 + nx*opt->params.lbfgs.m*2);
+            if (opt->params.past > 0) {
+                ctx_opt_params.mem_size += GGML_MEM_ALIGN + ggml_tensor_overhead() + ggml_type_size(GGML_TYPE_F32)*opt->params.past;
+            }
+        }
+        ctx_opt_params.mem_buffer = NULL;
+        ctx_opt_params.no_alloc   = false;
+
+        opt->ctx = ggml_init(ctx_opt_params);
+    }
     switch (opt->params.type) {
         case GGML_OPT_ADAM:
             {
-                opt->adam.x  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.g1 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.g2 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.m  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.v  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.mh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->adam.vh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
+                opt->adam.g  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->adam.m  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->adam.v  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->adam.pf = params.past > 0
-                    ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.past)
+                    ? ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, params.past)
                     : NULL;
-                ggml_set_zero(opt->adam.x);
-                ggml_set_zero(opt->adam.g1);
-                ggml_set_zero(opt->adam.g2);
                 ggml_set_zero(opt->adam.m);
                 ggml_set_zero(opt->adam.v);
-                ggml_set_zero(opt->adam.mh);
-                ggml_set_zero(opt->adam.vh);
                 if (opt->adam.pf) {
                     ggml_set_zero(opt->adam.pf);
                 }
             } break;
         case GGML_OPT_LBFGS:
             {
-                opt->lbfgs.x  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->lbfgs.xp = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->lbfgs.g  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->lbfgs.gp = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
-                opt->lbfgs.d  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx);
+                opt->lbfgs.x  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->lbfgs.xp = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->lbfgs.g  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->lbfgs.gp = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
+                opt->lbfgs.d  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->lbfgs.pf = params.past > 0
-                    ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.past)
+                    ? ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, params.past)
                     : NULL;
-                opt->lbfgs.lmal = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.lbfgs.m);
-                opt->lbfgs.lmys = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.lbfgs.m);
-                opt->lbfgs.lms  = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, params.lbfgs.m);
-                opt->lbfgs.lmy  = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, params.lbfgs.m);
+                opt->lbfgs.lmal = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, params.lbfgs.m);
+                opt->lbfgs.lmys = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, params.lbfgs.m);
+                opt->lbfgs.lms  = ggml_new_tensor_2d(opt->ctx, GGML_TYPE_F32, nx, params.lbfgs.m);
+                opt->lbfgs.lmy  = ggml_new_tensor_2d(opt->ctx, GGML_TYPE_F32, nx, params.lbfgs.m);
                 ggml_set_zero(opt->lbfgs.x);
                 ggml_set_zero(opt->lbfgs.xp);
                 ggml_set_zero(opt->lbfgs.g);
@@ -19142,7 +20509,7 @@ enum ggml_opt_result ggml_opt_resume(
     *gf = ggml_build_forward (f);
     *gb = ggml_build_backward(ctx, gf, true);
 
-    return ggml_opt_resume_g(ctx, opt, f, gf, gb);
+    return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL);
 }
 
 enum ggml_opt_result ggml_opt_resume_g(
@@ -19150,7 +20517,9 @@ enum ggml_opt_result ggml_opt_resume_g(
         struct ggml_opt_context * opt,
         struct ggml_tensor * f,
         struct ggml_cgraph * gf,
-        struct ggml_cgraph * gb) {
+        struct ggml_cgraph * gb,
+        ggml_opt_callback callback,
+        void * callback_data) {
 
     // build forward + backward compute graphs
     enum ggml_opt_result result = GGML_OPT_OK;
@@ -19158,11 +20527,11 @@ enum ggml_opt_result ggml_opt_resume_g(
     switch (opt->params.type) {
         case GGML_OPT_ADAM:
             {
-                result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb);
+                result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
             } break;
         case GGML_OPT_LBFGS:
             {
-                result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb);
+                result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
             } break;
     }
 
@@ -19394,7 +20763,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
 ////////////////////////////////////////////////////////////////////////////////
 
 struct gguf_str {
-    uint32_t n;
+    uint64_t n;  // GGUFv2
     char * data;
 };
 
@@ -19408,9 +20777,12 @@ static const size_t GGUF_TYPE_SIZE[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_FLOAT32] = sizeof(float),
     [GGUF_TYPE_BOOL]    = sizeof(bool),
     [GGUF_TYPE_STRING]  = sizeof(struct gguf_str),
+    [GGUF_TYPE_UINT64]  = sizeof(uint64_t),
+    [GGUF_TYPE_INT64]   = sizeof(int64_t),
+    [GGUF_TYPE_FLOAT64] = sizeof(double),
     [GGUF_TYPE_ARRAY]   = 0, // undefined
 };
-static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10");
+static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");
 
 static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_UINT8]   = "u8",
@@ -19423,8 +20795,11 @@ static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = {
     [GGUF_TYPE_BOOL]    = "bool",
     [GGUF_TYPE_STRING]  = "str",
     [GGUF_TYPE_ARRAY]   = "arr",
+    [GGUF_TYPE_UINT64]  = "u64",
+    [GGUF_TYPE_INT64]   = "i64",
+    [GGUF_TYPE_FLOAT64] = "f64",
 };
-static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10");
+static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13");
 
 union gguf_value {
     uint8_t  uint8;
@@ -19434,6 +20809,9 @@ union gguf_value {
     uint32_t uint32;
     int32_t  int32;
     float    float32;
+    uint64_t uint64;
+    int64_t  int64;
+    double   float64;
     bool     bool_;
 
     struct gguf_str str;
@@ -19441,7 +20819,7 @@ union gguf_value {
     struct {
         enum gguf_type type;
 
-        uint32_t n;
+        uint64_t n;  // GGUFv2
         void * data;
     } arr;
 };
@@ -19449,8 +20827,6 @@ union gguf_value {
 struct gguf_kv {
     struct gguf_str key;
 
-    uint32_t n_bytes; // TODO: is this actually needed?
-
     enum  gguf_type  type;
     union gguf_value value;
 };
@@ -19458,15 +20834,15 @@ struct gguf_kv {
 struct gguf_header {
     uint32_t magic;
     uint32_t version;
-    uint32_t n_tensors;
-    uint32_t n_kv;
+    uint64_t n_tensors; // GGUFv2
+    uint64_t n_kv;      // GGUFv2
 };
 
 struct gguf_tensor_info {
     struct gguf_str name;
 
     uint32_t n_dims;
-    uint32_t ne[GGML_MAX_DIMS];
+    uint64_t ne[GGML_MAX_DIMS];
 
     enum ggml_type type;
 
@@ -19497,19 +20873,32 @@ static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset)
     return n == size;
 }
 
-static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
+// NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset) {
     p->n    = 0;
     p->data = NULL;
 
     bool ok = true;
 
-    // TODO: how to avoid mallocs for strings?
     ok = ok && gguf_fread_el(file, &p->n,    sizeof(p->n), offset); p->data = calloc(p->n + 1, 1);
     ok = ok && gguf_fread_el(file,  p->data, p->n,         offset);
 
     return ok;
 }
 
+static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) {
+    p->n    = 0;
+    p->data = NULL;
+
+    bool ok = true;
+
+    uint32_t n = 0;
+    ok = ok && gguf_fread_el(file, &n,       sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n;
+    ok = ok && gguf_fread_el(file,  p->data, p->n,      offset);
+
+    return ok;
+}
+
 struct gguf_context * gguf_init_empty(void) {
     struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
 
@@ -19565,8 +20954,21 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         ctx->data  = NULL;
 
         ok = ok && gguf_fread_el(file, &ctx->header.version,   sizeof(ctx->header.version),   &offset);
-        ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset);
-        ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);
+
+        if (ctx->header.version == 1) {
+            // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+            uint32_t n_tensors = 0;
+            uint32_t n_kv      = 0;
+
+            ok = ok && gguf_fread_el(file, &n_tensors, sizeof(n_tensors), &offset);
+            ok = ok && gguf_fread_el(file, &n_kv,      sizeof(n_kv),      &offset);
+
+            ctx->header.n_tensors = n_tensors;
+            ctx->header.n_kv      = n_kv;
+        } else {
+            ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset);
+            ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);
+        }
 
         if (!ok) {
             fprintf(stderr, "%s: failed to read header\n", __func__);
@@ -19576,18 +20978,23 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         }
     }
 
+    // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+    bool (* gguf_fread_str)(FILE *, struct gguf_str *, size_t *) = gguf_fread_str_cur;
+    if (ctx->header.version == 1) {
+        gguf_fread_str = gguf_fread_str_v1;
+    }
+
     // read the kv pairs
     {
-        ctx->kv = GGML_ALIGNED_MALLOC(ctx->header.n_kv * sizeof(struct gguf_kv));
+        ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv));
 
         for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
             struct gguf_kv * kv = &ctx->kv[i];
 
             //fprintf(stderr, "%s: reading kv %d\n", __func__, i);
 
-            ok = ok && gguf_fread_str(file, &kv->key,                          &offset);
-          //ok = ok && gguf_fread_el (file, &kv->n_bytes, sizeof(kv->n_bytes), &offset);
-            ok = ok && gguf_fread_el (file, &kv->type,    sizeof(kv->type),    &offset);
+            ok = ok && gguf_fread_str(file, &kv->key,                    &offset);
+            ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset);
 
             //fprintf(stderr, "%s: reading kv with key %s\n", __func__, kv->key.data);
 
@@ -19599,12 +21006,23 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 case GGUF_TYPE_UINT32:  ok = ok && gguf_fread_el (file, &kv->value.uint32,  sizeof(kv->value.uint32),  &offset); break;
                 case GGUF_TYPE_INT32:   ok = ok && gguf_fread_el (file, &kv->value.int32,   sizeof(kv->value.int32),   &offset); break;
                 case GGUF_TYPE_FLOAT32: ok = ok && gguf_fread_el (file, &kv->value.float32, sizeof(kv->value.float32), &offset); break;
+                case GGUF_TYPE_UINT64:  ok = ok && gguf_fread_el (file, &kv->value.uint64,  sizeof(kv->value.uint64),  &offset); break;
+                case GGUF_TYPE_INT64:   ok = ok && gguf_fread_el (file, &kv->value.int64,   sizeof(kv->value.int64),   &offset); break;
+                case GGUF_TYPE_FLOAT64: ok = ok && gguf_fread_el (file, &kv->value.float64, sizeof(kv->value.float64), &offset); break;
                 case GGUF_TYPE_BOOL:    ok = ok && gguf_fread_el (file, &kv->value.bool_,   sizeof(kv->value.bool_),   &offset); break;
                 case GGUF_TYPE_STRING:  ok = ok && gguf_fread_str(file, &kv->value.str,                                &offset); break;
                 case GGUF_TYPE_ARRAY:
                     {
                         ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset);
-                        ok = ok && gguf_fread_el(file, &kv->value.arr.n,    sizeof(kv->value.arr.n),    &offset);
+
+                        if (ctx->header.version == 1) {
+                            // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+                            uint32_t n = 0;
+                            ok = ok && gguf_fread_el(file, &n, sizeof(n), &offset);
+                            kv->value.arr.n = n;
+                        } else {
+                            ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset);
+                        }
 
                         switch (kv->value.arr.type) {
                             case GGUF_TYPE_UINT8:
@@ -19614,6 +21032,9 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                             case GGUF_TYPE_UINT32:
                             case GGUF_TYPE_INT32:
                             case GGUF_TYPE_FLOAT32:
+                            case GGUF_TYPE_UINT64:
+                            case GGUF_TYPE_INT64:
+                            case GGUF_TYPE_FLOAT64:
                             case GGUF_TYPE_BOOL:
                                 {
                                     kv->value.arr.data = malloc(kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
@@ -19628,10 +21049,10 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                                 } break;
                             case GGUF_TYPE_ARRAY:
                             case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type"); break;
-                        };
+                        }
                     } break;
                 case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type");
-            };
+            }
 
             if (!ok) {
                 break;
@@ -19648,7 +21069,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     // read the tensor infos
     {
-        ctx->infos = GGML_ALIGNED_MALLOC(ctx->header.n_tensors * sizeof(struct gguf_tensor_info));
+        ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info));
 
         for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
@@ -19660,7 +21081,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
             ok = ok && gguf_fread_str(file, &info->name,                          &offset);
             ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims),  &offset);
             for (uint32_t j = 0; j < info->n_dims; ++j) {
-                ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
+                if (ctx->header.version == 1) {
+                    // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023
+                    uint32_t t = 0;
+                    ok = ok && gguf_fread_el(file, &t, sizeof(t), &offset);
+                    info->ne[j] = t;
+                } else {
+                    ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
+                }
             }
             ok = ok && gguf_fread_el (file, &info->type,   sizeof(info->type),    &offset);
             ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset),  &offset);
@@ -19744,7 +21172,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
         struct ggml_tensor * data = NULL;
 
-        if (params.no_alloc == false) {
+        if (!params.no_alloc) {
             data = ggml_new_tensor_1d(ctx_data, GGML_TYPE_I8, ctx->size);
 
             ok = ok && data != NULL;
@@ -19785,7 +21213,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
             }
 
             // point the data member to the appropriate location in the binary blob using the tensor infos
-            if (params.no_alloc == false) {
+            if (!params.no_alloc) {
               //cur->data = (char *) data->data + ctx->infos[i].offset - ctx->offset; // offset from start of file
                 cur->data = (char *) data->data + ctx->infos[i].offset;               // offset from data
             }
@@ -19842,7 +21270,7 @@ void gguf_free(struct gguf_context * ctx) {
             }
         }
 
-        GGML_ALIGNED_FREE(ctx->kv);
+        free(ctx->kv);
     }
 
     if (ctx->infos) {
@@ -19854,7 +21282,7 @@ void gguf_free(struct gguf_context * ctx) {
             }
         }
 
-        GGML_ALIGNED_FREE(ctx->infos);
+        free(ctx->infos);
     }
 
     GGML_ALIGNED_FREE(ctx);
@@ -19864,27 +21292,27 @@ const char * gguf_type_name(enum gguf_type type) {
     return GGUF_TYPE_NAME[type];
 }
 
-int gguf_get_version(struct gguf_context * ctx) {
+int gguf_get_version(const struct gguf_context * ctx) {
     return ctx->header.version;
 }
 
-size_t gguf_get_alignment(struct gguf_context * ctx) {
+size_t gguf_get_alignment(const struct gguf_context * ctx) {
     return ctx->alignment;
 }
 
-size_t gguf_get_data_offset(struct gguf_context * ctx) {
+size_t gguf_get_data_offset(const struct gguf_context * ctx) {
     return ctx->offset;
 }
 
-void * gguf_get_data(struct gguf_context * ctx) {
+void * gguf_get_data(const struct gguf_context * ctx) {
     return ctx->data;
 }
 
-int gguf_get_n_kv(struct gguf_context * ctx) {
+int gguf_get_n_kv(const struct gguf_context * ctx) {
     return ctx->header.n_kv;
 }
 
-int gguf_find_key(struct gguf_context * ctx, const char * key) {
+int gguf_find_key(const struct gguf_context * ctx, const char * key) {
     // return -1 if key not found
     int keyfound = -1;
 
@@ -19900,73 +21328,101 @@ int gguf_find_key(struct gguf_context * ctx, const char * key) {
     return keyfound;
 }
 
-const char * gguf_get_key(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].key.data;
+const char * gguf_get_key(const struct gguf_context * ctx, int key_id) {
+    return ctx->kv[key_id].key.data;
 }
 
-enum gguf_type gguf_get_kv_type(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].type;
+enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int key_id) {
+    return ctx->kv[key_id].type;
 }
 
-enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.arr.type;
+enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
+    return ctx->kv[key_id].value.arr.type;
 }
 
-const void * gguf_get_arr_data(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.arr.data;
+const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
+    return ctx->kv[key_id].value.arr.data;
 }
 
-const char * gguf_get_arr_str(struct gguf_context * ctx, int key_id, int i) {
+const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
     struct gguf_kv * kv = &ctx->kv[key_id];
     struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i];
     return str->data;
 }
 
-int gguf_get_arr_n(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.arr.n;
+int gguf_get_arr_n(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
+    return ctx->kv[key_id].value.arr.n;
 }
 
-uint8_t gguf_get_val_u8(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.uint8;
+uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT8);
+    return ctx->kv[key_id].value.uint8;
 }
 
-int8_t gguf_get_val_i8(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.int8;
+int8_t gguf_get_val_i8(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT8);
+    return ctx->kv[key_id].value.int8;
 }
 
-uint16_t gguf_get_val_u16(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.uint16;
+uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT16);
+    return ctx->kv[key_id].value.uint16;
 }
 
-int16_t gguf_get_val_i16(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.int16;
+int16_t gguf_get_val_i16(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT16);
+    return ctx->kv[key_id].value.int16;
 }
 
-uint32_t gguf_get_val_u32(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.uint32;
+uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT32);
+    return ctx->kv[key_id].value.uint32;
 }
 
-int32_t gguf_get_val_i32(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.int32;
+int32_t gguf_get_val_i32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT32);
+    return ctx->kv[key_id].value.int32;
 }
 
-float gguf_get_val_f32(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.float32;
+float gguf_get_val_f32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT32);
+    return ctx->kv[key_id].value.float32;
 }
 
-bool gguf_get_val_bool(struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.bool_;
+uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT64);
+    return ctx->kv[key_id].value.uint64;
 }
 
-const char * gguf_get_val_str (struct gguf_context * ctx, int i) {
-    return ctx->kv[i].value.str.data;
+int64_t gguf_get_val_i64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT64);
+    return ctx->kv[key_id].value.int64;
 }
 
-int gguf_get_n_tensors(struct gguf_context * ctx) {
+double gguf_get_val_f64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT64);
+    return ctx->kv[key_id].value.float64;
+}
+
+bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_BOOL);
+    return ctx->kv[key_id].value.bool_;
+}
+
+const char * gguf_get_val_str(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_STRING);
+    return ctx->kv[key_id].value.str.data;
+}
+
+int gguf_get_n_tensors(const struct gguf_context * ctx) {
     return ctx->header.n_tensors;
 }
 
-int gguf_find_tensor(struct gguf_context * ctx, const char * name) {
+int gguf_find_tensor(const struct gguf_context * ctx, const char * name) {
     // return -1 if tensor not found
     int tensorfound = -1;
 
@@ -19982,11 +21438,11 @@ int gguf_find_tensor(struct gguf_context * ctx, const char * name) {
     return tensorfound;
 }
 
-size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i) {
+size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].offset;
 }
 
-char * gguf_get_tensor_name(struct gguf_context * ctx, int i) {
+char * gguf_get_tensor_name(const struct gguf_context * ctx, int i) {
     return ctx->infos[i].name.data;
 }
 
@@ -20000,7 +21456,7 @@ static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     const int n_kv = gguf_get_n_kv(ctx);
 
     ctx->kv = realloc(ctx->kv, (n_kv + 1) * sizeof(struct gguf_kv));
-    ctx->kv[n_kv].key.n    = strlen(key) + 1;
+    ctx->kv[n_kv].key.n    = strlen(key);
     ctx->kv[n_kv].key.data = strdup(key);
     ctx->header.n_kv++;
 
@@ -20056,6 +21512,27 @@ void gguf_set_val_f32(struct gguf_context * ctx, const char * key, float val) {
     ctx->kv[idx].value.float32 = val;
 }
 
+void gguf_set_val_u64(struct gguf_context * ctx, const char * key, uint64_t val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type         = GGUF_TYPE_UINT64;
+    ctx->kv[idx].value.uint64 = val;
+}
+
+void gguf_set_val_i64(struct gguf_context * ctx, const char * key, int64_t val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type        = GGUF_TYPE_INT64;
+    ctx->kv[idx].value.int64 = val;
+}
+
+void gguf_set_val_f64(struct gguf_context * ctx, const char * key, double val) {
+    const int idx = gguf_get_or_add_key(ctx, key);
+
+    ctx->kv[idx].type          = GGUF_TYPE_FLOAT64;
+    ctx->kv[idx].value.float64 = val;
+}
+
 void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val) {
     const int idx = gguf_get_or_add_key(ctx, key);
 
@@ -20067,7 +21544,7 @@ void gguf_set_val_str(struct gguf_context * ctx, const char * key, const char *
     const int idx = gguf_get_or_add_key(ctx, key);
 
     ctx->kv[idx].type           = GGUF_TYPE_STRING;
-    ctx->kv[idx].value.str.n    = strlen(val) + 1;
+    ctx->kv[idx].value.str.n    = strlen(val);
     ctx->kv[idx].value.str.data = strdup(val);
 }
 
@@ -20090,7 +21567,7 @@ void gguf_set_arr_str(struct gguf_context * ctx, const char * key, const char **
     ctx->kv[idx].value.arr.data = malloc(n*sizeof(struct gguf_str));
     for (int i = 0; i < n; i++) {
         struct gguf_str * str = &((struct gguf_str *)ctx->kv[idx].value.arr.data)[i];
-        str->n    = strlen(data[i]) + 1;
+        str->n    = strlen(data[i]);
         str->data = strdup(data[i]);
     }
 }
@@ -20106,6 +21583,9 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
             case GGUF_TYPE_UINT32:  gguf_set_val_u32 (ctx, src->kv[i].key.data, src->kv[i].value.uint32);   break;
             case GGUF_TYPE_INT32:   gguf_set_val_i32 (ctx, src->kv[i].key.data, src->kv[i].value.int32);    break;
             case GGUF_TYPE_FLOAT32: gguf_set_val_f32 (ctx, src->kv[i].key.data, src->kv[i].value.float32);  break;
+            case GGUF_TYPE_UINT64:  gguf_set_val_u64 (ctx, src->kv[i].key.data, src->kv[i].value.uint64);   break;
+            case GGUF_TYPE_INT64:   gguf_set_val_i64 (ctx, src->kv[i].key.data, src->kv[i].value.int64);    break;
+            case GGUF_TYPE_FLOAT64: gguf_set_val_f64 (ctx, src->kv[i].key.data, src->kv[i].value.float64);  break;
             case GGUF_TYPE_BOOL:    gguf_set_val_bool(ctx, src->kv[i].key.data, src->kv[i].value.bool_);    break;
             case GGUF_TYPE_STRING:  gguf_set_val_str (ctx, src->kv[i].key.data, src->kv[i].value.str.data); break;
             case GGUF_TYPE_ARRAY:
@@ -20134,7 +21614,7 @@ void gguf_add_tensor(
     const int idx = ctx->header.n_tensors;
     ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info));
 
-    ctx->infos[idx].name.n    = strlen(tensor->name) + 1;
+    ctx->infos[idx].name.n    = strlen(tensor->name);
     ctx->infos[idx].name.data = strdup(tensor->name);
 
     for (int i = 0; i < GGML_MAX_DIMS; ++i) {
@@ -20245,7 +21725,7 @@ static void gguf_bwrite_el(struct gguf_buf * buf, const void * val, size_t el_si
     buf->offset += el_size;
 }
 
-static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta) {
+static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta) {
     // write header
     gguf_bwrite_el(buf, &ctx->header.magic,     sizeof(ctx->header.magic));
     gguf_bwrite_el(buf, &ctx->header.version,   sizeof(ctx->header.version));
@@ -20267,6 +21747,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
             case GGUF_TYPE_UINT32:  gguf_bwrite_el (buf, &kv->value.uint32,  sizeof(kv->value.uint32) ); break;
             case GGUF_TYPE_INT32:   gguf_bwrite_el (buf, &kv->value.int32,   sizeof(kv->value.int32)  ); break;
             case GGUF_TYPE_FLOAT32: gguf_bwrite_el (buf, &kv->value.float32, sizeof(kv->value.float32)); break;
+            case GGUF_TYPE_UINT64:  gguf_bwrite_el (buf, &kv->value.uint64,  sizeof(kv->value.uint64) ); break;
+            case GGUF_TYPE_INT64:   gguf_bwrite_el (buf, &kv->value.int64,   sizeof(kv->value.int64)  ); break;
+            case GGUF_TYPE_FLOAT64: gguf_bwrite_el (buf, &kv->value.float64, sizeof(kv->value.float64)); break;
             case GGUF_TYPE_BOOL:    gguf_bwrite_el (buf, &kv->value.bool_,   sizeof(kv->value.bool_)  ); break;
             case GGUF_TYPE_STRING:  gguf_bwrite_str(buf, &kv->value.str                               ); break;
             case GGUF_TYPE_ARRAY:
@@ -20282,6 +21765,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
                         case GGUF_TYPE_UINT32:
                         case GGUF_TYPE_INT32:
                         case GGUF_TYPE_FLOAT32:
+                        case GGUF_TYPE_UINT64:
+                        case GGUF_TYPE_INT64:
+                        case GGUF_TYPE_FLOAT64:
                         case GGUF_TYPE_BOOL:
                             {
                                 gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
@@ -20294,10 +21780,10 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
                             } break;
                         case GGUF_TYPE_ARRAY:
                         case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type"); break;
-                    };
+                    }
                 } break;
             case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type");
-        };
+        }
     }
 
     // write tensor infos
@@ -20354,7 +21840,7 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf,
     }
 }
 
-void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta) {
+void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
     FILE * file = fopen(fname, "wb");
     if (!file) {
         GGML_ASSERT(false && "failed to open file for writing");
@@ -20371,7 +21857,7 @@ void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only
     fclose(file);
 }
 
-size_t gguf_get_meta_size(struct gguf_context * ctx) {
+size_t gguf_get_meta_size(const struct gguf_context * ctx) {
     // no allocs - only compute size
     struct gguf_buf buf = gguf_buf_init(0);
 
@@ -20380,7 +21866,7 @@ size_t gguf_get_meta_size(struct gguf_context * ctx) {
     return buf.offset;
 }
 
-void gguf_get_meta_data(struct gguf_context * ctx, void * data) {
+void gguf_get_meta_data(const struct gguf_context * ctx, void * data) {
     struct gguf_buf buf = gguf_buf_init(16*1024);
 
     gguf_write_to_buf(ctx, &buf, true);
@@ -20456,6 +21942,14 @@ int ggml_cpu_has_arm_fma(void) {
 #endif
 }
 
+int ggml_cpu_has_metal(void) {
+#if defined(GGML_USE_METAL)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_f16c(void) {
 #if defined(__F16C__)
     return 1;
@@ -20516,6 +22010,14 @@ int ggml_cpu_has_sse3(void) {
 #endif
 }
 
+int ggml_cpu_has_ssse3(void) {
+#if defined(__SSSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_vsx(void) {
 #if defined(__POWER9_VECTOR__)
     return 1;
diff --git a/ggml.h b/ggml.h
index 3c48fd27f..a9d4e33d9 100644
--- a/ggml.h
+++ b/ggml.h
@@ -130,13 +130,16 @@
 // The data of the tensor is accessed via the "data" pointer. For example:
 //
 //   {
-//       struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3);
+//       const int nx = 2;
+//       const int ny = 3;
 //
-//       // a[2, 1] = 1.0f;
-//       *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f;
+//       struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, ny);
 //
-//       // a[0, 2] = 2.0f;
-//       *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f;
+//       for (int y = 0; y < ny; y++) {
+//           for (int x = 0; x < nx; x++) {
+//               *(float *) ((char *) a->data + y*a->nb[1] + x*a->nb[0]) = x + y;
+//           }
+//       }
 //
 //       ...
 //   }
@@ -192,6 +195,14 @@
 #    define GGML_DEPRECATED(func, hint) func
 #endif
 
+#ifndef __GNUC__
+#    define GGML_ATTRIBUTE_FORMAT(...)
+#elif defined(__MINGW32__)
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+
 #include <stdint.h>
 #include <stddef.h>
 #include <stdbool.h>
@@ -203,20 +214,25 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define GGML_MAX_DIMS          4
-#define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        256
+#define GGML_MAX_NODES         16384
+#define GGML_MAX_PARAMS        1024
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          64
 #define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
+#if UINTPTR_MAX == 0xFFFFFFFF
+    #define GGML_MEM_ALIGN 4
+#else
+    #define GGML_MEM_ALIGN 16
+#endif
 
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
 #define GGUF_MAGIC   0x46554747 // "GGUF"
-#define GGUF_VERSION 1
+#define GGUF_VERSION 2
 
 #define GGUF_DEFAULT_ALIGNMENT 32
 
@@ -232,6 +248,14 @@
         } \
     } while (0)
 
+#ifndef NDEBUG
+#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
+#elif defined(__GNUC__)
+#define GGML_UNREACHABLE() __builtin_unreachable()
+#else
+#define GGML_UNREACHABLE() ((void) 0)
+#endif
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -377,10 +401,14 @@ extern "C" {
         GGML_OP_CLAMP,
         GGML_OP_CONV_1D,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
 
+        GGML_OP_CONV_1D_STAGE_0,  // internal
+        GGML_OP_CONV_1D_STAGE_1,  // internal
+
         GGML_OP_UPSCALE, // nearest interpolate
 
         GGML_OP_FLASH_ATTN,
@@ -429,6 +457,12 @@ extern "C" {
         GGML_OBJECT_WORK_BUFFER
     };
 
+    enum ggml_log_level {
+        GGML_LOG_LEVEL_ERROR = 2,
+        GGML_LOG_LEVEL_WARN = 3,
+        GGML_LOG_LEVEL_INFO = 4
+    };
+
     // ggml object
     struct ggml_object {
         size_t offs;
@@ -451,8 +485,8 @@ extern "C" {
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
         size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
-                                   // nb[0] = sizeof(type)
-                                   // nb[1] = nb[0]   * ne[0] + padding
+                                   // nb[0] = ggml_type_size(type)
+                                   // nb[1] = nb[0]   * (ne[0] / ggml_blck_size(type)) + padding
                                    // nb[i] = nb[i-1] * ne[i-1]
 
         // compute data
@@ -471,6 +505,9 @@ extern "C" {
         int64_t perf_cycles;
         int64_t perf_time_us;
 
+        struct ggml_tensor * view_src;
+        size_t               view_offs;
+
         void * data;
 
         char name[GGML_MAX_NAME];
@@ -501,7 +538,15 @@ extern "C" {
     // next prime after GGML_MAX_NODES
     // #define GGML_GRAPH_HASHTABLE_SIZE 4099
     // next prime after GGML_MAX_NODES * 2 (nodes + leafs)
-    #define GGML_GRAPH_HASHTABLE_SIZE 8273
+    // #define GGML_GRAPH_HASHTABLE_SIZE 8273
+    // #define GGML_GRAPH_HASHTABLE_SIZE 16411
+    #define GGML_GRAPH_HASHTABLE_SIZE 32771
+
+    enum ggml_cgraph_eval_order {
+        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+        GGML_CGRAPH_EVAL_ORDER_COUNT
+    };
 
     // computation graph
     struct ggml_cgraph {
@@ -514,6 +559,8 @@ extern "C" {
 
         void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE];
 
+        enum ggml_cgraph_eval_order order;
+
         // performance
         int     perf_runs;
         int64_t perf_cycles;
@@ -653,7 +700,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
-    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
+    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
@@ -661,12 +708,21 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
     GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
 
+    // Converts a flat index into coordinates
+    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+
     GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
     GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
 
+    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
     GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
     GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
 
+    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
@@ -674,6 +730,7 @@ extern "C" {
 
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
+    GGML_ATTRIBUTE_FORMAT(2, 3)
     GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
     //
@@ -699,6 +756,12 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add_cast(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            enum   ggml_type      type);
+
     GGML_API struct ggml_tensor * ggml_add1(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -808,6 +871,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // sums repetitions in a into shape of b
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -909,14 +973,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // normalize along rows
-    // TODO: eps is hardcoded to 1e-5 for now
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_norm_inplace(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_rms_norm(
             struct ggml_context * ctx,
@@ -943,11 +1008,11 @@ extern "C" {
 
     // a - x
     // b - dy
-    // TODO: update with configurable eps
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            float                 eps);
 
     // A: n columns, m rows
     // B: n columns, p rows  (i.e. we transpose it internally)
@@ -1028,7 +1093,6 @@ extern "C" {
             size_t                nb1,
             size_t                offset);
 
-
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
             struct ggml_context * ctx,
@@ -1051,6 +1115,33 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // make contiguous, with new shape
+    GGML_API struct ggml_tensor * ggml_cont_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
+    GGML_API struct ggml_tensor * ggml_cont_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1);
+
+    GGML_API struct ggml_tensor * ggml_cont_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2);
+
+    GGML_API struct ggml_tensor * ggml_cont_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
     // return view(a), b specifies the new shape
     // TODO: when we start computing gradient, make a copy instead of view
     GGML_API struct ggml_tensor * ggml_reshape(
@@ -1198,14 +1289,15 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements
+    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
     // if mode & 2 == 1, GPT-NeoX style
     // if mode & 4 == 1, ChatGLM style
-    // TODO: avoid creating a new tensor every time
+    //
+    // b is an int32 vector with size a->ne[2], it contains the positions
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1214,7 +1306,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx);
@@ -1223,7 +1315,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1234,7 +1326,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1245,7 +1337,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             float                 base,
             bool                  down);
@@ -1255,7 +1347,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_rope_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_past,
+            struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1298,6 +1390,14 @@ extern "C" {
             int                   s,
             int                   d);
 
+    GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   s0,
+            int                   p0,
+            int                   d0);
+
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1603,7 +1703,8 @@ extern "C" {
             struct ggml_tensor  * tensor);
 
 
-    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
 
     GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
     GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
@@ -1634,6 +1735,16 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
+    // build gradient checkpointing backward graph gb for gf using provided checkpoints
+    // gb_tmp will contain original backward graph with rewritten backward process nodes,
+    // but without the second forward pass nodes.
+    GGML_API void ggml_build_backward_gradient_checkpointing(
+            struct ggml_context   * ctx,
+            struct ggml_cgraph    * gf,
+            struct ggml_cgraph    * gb,
+            struct ggml_cgraph    * gb_tmp,
+            struct ggml_tensor  * * checkpoints,
+            int                     n_checkpoints);
     //
     // optimization
     //
@@ -1660,6 +1771,7 @@ extern "C" {
         GGML_OPT_NO_CONTEXT,
         GGML_OPT_INVALID_WOLFE,
         GGML_OPT_FAIL,
+        GGML_OPT_CANCEL,
 
         GGML_LINESEARCH_FAIL = -128,
         GGML_LINESEARCH_MINIMUM_STEP,
@@ -1668,6 +1780,9 @@ extern "C" {
         GGML_LINESEARCH_INVALID_PARAMETERS,
     };
 
+    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
+
     // optimization parameters
     //
     //   see ggml.c (ggml_opt_default_params) for default values
@@ -1697,18 +1812,22 @@ extern "C" {
         bool print_forward_graph;
         bool print_backward_graph;
 
+        int n_gradient_accumulation;
+
         // ADAM parameters
         struct {
             int n_iter;
 
             float sched; // schedule multiplier (fixed, decay or warmup)
             float decay; // weight decay for AdamW, use 0.0f to disable
+            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
             float alpha; // learning rate
             float beta1;
             float beta2;
             float eps;   // epsilon for numerical stability
             float eps_f; // epsilon for convergence test
             float eps_g; // epsilon for convergence test
+            float gclip; // gradient clipping
         } adam;
 
         // LBFGS parameters
@@ -1736,14 +1855,13 @@ extern "C" {
 
         bool just_initialized;
 
+        float loss_before;
+        float loss_after;
+
         struct {
-            struct ggml_tensor * x;  // view of the parameters
-            struct ggml_tensor * g1; // gradient
-            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * g;  // current gradient
             struct ggml_tensor * m;  // first moment
             struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * mh; // first moment hat
-            struct ggml_tensor * vh; // second moment hat
             struct ggml_tensor * pf; // past function values
             float fx_best;
             float fx_prev;
@@ -1780,10 +1898,10 @@ extern "C" {
 
     // initialize optimizer context
     GGML_API void ggml_opt_init(
-            struct ggml_context * ctx,
+            struct ggml_context     * ctx,
             struct ggml_opt_context * opt,
-            struct ggml_opt_params params,
-            int64_t nx);
+            struct ggml_opt_params    params,
+            int64_t                   nx);
 
     // continue optimizing the function defined by the tensor f
     GGML_API enum ggml_opt_result ggml_opt_resume(
@@ -1797,7 +1915,9 @@ extern "C" {
             struct ggml_opt_context * opt,
             struct ggml_tensor * f,
             struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb);
+            struct ggml_cgraph * gb,
+            ggml_opt_callback callback,
+            void * callback_data);
 
     //
     // quantization
@@ -1826,6 +1946,9 @@ extern "C" {
         GGUF_TYPE_BOOL    = 7,
         GGUF_TYPE_STRING  = 8,
         GGUF_TYPE_ARRAY   = 9,
+        GGUF_TYPE_UINT64  = 10,
+        GGUF_TYPE_INT64   = 11,
+        GGUF_TYPE_FLOAT64 = 12,
         GGUF_TYPE_COUNT,       // marks the end of the enum
     };
 
@@ -1846,36 +1969,39 @@ extern "C" {
 
     GGML_API const char * gguf_type_name(enum gguf_type type);
 
-    GGML_API int    gguf_get_version    (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_alignment  (struct gguf_context * ctx);
-    GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx);
-    GGML_API void * gguf_get_data       (struct gguf_context * ctx);
+    GGML_API int    gguf_get_version    (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_alignment  (const struct gguf_context * ctx);
+    GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx);
+    GGML_API void * gguf_get_data       (const struct gguf_context * ctx);
 
-    GGML_API int          gguf_get_n_kv(struct gguf_context * ctx);
-    GGML_API int          gguf_find_key(struct gguf_context * ctx, const char * key);
-    GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i);
+    GGML_API int          gguf_get_n_kv(const struct gguf_context * ctx);
+    GGML_API int          gguf_find_key(const struct gguf_context * ctx, const char * key);
+    GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int key_id);
 
-    GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i);
-    GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i);
+    GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int key_id);
+    GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id);
 
-    // results are undefined if the wrong type is used for the key
-    GGML_API uint8_t      gguf_get_val_u8  (struct gguf_context * ctx, int i);
-    GGML_API int8_t       gguf_get_val_i8  (struct gguf_context * ctx, int i);
-    GGML_API uint16_t     gguf_get_val_u16 (struct gguf_context * ctx, int i);
-    GGML_API int16_t      gguf_get_val_i16 (struct gguf_context * ctx, int i);
-    GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);
-    GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);
-    GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);
-    GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);
-    GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);
-    GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i);
-    GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i);
+    // will abort if the wrong type is used for the key
+    GGML_API uint8_t      gguf_get_val_u8  (const struct gguf_context * ctx, int key_id);
+    GGML_API int8_t       gguf_get_val_i8  (const struct gguf_context * ctx, int key_id);
+    GGML_API uint16_t     gguf_get_val_u16 (const struct gguf_context * ctx, int key_id);
+    GGML_API int16_t      gguf_get_val_i16 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint32_t     gguf_get_val_u32 (const struct gguf_context * ctx, int key_id);
+    GGML_API int32_t      gguf_get_val_i32 (const struct gguf_context * ctx, int key_id);
+    GGML_API float        gguf_get_val_f32 (const struct gguf_context * ctx, int key_id);
+    GGML_API uint64_t     gguf_get_val_u64 (const struct gguf_context * ctx, int key_id);
+    GGML_API int64_t      gguf_get_val_i64 (const struct gguf_context * ctx, int key_id);
+    GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
+    GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
+    GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);
 
-    GGML_API int    gguf_get_n_tensors    (struct gguf_context * ctx);
-    GGML_API int    gguf_find_tensor      (struct gguf_context * ctx, const char * name);
-    GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i);
-    GGML_API char * gguf_get_tensor_name  (struct gguf_context * ctx, int i);
+    GGML_API int    gguf_get_n_tensors    (const struct gguf_context * ctx);
+    GGML_API int    gguf_find_tensor      (const struct gguf_context * ctx, const char * name);
+    GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int i);
+    GGML_API char * gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
 
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
@@ -1885,6 +2011,9 @@ extern "C" {
     GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);
     GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t  val);
     GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float    val);
+    GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val);
+    GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t  val);
+    GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double   val);
     GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool     val);
     GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);
     GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);
@@ -1917,11 +2046,11 @@ extern "C" {
     //
 
     // write the entire context to a binary file
-    GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta);
+    GGML_API void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta);
 
     // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
-    GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx);
-    GGML_API void   gguf_get_meta_data(struct gguf_context * ctx, void * data);
+    GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
+    GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
     //
     // system info
@@ -1935,6 +2064,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
+    GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);
     GGML_API int ggml_cpu_has_fp16_va    (void);
     GGML_API int ggml_cpu_has_wasm_simd  (void);
@@ -1943,6 +2073,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_clblast    (void);
     GGML_API int ggml_cpu_has_gpublas    (void);
     GGML_API int ggml_cpu_has_sse3       (void);
+    GGML_API int ggml_cpu_has_ssse3      (void);
     GGML_API int ggml_cpu_has_vsx        (void);
 
     //
diff --git a/gguf-py/LICENSE b/gguf-py/LICENSE
new file mode 100644
index 000000000..76f67efdc
--- /dev/null
+++ b/gguf-py/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/gguf-py/README.md b/gguf-py/README.md
new file mode 100644
index 000000000..a28d8c57a
--- /dev/null
+++ b/gguf-py/README.md
@@ -0,0 +1,71 @@
+## gguf
+
+This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302)
+(GGML Universal File) format.
+
+See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py)
+as an example for its usage.
+
+## Installation
+```sh
+pip install gguf
+```
+
+## Development
+Maintainers who participate in development of this package are advised to install it in editable mode:
+
+```sh
+cd /path/to/llama.cpp/gguf-py
+
+pip install --editable .
+```
+
+**Note**: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires `setup.py`.
+In this case, upgrade Pip to the latest:
+
+```sh
+pip install --upgrade pip
+```
+
+## Automatic publishing with CI
+
+There's a GitHub workflow to make a release automatically upon creation of tags in a specified format.
+
+1. Bump the version in `pyproject.toml`.
+2. Create a tag named `gguf-vx.x.x` where `x.x.x` is the semantic version number.
+
+```sh
+git tag -a gguf-v1.0.0 -m "Version 1.0 release"
+```
+
+3. Push the tags.
+
+```sh
+git push origin --tags
+```
+
+## Manual publishing
+If you want to publish the package manually for any reason, you need to have `twine` and `build` installed:
+
+```sh
+pip install build twine
+```
+
+Then, folow these steps to release a new version:
+
+1. Bump the version in `pyproject.toml`.
+2. Build the package:
+
+```sh
+python -m build
+```
+
+3. Upload the generated distribution archives:
+
+```sh
+python -m twine upload dist/*
+```
+
+## TODO
+- [ ] Add tests
+- [ ] Include conversion scripts as command line entry points in this package.
diff --git a/gguf-py/gguf/__init__.py b/gguf-py/gguf/__init__.py
new file mode 100644
index 000000000..f9b70a85b
--- /dev/null
+++ b/gguf-py/gguf/__init__.py
@@ -0,0 +1 @@
+from .gguf import *
diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py
new file mode 100644
index 000000000..fb677a6ed
--- /dev/null
+++ b/gguf-py/gguf/gguf.py
@@ -0,0 +1,1042 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import struct
+import sys
+import tempfile
+from enum import IntEnum, auto
+from io import BufferedWriter
+from pathlib import Path
+from typing import IO, Any, BinaryIO, Callable, Sequence
+
+import numpy as np
+
+#
+# constants
+#
+
+GGUF_MAGIC             = 0x46554747
+GGUF_VERSION           = 2
+GGUF_DEFAULT_ALIGNMENT = 32
+
+# general
+KEY_GENERAL_ARCHITECTURE         = "general.architecture"
+KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
+KEY_GENERAL_ALIGNMENT            = "general.alignment"
+KEY_GENERAL_NAME                 = "general.name"
+KEY_GENERAL_AUTHOR               = "general.author"
+KEY_GENERAL_URL                  = "general.url"
+KEY_GENERAL_DESCRIPTION          = "general.description"
+KEY_GENERAL_LICENSE              = "general.license"
+KEY_GENERAL_SOURCE_URL           = "general.source.url"
+KEY_GENERAL_SOURCE_HF_REPO       = "general.source.huggingface.repository"
+KEY_GENERAL_FILE_TYPE            = "general.file_type"
+
+# LLM
+KEY_CONTEXT_LENGTH        = "{arch}.context_length"
+KEY_EMBEDDING_LENGTH      = "{arch}.embedding_length"
+KEY_BLOCK_COUNT           = "{arch}.block_count"
+KEY_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
+KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
+KEY_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
+
+# attention
+KEY_ATTENTION_HEAD_COUNT        = "{arch}.attention.head_count"
+KEY_ATTENTION_HEAD_COUNT_KV     = "{arch}.attention.head_count_kv"
+KEY_ATTENTION_MAX_ALIBI_BIAS    = "{arch}.attention.max_alibi_bias"
+KEY_ATTENTION_CLAMP_KQV         = "{arch}.attention.clamp_kqv"
+KEY_ATTENTION_LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
+KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
+
+# RoPE
+KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count"
+KEY_ROPE_FREQ_BASE       = "{arch}.rope.freq_base"
+KEY_ROPE_SCALE_LINEAR    = "{arch}.rope.scale_linear"
+
+# tokenization
+KEY_TOKENIZER_MODEL      = "tokenizer.ggml.model"
+KEY_TOKENIZER_LIST       = "tokenizer.ggml.tokens"
+KEY_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type"
+KEY_TOKENIZER_SCORES     = "tokenizer.ggml.scores"
+KEY_TOKENIZER_MERGES     = "tokenizer.ggml.merges"
+KEY_TOKENIZER_BOS_ID     = "tokenizer.ggml.bos_token_id"
+KEY_TOKENIZER_EOS_ID     = "tokenizer.ggml.eos_token_id"
+KEY_TOKENIZER_UNK_ID     = "tokenizer.ggml.unknown_token_id"
+KEY_TOKENIZER_SEP_ID     = "tokenizer.ggml.seperator_token_id"
+KEY_TOKENIZER_PAD_ID     = "tokenizer.ggml.padding_token_id"
+KEY_TOKENIZER_HF_JSON    = "tokenizer.huggingface.json"
+KEY_TOKENIZER_RWKV       = "tokenizer.rwkv.world"
+
+
+#
+# recommended mapping of model tensor names for storage in gguf
+#
+
+
+class MODEL_ARCH(IntEnum):
+    LLAMA         : int = auto()
+    FALCON        : int = auto()
+    BAICHUAN      : int = auto()
+    GPT2          : int = auto()
+    GPTJ          : int = auto()
+    GPTNEOX       : int = auto()
+    MPT           : int = auto()
+    STARCODER     : int = auto()
+    PERSIMMON     : int = auto()
+    REFACT        : int = auto()
+    BERT          : int = auto()
+
+
+class MODEL_TENSOR(IntEnum):
+    TOKEN_EMBD   : int = auto()
+    TOKEN_TYPES  : int = auto()
+    POS_EMBD     : int = auto()
+    OUTPUT       : int = auto()
+    OUTPUT_NORM  : int = auto()
+    ROPE_FREQS   : int = auto()
+    ATTN_Q       : int = auto()
+    ATTN_K       : int = auto()
+    ATTN_V       : int = auto()
+    ATTN_QKV     : int = auto()
+    ATTN_OUT     : int = auto()
+    ATTN_NORM    : int = auto()
+    ATTN_NORM_2  : int = auto()
+    ATTN_ROT_EMBD: int = auto()
+    FFN_GATE     : int = auto()
+    FFN_DOWN     : int = auto()
+    FFN_UP       : int = auto()
+    FFN_NORM     : int = auto()
+    ATTN_Q_NORM  : int = auto()
+    ATTN_K_NORM  : int = auto()
+
+
+MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+    MODEL_ARCH.LLAMA:          "llama",
+    MODEL_ARCH.FALCON:         "falcon",
+    MODEL_ARCH.BAICHUAN:       "baichuan",
+    MODEL_ARCH.GPT2:           "gpt2",
+    MODEL_ARCH.GPTJ:           "gptj",
+    MODEL_ARCH.GPTNEOX:        "gptneox",
+    MODEL_ARCH.MPT:            "mpt",
+    MODEL_ARCH.STARCODER:      "starcoder",
+    MODEL_ARCH.PERSIMMON:      "persimmon",
+    MODEL_ARCH.REFACT:         "refact",
+    MODEL_ARCH.BERT:           "bert",
+}
+
+TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
+    MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
+    MODEL_TENSOR.TOKEN_TYPES:   "token_types",
+    MODEL_TENSOR.POS_EMBD:      "position_embd",
+    MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
+    MODEL_TENSOR.OUTPUT:        "output",
+    MODEL_TENSOR.ROPE_FREQS:    "rope_freqs",
+    MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
+    MODEL_TENSOR.ATTN_NORM_2:   "blk.{bid}.attn_norm_2",
+    MODEL_TENSOR.ATTN_QKV:      "blk.{bid}.attn_qkv",
+    MODEL_TENSOR.ATTN_Q:        "blk.{bid}.attn_q",
+    MODEL_TENSOR.ATTN_K:        "blk.{bid}.attn_k",
+    MODEL_TENSOR.ATTN_V:        "blk.{bid}.attn_v",
+    MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
+    MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
+    MODEL_TENSOR.ATTN_Q_NORM:   "blk.{bid}.attn_q_norm",
+    MODEL_TENSOR.ATTN_K_NORM:   "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
+    MODEL_TENSOR.FFN_GATE:      "blk.{bid}.ffn_gate",
+    MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
+    MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
+}
+
+MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+    MODEL_ARCH.LLAMA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GPTNEOX: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.FALCON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.STARCODER: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.MPT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GPTJ: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.PERSIMMON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.REFACT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GPT2: [
+        # TODO
+    ],
+    # TODO
+}
+
+# tensors that will not be serialized
+MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+    MODEL_ARCH.LLAMA: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.BAICHUAN: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
+    MODEL_ARCH.PERSIMMON: [
+        MODEL_TENSOR.ROPE_FREQS,
+    ]
+}
+
+
+class TensorNameMap:
+    mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
+        # Token embeddings
+        MODEL_TENSOR.TOKEN_EMBD: (
+            "gpt_neox.embed_in",                        # gptneox
+            "transformer.wte",                          # gpt2 gpt-j mpt refact
+            "transformer.word_embeddings",              # falcon
+            "model.embed_tokens",                       # llama-hf
+            "tok_embeddings",                           # llama-pth
+            "embeddings.word_embeddings",               # bert
+            "language_model.embedding.word_embeddings", # persimmon
+        ),
+
+        # Token type embeddings
+        MODEL_TENSOR.TOKEN_TYPES: (
+            "embeddings.token_type_embeddings",  # bert
+        ),
+
+        # Position embeddings
+        MODEL_TENSOR.POS_EMBD: (
+            "transformer.wpe",                 # gpt2
+            "embeddings.position_embeddings",  # bert
+        ),
+
+        # Output
+        MODEL_TENSOR.OUTPUT: (
+            "embed_out",                # gptneox
+            "lm_head",                  # gpt2 mpt falcon llama-hf baichuan
+            "output",                   # llama-pth
+            "word_embeddings_for_head", # persimmon
+        ),
+
+        # Output norm
+        MODEL_TENSOR.OUTPUT_NORM: (
+            "gpt_neox.final_layer_norm",              # gptneox
+            "transformer.ln_f",                       # gpt2 gpt-j falcon
+            "model.norm",                             # llama-hf baichuan
+            "norm",                                   # llama-pth
+            "embeddings.LayerNorm",                   # bert
+            "transformer.norm_f",                     # mpt
+            "ln_f",                                   # refact
+            "language_model.encoder.final_layernorm", # persimmon
+        ),
+
+        # Rope frequencies
+        MODEL_TENSOR.ROPE_FREQS: (
+            "rope.freqs", # llama-pth
+        ),
+    }
+
+    block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
+        # Attention norm
+        MODEL_TENSOR.ATTN_NORM: (
+            "gpt_neox.layers.{bid}.input_layernorm",               # gptneox
+            "transformer.h.{bid}.ln_1",                            # gpt2 gpt-j refact
+            "transformer.blocks.{bid}.norm_1",                     # mpt
+            "transformer.h.{bid}.input_layernorm",                 # falcon7b
+            "transformer.h.{bid}.ln_mlp",                          # falcon40b
+            "model.layers.{bid}.input_layernorm",                  # llama-hf
+            "layers.{bid}.attention_norm",                         # llama-pth
+            "encoder.layer.{bid}.attention.output.LayerNorm",      # bert
+            "language_model.encoder.layers.{bid}.input_layernorm", # persimmon
+        ),
+
+        # Attention norm 2
+        MODEL_TENSOR.ATTN_NORM_2: (
+            "transformer.h.{bid}.ln_attn", # falcon40b
+        ),
+
+        # Attention query-key-value
+        MODEL_TENSOR.ATTN_QKV: (
+            "gpt_neox.layers.{bid}.attention.query_key_value",                    # gptneox
+            "transformer.h.{bid}.attn.c_attn",                                    # gpt2
+            "transformer.blocks.{bid}.attn.Wqkv",                                 # mpt
+            "transformer.h.{bid}.self_attention.query_key_value",                 # falcon
+            "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
+        ),
+
+        # Attention query
+        MODEL_TENSOR.ATTN_Q: (
+            "model.layers.{bid}.self_attn.q_proj",       # llama-hf
+            "layers.{bid}.attention.wq",                 # llama-pth
+            "encoder.layer.{bid}.attention.self.query",  # bert
+            "transformer.h.{bid}.attn.q_proj",           # gpt-j
+        ),
+
+        # Attention key
+        MODEL_TENSOR.ATTN_K: (
+            "model.layers.{bid}.self_attn.k_proj",     # llama-hf
+            "layers.{bid}.attention.wk",               # llama-pth
+            "encoder.layer.{bid}.attention.self.key",  # bert
+            "transformer.h.{bid}.attn.k_proj",         # gpt-j
+        ),
+
+        # Attention value
+        MODEL_TENSOR.ATTN_V: (
+            "model.layers.{bid}.self_attn.v_proj",       # llama-hf
+            "layers.{bid}.attention.wv",                 # llama-pth
+            "encoder.layer.{bid}.attention.self.value",  # bert
+            "transformer.h.{bid}.attn.v_proj",           # gpt-j
+        ),
+
+        # Attention output
+        MODEL_TENSOR.ATTN_OUT: (
+            "gpt_neox.layers.{bid}.attention.dense",                   # gptneox
+            "transformer.h.{bid}.attn.c_proj",                         # gpt2 refact
+            "transformer.blocks.{bid}.attn.out_proj",                  # mpt
+            "transformer.h.{bid}.self_attention.dense",                # falcon
+            "model.layers.{bid}.self_attn.o_proj",                     # llama-hf
+            "layers.{bid}.attention.wo",                               # llama-pth
+            "encoder.layer.{bid}.attention.output.dense",              # bert
+            "transformer.h.{bid}.attn.out_proj",                       # gpt-j
+            "language_model.encoder.layers.{bid}.self_attention.dense" # persimmon
+        ),
+
+        # Rotary embeddings
+        MODEL_TENSOR.ATTN_ROT_EMBD: (
+            "model.layers.{bid}.self_attn.rotary_emb.inv_freq",  # llama-hf
+            "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth
+        ),
+
+        # Feed-forward norm
+        MODEL_TENSOR.FFN_NORM: (
+            "gpt_neox.layers.{bid}.post_attention_layernorm",               # gptneox
+            "transformer.h.{bid}.ln_2",                                     # gpt2 refact
+            "transformer.blocks.{bid}.norm_2",                              # mpt
+            "model.layers.{bid}.post_attention_layernorm",                  # llama-hf
+            "layers.{bid}.ffn_norm",                                        # llama-pth
+            "encoder.layer.{bid}.output.LayerNorm",                         # bert
+            "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
+        ),
+
+        # Feed-forward up
+        MODEL_TENSOR.FFN_UP: (
+            "gpt_neox.layers.{bid}.mlp.dense_h_to_4h",               # gptneox
+            "transformer.h.{bid}.mlp.c_fc",                          # gpt2
+            "transformer.blocks.{bid}.ffn.up_proj",                  # mpt
+            "transformer.h.{bid}.mlp.dense_h_to_4h",                 # falcon
+            "model.layers.{bid}.mlp.up_proj",                        # llama-hf refact
+            "layers.{bid}.feed_forward.w3",                          # llama-pth
+            "encoder.layer.{bid}.intermediate.dense",                # bert
+            "transformer.h.{bid}.mlp.fc_in",                         # gpt-j
+            "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
+        ),
+
+        # Feed-forward gate
+        MODEL_TENSOR.FFN_GATE: (
+            "model.layers.{bid}.mlp.gate_proj", # llama-hf refact
+            "layers.{bid}.feed_forward.w1",     # llama-pth
+        ),
+
+        # Feed-forward down
+        MODEL_TENSOR.FFN_DOWN: (
+            "gpt_neox.layers.{bid}.mlp.dense_4h_to_h",               # gptneox
+            "transformer.h.{bid}.mlp.c_proj",                        # gpt2 refact
+            "transformer.blocks.{bid}.ffn.down_proj",                # mpt
+            "transformer.h.{bid}.mlp.dense_4h_to_h",                 # falcon
+            "model.layers.{bid}.mlp.down_proj",                      # llama-hf
+            "layers.{bid}.feed_forward.w2",                          # llama-pth
+            "encoder.layer.{bid}.output.dense",                      # bert
+            "transformer.h.{bid}.mlp.fc_out",                        # gpt-j
+            "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
+        ),
+
+        MODEL_TENSOR.ATTN_Q_NORM: (
+            "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
+        ),
+
+        MODEL_TENSOR.ATTN_K_NORM: (
+            "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
+        ),
+
+        MODEL_TENSOR.ROPE_FREQS: (
+            "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon
+        )
+    }
+
+    mapping: dict[str, tuple[MODEL_TENSOR, str]]
+
+    def __init__(self, arch: MODEL_ARCH, n_blocks: int):
+        self.mapping = {}
+        for tensor, keys in self.mappings_cfg.items():
+            if tensor not in MODEL_TENSORS[arch]:
+                continue
+            tensor_name = TENSOR_NAMES[tensor]
+            self.mapping[tensor_name] = (tensor, tensor_name)
+            for key in keys:
+                self.mapping[key] = (tensor, tensor_name)
+        for bid in range(n_blocks):
+            for tensor, keys in self.block_mappings_cfg.items():
+                if tensor not in MODEL_TENSORS[arch]:
+                    continue
+                tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
+                self.mapping[tensor_name] = (tensor, tensor_name)
+                for key in keys:
+                    key = key.format(bid = bid)
+                    self.mapping[key] = (tensor, tensor_name)
+
+    def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
+        result = self.mapping.get(key)
+        if result is not None:
+            return result
+        for suffix in try_suffixes:
+            if key.endswith(suffix):
+                result = self.mapping.get(key[:-len(suffix)])
+                if result is not None:
+                    return (result[0], result[1] + suffix)
+        return None
+
+    def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[1]
+
+    def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None:
+        result = self.get_type_and_name(key, try_suffixes = try_suffixes)
+        if result is None:
+            return None
+        return result[0]
+
+    def __getitem__(self, key: str) -> str:
+        try:
+            return self.mapping[key][1]
+        except KeyError:
+            raise KeyError(key)
+
+    def __contains__(self, key: str) -> bool:
+        return key in self.mapping
+
+    def __repr__(self) -> str:
+        return repr(self.mapping)
+
+def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap:
+    return TensorNameMap(arch, n_blocks)
+
+class TokenType(IntEnum):
+    NORMAL       = 1
+    UNKNOWN      = 2
+    CONTROL      = 3
+    USER_DEFINED = 4
+    UNUSED       = 5
+    BYTE         = 6
+
+#
+# implementation
+#
+
+
+class GGMLQuantizationType(IntEnum):
+    F32  = 0
+    F16  = 1
+    Q4_0 = 2
+    Q4_1 = 3
+    Q5_0 = 6
+    Q5_1 = 7
+    Q8_0 = 8
+    Q8_1 = 9
+    Q2_K = 10
+    Q3_K = 11
+    Q4_K = 12
+    Q5_K = 13
+    Q6_K = 14
+    Q8_K = 15
+
+
+class GGUFValueType(IntEnum):
+    UINT8   = 0
+    INT8    = 1
+    UINT16  = 2
+    INT16   = 3
+    UINT32  = 4
+    INT32   = 5
+    FLOAT32 = 6
+    BOOL    = 7
+    STRING  = 8
+    ARRAY   = 9
+    UINT64  = 10
+    INT64   = 11
+    FLOAT64 = 12
+
+    @staticmethod
+    def get_type(val):
+        if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray):
+            return GGUFValueType.STRING
+        elif isinstance(val, list):
+            return GGUFValueType.ARRAY
+        elif isinstance(val, float):
+            return GGUFValueType.FLOAT32
+        elif isinstance(val, bool):
+            return GGUFValueType.BOOL
+        elif isinstance(val, int):
+            return GGUFValueType.INT32
+        # TODO: need help with 64-bit types in Python
+        else:
+            print("Unknown type: "+str(type(val)))
+            sys.exit()
+
+
+class GGUFWriter:
+    fout: BufferedWriter
+    arch: str
+    offset_tensor = 0
+    data_alignment = GGUF_DEFAULT_ALIGNMENT
+    kv_data = b""
+    kv_data_count = 0
+    ti_data = b""
+    ti_data_count = 0
+    use_temp_file: bool
+    temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None
+    tensors: list[tuple[np.ndarray[Any, Any], int]]
+
+    def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True):
+        self.fout = open(path, "wb")
+        self.arch = arch
+        self.add_architecture()
+        self.use_temp_file = use_temp_file
+        self.tensors = []
+
+    def write_header_to_file(self):
+        self.fout.write(struct.pack("<I", GGUF_MAGIC))
+        self.fout.write(struct.pack("<I", GGUF_VERSION))
+        self.fout.write(struct.pack("<Q", self.ti_data_count))
+        self.fout.write(struct.pack("<Q", self.kv_data_count))
+        self.flush()
+#        print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
+
+    def write_kv_data_to_file(self):
+        self.fout.write(self.kv_data)
+        self.flush()
+
+    def write_ti_data_to_file(self):
+        self.fout.write(self.ti_data)
+        self.flush()
+
+    def add_key(self, key: str):
+        self.add_val(key, GGUFValueType.STRING, add_vtype=False)
+
+    def add_uint8(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT8)
+
+    def add_int8(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT8)
+
+    def add_uint16(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT16)
+
+    def add_int16(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT16)
+
+    def add_uint32(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT32)
+
+    def add_int32(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT32)
+
+    def add_float32(self, key: str, val: float):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.FLOAT32)
+
+    def add_uint64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.UINT64)
+
+    def add_int64(self, key: str, val: int):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.INT64)
+
+    def add_float64(self, key: str, val: float):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.FLOAT64)
+
+    def add_bool(self, key: str, val: bool):
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.BOOL)
+
+    def add_string(self, key: str, val: str):
+        if len(val) == 0:
+            return
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.STRING)
+
+    def add_array(self, key: str, val: Sequence[Any]):
+        if not isinstance(val, Sequence):
+            raise ValueError("Value must be a sequence for array type")
+
+        self.add_key(key)
+        self.add_val(val, GGUFValueType.ARRAY)
+
+    _simple_value_packing = {
+        GGUFValueType.UINT8:   "<B",
+        GGUFValueType.INT8:    "<b",
+        GGUFValueType.UINT16:  "<H",
+        GGUFValueType.INT16:   "<h",
+        GGUFValueType.UINT32:  "<I",
+        GGUFValueType.INT32:   "<i",
+        GGUFValueType.FLOAT32: "<f",
+        GGUFValueType.UINT64:  "<Q",
+        GGUFValueType.INT64:   "<q",
+        GGUFValueType.FLOAT64: "<d",
+        GGUFValueType.BOOL:    "?" ,
+    }
+    def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True):
+        if vtype is None:
+            vtype = GGUFValueType.get_type(val)
+
+        if add_vtype:
+            self.kv_data += struct.pack("<I", vtype)
+            self.kv_data_count += 1
+
+        pack_fmt = self._simple_value_packing.get(vtype)
+        if pack_fmt is not None:
+            self.kv_data += struct.pack(pack_fmt, val)
+        elif vtype == GGUFValueType.STRING:
+            encoded_val = val.encode("utf8") if isinstance(val, str) else val
+            self.kv_data += struct.pack("<Q", len(encoded_val))
+            self.kv_data += encoded_val
+        elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and len(val) > 0:
+            ltype = GGUFValueType.get_type(val[0])
+            if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
+                raise ValueError("All items in a GGUF array should be of the same type")
+            self.kv_data += struct.pack("<I", ltype)
+            self.kv_data += struct.pack("<Q", len(val))
+            for item in val:
+                self.add_val(item, add_vtype=False)
+        else:
+            raise ValueError("Invalid GGUF metadata value type or value")
+
+    @staticmethod
+    def ggml_pad(x: int, n: int) -> int:
+        return ((x + n - 1) // n) * n
+
+    def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None):
+        assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
+
+        encoded_name = name.encode("utf8")
+        self.ti_data += struct.pack("<Q", len(encoded_name))
+        self.ti_data += encoded_name
+        n_dims = len(tensor_shape)
+        self.ti_data += struct.pack("<I", n_dims)
+        for i in range(n_dims):
+            self.ti_data += struct.pack("<Q", tensor_shape[n_dims - 1 - i])
+        if raw_dtype is None:
+            dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
+        else:
+            dtype = raw_dtype
+        self.ti_data += struct.pack("<I", dtype)
+        self.ti_data += struct.pack("<Q", self.offset_tensor)
+        self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
+        self.ti_data_count += 1
+
+    def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, raw_dtype: GGMLQuantizationType | None = None):
+        if self.use_temp_file and self.temp_file is None:
+            fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
+            fp.seek(0)
+            self.temp_file = fp
+
+        shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape
+        self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype)
+
+        pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
+
+        if  self.temp_file is None:
+            self.tensors.append((tensor, pad))
+            return
+
+        tensor.tofile(self.temp_file)
+
+        if pad != 0:
+            self.temp_file.write(bytes([0] * pad))
+
+    def write_padding(self, fp: BinaryIO, n: int, align: int | None = None):
+        pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n
+        if pad != 0:
+            fp.write(bytes([0] * pad))
+
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any]):
+        self.write_padding(self.fout, self.fout.tell())
+        tensor.tofile(self.fout)
+        self.write_padding(self.fout, tensor.nbytes)
+
+    def write_tensors_to_file(self):
+        self.write_ti_data_to_file()
+
+        self.write_padding(self.fout, self.fout.tell())
+
+        if self.temp_file is None:
+            for (currtensor, currpad) in self.tensors:
+                currtensor.tofile(self.fout)
+                if currpad != 0:
+                    self.fout.write(bytes([0] * currpad))
+            return
+
+        self.temp_file.seek(0)
+
+        shutil.copyfileobj(self.temp_file, self.fout)
+        self.flush()
+        self.temp_file.close()
+
+    def flush(self):
+        self.fout.flush()
+
+    def close(self):
+        self.fout.close()
+
+    def add_architecture(self):
+        self.add_string(KEY_GENERAL_ARCHITECTURE, self.arch)
+
+    def add_author(self, author: str):
+        self.add_string(KEY_GENERAL_AUTHOR, author)
+
+    def add_tensor_data_layout(self, layout: str):
+        self.add_string(KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
+
+    def add_url(self, url: str):
+        self.add_string(KEY_GENERAL_URL, url)
+
+    def add_description(self, description: str):
+        self.add_string(KEY_GENERAL_DESCRIPTION, description)
+
+    def add_source_url(self, url: str):
+        self.add_string(KEY_GENERAL_SOURCE_URL, url)
+
+    def add_source_hf_repo(self, repo: str):
+        self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
+
+    def add_file_type(self, ftype: int):
+        self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
+
+    def add_name(self, name: str):
+        self.add_string(KEY_GENERAL_NAME, name)
+
+    def add_quantization_version(self, quantization_version: GGMLQuantizationType):
+        self.add_uint32(
+            KEY_GENERAL_QUANTIZATION_VERSION, quantization_version)
+
+    def add_custom_alignment(self, alignment: int):
+        self.data_alignment = alignment
+        self.add_uint32(KEY_GENERAL_ALIGNMENT, alignment)
+
+    def add_context_length(self, length: int):
+        self.add_uint32(
+            KEY_CONTEXT_LENGTH.format(arch=self.arch), length)
+
+    def add_embedding_length(self, length: int):
+        self.add_uint32(
+            KEY_EMBEDDING_LENGTH.format(arch=self.arch), length)
+
+    def add_block_count(self, length: int):
+        self.add_uint32(
+            KEY_BLOCK_COUNT.format(arch=self.arch), length)
+
+    def add_feed_forward_length(self, length: int):
+        self.add_uint32(
+            KEY_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
+
+    def add_parallel_residual(self, use: bool):
+        self.add_bool(
+            KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
+
+    def add_head_count(self, count: int):
+        self.add_uint32(
+            KEY_ATTENTION_HEAD_COUNT.format(arch=self.arch), count)
+
+    def add_head_count_kv(self, count: int):
+        self.add_uint32(
+            KEY_ATTENTION_HEAD_COUNT_KV.format(arch=self.arch), count)
+
+    def add_max_alibi_bias(self, bias: float):
+        self.add_float32(
+            KEY_ATTENTION_MAX_ALIBI_BIAS.format(arch=self.arch), bias)
+
+    def add_clamp_kqv(self, value: float):
+        self.add_float32(
+            KEY_ATTENTION_CLAMP_KQV.format(arch=self.arch), value)
+
+    def add_layer_norm_eps(self, value: float):
+        self.add_float32(
+            KEY_ATTENTION_LAYERNORM_EPS.format(arch=self.arch), value)
+
+    def add_layer_norm_rms_eps(self, value: float):
+        self.add_float32(
+            KEY_ATTENTION_LAYERNORM_RMS_EPS.format(arch=self.arch), value)
+
+    def add_rope_dimension_count(self, count: int):
+        self.add_uint32(
+            KEY_ROPE_DIMENSION_COUNT.format(arch=self.arch), count)
+
+    def add_rope_freq_base(self, value: float):
+        self.add_float32(KEY_ROPE_FREQ_BASE.format(arch=self.arch), value)
+
+    def add_rope_scale_linear(self, value: float):
+        self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value)
+
+    def add_tokenizer_model(self, model: str):
+        self.add_string(KEY_TOKENIZER_MODEL, model)
+
+    def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]):
+        self.add_array(KEY_TOKENIZER_LIST, tokens)
+
+    def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]):
+        self.add_array(KEY_TOKENIZER_MERGES, merges)
+
+    def add_token_types(self, types: Sequence[TokenType] | Sequence[int]):
+        self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types)
+
+    def add_token_scores(self, scores: Sequence[float]):
+        self.add_array(KEY_TOKENIZER_SCORES, scores)
+
+    def add_bos_token_id(self, id: int):
+        self.add_uint32(KEY_TOKENIZER_BOS_ID, id)
+
+    def add_eos_token_id(self, id: int):
+        self.add_uint32(KEY_TOKENIZER_EOS_ID, id)
+
+    def add_unk_token_id(self, id: int):
+        self.add_uint32(KEY_TOKENIZER_UNK_ID, id)
+
+    def add_sep_token_id(self, id: int):
+        self.add_uint32(KEY_TOKENIZER_SEP_ID, id)
+
+    def add_pad_token_id(self, id: int):
+        self.add_uint32(KEY_TOKENIZER_PAD_ID, id)
+
+
+class SpecialVocab:
+    load_merges: bool = False
+    merges: list[str] = []
+    special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad')
+    special_token_ids: dict[str, int] = {}
+
+    def __init__(
+        self, path: str | os.PathLike[str], load_merges: bool = False,
+        special_token_types: tuple[str, ...] | None = None,
+    ):
+        self.special_token_ids = {}
+        self.load_merges = load_merges
+        if special_token_types is not None:
+            self.special_token_types = special_token_types
+        self._load(Path(path))
+
+    def _load(self, path: Path) -> None:
+        if not self._try_load_from_tokenizer_json(path):
+            self._try_load_from_config_json(path)
+
+    def _try_load_from_tokenizer_json(self, path: Path) -> bool:
+        tokenizer_file = path / 'tokenizer.json'
+        if not tokenizer_file.is_file():
+            return False
+        with open(tokenizer_file, encoding = 'utf-8') as f:
+            tokenizer = json.load(f)
+        if self.load_merges:
+            merges = tokenizer.get('model', {}).get('merges')
+            if isinstance(merges, list) and len(merges) > 0 and isinstance(merges[0], str):
+                self.merges = merges
+        tokenizer_config_file = path / 'tokenizer_config.json'
+        added_tokens = tokenizer.get('added_tokens')
+        if added_tokens is None or not tokenizer_config_file.is_file():
+            return True
+        with open(tokenizer_config_file, encoding = 'utf-8') as f:
+            tokenizer_config = json.load(f)
+        for typ in self.special_token_types:
+            entry = tokenizer_config.get(f'{typ}_token')
+            if isinstance(entry, str):
+                tc_content = entry
+            elif isinstance(entry, dict):
+                entry_content = entry.get('content')
+                if not isinstance(entry_content, str):
+                    continue
+                tc_content = entry_content
+            else:
+                continue
+            for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content):
+                if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
+                    self.special_token_ids[typ] = maybe_token_id
+                break
+        return True
+
+    def _try_load_from_config_json(self, path: Path) -> bool:
+        config_file = path / 'config.json'
+        if not config_file.is_file():
+            return False
+        with open(config_file, encoding = 'utf-8') as f:
+            config = json.load(f)
+        for typ in self.special_token_types:
+            maybe_token_id = config.get(f'{typ}_token_id')
+            if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
+                self.special_token_ids[typ] = maybe_token_id
+        return True
+
+    def add_to_gguf(self, gw: GGUFWriter) -> None:
+        if len(self.merges) > 0:
+            print(f'gguf: Adding {len(self.merges)} merge(s).')
+            gw.add_token_merges(self.merges)
+        for typ, tokid in self.special_token_ids.items():
+            handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
+            if handler is None:
+                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping')
+                continue
+            print(f'gguf: Setting special token type {typ} to {tokid}')
+            handler(tokid)
+
+    def __repr__(self) -> str:
+        return f'<SpecialVocab with {len(self.merges)} merges and special tokens {self.special_token_ids or "unset"}>'
+
+
+# Example usage:
+if __name__ == "__main__":
+    # Example usage with a file
+    gguf_writer = GGUFWriter("example.gguf", "llama")
+
+    gguf_writer.add_architecture()
+    gguf_writer.add_block_count(12)
+    gguf_writer.add_uint32("answer", 42)  # Write a 32-bit integer
+    gguf_writer.add_float32("answer_in_float", 42.0)  # Write a 32-bit float
+    gguf_writer.add_custom_alignment(64)
+
+    tensor1 = np.ones((32,), dtype=np.float32) * 100.0
+    tensor2 = np.ones((64,), dtype=np.float32) * 101.0
+    tensor3 = np.ones((96,), dtype=np.float32) * 102.0
+
+    gguf_writer.add_tensor("tensor1", tensor1)
+    gguf_writer.add_tensor("tensor2", tensor2)
+    gguf_writer.add_tensor("tensor3", tensor3)
+
+    gguf_writer.write_header_to_file()
+    gguf_writer.write_kv_data_to_file()
+    gguf_writer.write_tensors_to_file()
+
+    gguf_writer.close()
diff --git a/gguf-py/gguf/py.typed b/gguf-py/gguf/py.typed
new file mode 100644
index 000000000..e69de29bb
diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml
new file mode 100644
index 000000000..07a7ab4dd
--- /dev/null
+++ b/gguf-py/pyproject.toml
@@ -0,0 +1,29 @@
+[tool.poetry]
+name = "gguf"
+version = "0.4.4"
+description = "Write ML models in GGUF for GGML"
+authors = ["GGML <ggml@ggml.ai>"]
+packages = [
+    {include = "gguf"},
+    {include = "gguf/py.typed"},
+]
+readme = "README.md"
+homepage = "https://ggml.ai"
+repository = "https://github.com/ggerganov/llama.cpp"
+keywords = ["ggml", "gguf", "llama.cpp"]
+classifiers = [
+    "Programming Language :: Python :: 3",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+]
+
+[tool.poetry.dependencies]
+python = ">=3.8"
+numpy = ">=1.17"
+
+[tool.poetry.dev-dependencies]
+pytest = "^5.2"
+
+[build-system]
+requires = ["poetry-core>=1.0.0"]
+build-backend = "poetry.core.masonry.api"
diff --git a/gguf-py/tests/test_gguf.py b/gguf-py/tests/test_gguf.py
new file mode 100644
index 000000000..512531dd2
--- /dev/null
+++ b/gguf-py/tests/test_gguf.py
@@ -0,0 +1,7 @@
+import gguf
+
+# TODO: add tests
+
+
+def test_write_gguf():
+    pass
diff --git a/gguf.py b/gguf.py
deleted file mode 100755
index 9421080b8..000000000
--- a/gguf.py
+++ /dev/null
@@ -1,723 +0,0 @@
-#!/usr/bin/env python3
-import shutil
-import sys
-import struct
-import tempfile
-import numpy as np
-
-from enum import IntEnum, auto
-from typing import Any, IO, List, Optional
-
-#
-# constants
-#
-
-GGUF_MAGIC             = 0x46554747
-GGUF_VERSION           = 1
-GGUF_DEFAULT_ALIGNMENT = 32
-
-# general
-KEY_GENERAL_ARCHITECTURE         = "general.architecture"
-KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
-KEY_GENERAL_ALIGNMENT            = "general.alignment"
-KEY_GENERAL_NAME                 = "general.name"
-KEY_GENERAL_AUTHOR               = "general.author"
-KEY_GENERAL_URL                  = "general.url"
-KEY_GENERAL_DESCRIPTION          = "general.description"
-KEY_GENERAL_LICENSE              = "general.license"
-KEY_GENERAL_SOURCE_URL           = "general.source.url"
-KEY_GENERAL_SOURCE_HF_REPO       = "general.source.hugginface.repository"
-KEY_GENERAL_FILE_TYPE            = "general.file_type"
-
-# LLM
-KEY_LLM_CONTEXT_LENGTH        = "{arch}.context_length"
-KEY_LLM_EMBEDDING_LENGTH      = "{arch}.embedding_length"
-KEY_LLM_BLOCK_COUNT           = "{arch}.block_count"
-KEY_LLM_FEED_FORWARD_LENGTH   = "{arch}.feed_forward_length"
-KEY_LLM_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual"
-KEY_LLM_TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
-
-# attention
-KEY_ATTENTION_HEAD_COUNT        = "{arch}.attention.head_count"
-KEY_ATTENTION_HEAD_COUNT_KV     = "{arch}.attention.head_count_kv"
-KEY_ATTENTION_MAX_ALIBI_BIAS    = "{arch}.attention.max_alibi_bias"
-KEY_ATTENTION_CLAMP_KQV         = "{arch}.attention.clamp_kqv"
-KEY_ATTENTION_LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
-KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
-
-# RoPE
-KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count"
-KEY_ROPE_SCALE_LINEAR    = "{arch}.rope.scale_linear"
-
-# tokenization
-KEY_TOKENIZER_MODEL      = "tokenizer.ggml.model"
-KEY_TOKENIZER_LIST       = "tokenizer.ggml.tokens"
-KEY_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type"
-KEY_TOKENIZER_SCORES     = "tokenizer.ggml.scores"
-KEY_TOKENIZER_MERGES     = "tokenizer.ggml.merges"
-KEY_TOKENIZER_BOS_ID     = "tokenizer.ggml.bos_token_id"
-KEY_TOKENIZER_EOS_ID     = "tokenizer.ggml.eos_token_id"
-KEY_TOKENIZER_UNK_ID     = "tokenizer.ggml.unknown_token_id"
-KEY_TOKENIZER_SEP_ID     = "tokenizer.ggml.seperator_token_id"
-KEY_TOKENIZER_PAD_ID     = "tokenizer.ggml.padding_token_id"
-KEY_TOKENIZER_HF_JSON    = "tokenizer.huggingface.json"
-KEY_TOKENIZER_RWKV       = "tokenizer.rwkv.world"
-
-
-#
-# recommended mapping of model tensor names for storage in gguf
-#
-
-
-class MODEL_ARCH(IntEnum):
-    LLAMA   = auto()
-    FALCON  = auto()
-    GPT2    = auto()
-    GPTJ    = auto()
-    GPTNEOX = auto()
-    MPT     = auto()
-
-
-class MODEL_TENSOR(IntEnum):
-    TOKEN_EMBD    = auto()
-    POS_EMBD      = auto()
-    OUTPUT        = auto()
-    OUTPUT_NORM   = auto()
-    ROPE_FREQS    = auto()
-    ATTN_Q        = auto()
-    ATTN_K        = auto()
-    ATTN_V        = auto()
-    ATTN_QKV      = auto()
-    ATTN_OUT      = auto()
-    ATTN_NORM     = auto()
-    ATTN_NORM_2   = auto()
-    ATTN_ROT_EMBD = auto()
-    FFN_GATE      = auto()
-    FFN_DOWN      = auto()
-    FFN_UP        = auto()
-    FFN_NORM      = auto()
-
-
-MODEL_ARCH_NAMES = {
-    MODEL_ARCH.LLAMA:   "llama",
-    MODEL_ARCH.FALCON:  "falcon",
-    MODEL_ARCH.GPT2:    "gpt2",
-    MODEL_ARCH.GPTJ:    "gptj",
-    MODEL_ARCH.GPTNEOX: "gptneox",
-    MODEL_ARCH.MPT:     "mpt",
-}
-
-MODEL_TENSOR_NAMES = {
-    MODEL_ARCH.LLAMA: {
-        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
-        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
-        MODEL_TENSOR.OUTPUT:        "output",
-        MODEL_TENSOR.ROPE_FREQS:    "rope_freqs",
-        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
-        MODEL_TENSOR.ATTN_Q:        "blk.{bid}.attn_q",
-        MODEL_TENSOR.ATTN_K:        "blk.{bid}.attn_k",
-        MODEL_TENSOR.ATTN_V:        "blk.{bid}.attn_v",
-        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
-        MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
-        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
-        MODEL_TENSOR.FFN_GATE:      "blk.{bid}.ffn_gate",
-        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
-        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
-    },
-    MODEL_ARCH.GPTNEOX: {
-        MODEL_TENSOR.TOKEN_EMBD:    "token_embd",
-        MODEL_TENSOR.OUTPUT_NORM:   "output_norm",
-        MODEL_TENSOR.OUTPUT:        "output",
-        MODEL_TENSOR.ATTN_NORM:     "blk.{bid}.attn_norm",
-        MODEL_TENSOR.ATTN_QKV:      "blk.{bid}.attn_qkv",
-        MODEL_TENSOR.ATTN_OUT:      "blk.{bid}.attn_output",
-        MODEL_TENSOR.FFN_NORM:      "blk.{bid}.ffn_norm",
-        MODEL_TENSOR.FFN_DOWN:      "blk.{bid}.ffn_down",
-        MODEL_TENSOR.FFN_UP:        "blk.{bid}.ffn_up",
-    },
-    MODEL_ARCH.FALCON: {
-        MODEL_TENSOR.TOKEN_EMBD:  "token_embd",
-        MODEL_TENSOR.OUTPUT_NORM: "output_norm",
-        MODEL_TENSOR.OUTPUT:      "output",
-        MODEL_TENSOR.ATTN_NORM:   "blk.{bid}.attn_norm",
-        MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2",
-        MODEL_TENSOR.ATTN_QKV:    "blk.{bid}.attn_qkv",
-        MODEL_TENSOR.ATTN_OUT:    "blk.{bid}.attn_output",
-        MODEL_TENSOR.FFN_DOWN:    "blk.{bid}.ffn_down",
-        MODEL_TENSOR.FFN_UP:      "blk.{bid}.ffn_up",
-    },
-    MODEL_ARCH.GPT2: {
-        # TODO
-    },
-    # TODO
-}
-
-# tensors that will not be serialized
-MODEL_TENSOR_SKIP = {
-    MODEL_ARCH.LLAMA: [
-        MODEL_TENSOR.ROPE_FREQS,
-        MODEL_TENSOR.ATTN_ROT_EMBD,
-    ],
-}
-
-
-# TODO: the following helper functions should be removed
-#       instead, get_tensor_name_map should return tuples of (name, MODEL_TENSOR)
-#       however, my Python is very bad, and I couldn't figure out how to do this, hence these functions
-# REMOVE
-def should_skip_tensor_TMP(arch: MODEL_ARCH, n_blocks: int, name: str) -> bool:
-    for skip in MODEL_TENSOR_SKIP.get(arch, []):
-        for i in range(n_blocks):
-            if name == MODEL_TENSOR_NAMES[arch][skip].format(bid=i):
-                return True
-
-    return False
-
-
-def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> dict:
-    tensor_map = {}
-
-    # Token embeddings
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.TOKEN_EMBD, None)
-
-    tensor_map["gpt_neox.embed_in"]           = mapped_to  # gptneox
-    tensor_map["transformer.wte"]             = mapped_to  # gpt2 mpt
-    tensor_map["transformer.word_embeddings"] = mapped_to  # falcon
-    tensor_map["model.embed_tokens"]          = mapped_to  # llama-hf
-    tensor_map["tok_embeddings"]              = mapped_to  # llama-pth
-
-    # Position embeddings
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.POS_EMBD, None)
-
-    tensor_map["transformer.wpe"] = mapped_to  # gpt2
-
-    # Output
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT, None)
-
-    tensor_map["embed_out"] = mapped_to  # gptneox
-    tensor_map["lm_head"]   = mapped_to  # gpt2 mpt falcon llama-hf
-    tensor_map["output"]    = mapped_to  # llama-pth
-
-    # Output norm
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT_NORM, None)
-
-    tensor_map["gpt_neox.final_layer_norm"] = mapped_to  # gptneox
-    tensor_map["transformer.ln_f"]          = mapped_to  # gpt2 falcon
-    tensor_map["transformer.norm_f"]        = mapped_to  # mpt
-    tensor_map["model.norm"]                = mapped_to  # llama-hf
-    tensor_map["norm"]                      = mapped_to  # llama-pth
-
-    # Rope frequencies
-    mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ROPE_FREQS, None)
-
-    tensor_map["rope.freqs"] = mapped_to  # llama-pth
-
-    # Attention and feed-forward blocks
-    for i in range(0, n_blocks):
-        # Attention norm
-        # TODO: is there are simpler way to write these 2 lines in Python?
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".input_layernorm"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".ln_1"]              = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".norm_1"]       = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".input_layernorm"]   = mapped_to  # falcon7b
-        tensor_map["transformer.h."+str(i)+".ln_mlp"]            = mapped_to  # falcon40b
-        tensor_map["model.layers."+str(i)+".input_layernorm"]    = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention_norm"]           = mapped_to  # llama-pth
-
-        # Attention norm 2
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM_2, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["transformer.h."+str(i)+".ln_attn"] = mapped_to  # falcon40b
-
-        # Attention query-key-value
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_QKV, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".attention.query_key_value"]    = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".attn.c_attn"]                    = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".attn.Wqkv"]                 = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".self_attention.query_key_value"] = mapped_to  # falcon
-
-        # Attention query
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_Q, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.q_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wq"]           = mapped_to  # llama-pth
-
-        # Attention key
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_K, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.k_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wk"]           = mapped_to  # llama-pth
-
-        # Attention value
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_V, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.v_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wv"]           = mapped_to  # llama-pth
-
-        # Attention output
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_OUT, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".attention.dense"]    = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".attn.c_proj"]          = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".attn.out_proj"]   = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".self_attention.dense"] = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".self_attn.o_proj"]      = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.wo"]                = mapped_to  # llama-pth
-
-        # Rotary embeddings
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_ROT_EMBD, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".self_attn.rotary_emb.inv_freq"]  = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".attention.inner_attention.rope.freqs"] = mapped_to  # llama-pth
-
-        # Feed-forward norm
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_NORM, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".post_attention_layernorm"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".ln_2"]                       = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".norm_2"]                = mapped_to  # mpt
-        tensor_map["model.layers."+str(i)+".post_attention_layernorm"]    = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".ffn_norm"]                          = mapped_to  # llama-pth
-
-        # Feed-forward up
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_UP, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_h_to_4h"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".mlp.c_fc"]            = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".ffn.up_proj"]    = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".mlp.dense_h_to_4h"]   = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".mlp.up_proj"]          = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w3"]            = mapped_to  # llama-pth
-
-        # Feed-forward gate
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_GATE, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["model.layers."+str(i)+".mlp.gate_proj"] = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w1"]     = mapped_to  # llama-pth
-
-        # Feed-forward down
-        mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_DOWN, None)
-        mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None
-
-        tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_4h_to_h"] = mapped_to  # gptneox
-        tensor_map["transformer.h."+str(i)+".mlp.c_proj"]          = mapped_to  # gpt2
-        tensor_map["transformer.blocks."+str(i)+".ffn.down_proj"]  = mapped_to  # mpt
-        tensor_map["transformer.h."+str(i)+".mlp.dense_4h_to_h"]   = mapped_to  # falcon
-        tensor_map["model.layers."+str(i)+".mlp.down_proj"]        = mapped_to  # llama-hf
-        tensor_map["layers."+str(i)+".feed_forward.w2"]            = mapped_to  # llama-pth
-
-    return tensor_map
-
-
-class TokenType(IntEnum):
-    NORMAL       = 1
-    UNKNOWN      = 2
-    CONTROL      = 3
-    USER_DEFINED = 4
-    UNUSED       = 5
-    BYTE         = 6
-
-#
-# implementation
-#
-
-
-class GGMLQuantizationType(IntEnum):
-    F32  = 0
-    F16  = 1
-    Q4_0 = 2
-    Q4_1 = 3
-    Q5_0 = 6
-    Q5_1 = 7
-    Q8_0 = 8
-    Q8_1 = 9
-    Q2_K = 10
-    Q3_K = 11
-    Q4_K = 12
-    Q5_K = 13
-    Q6_K = 14
-    Q8_K = 15
-
-
-class GGUFValueType(IntEnum):
-    UINT8   = 0
-    INT8    = 1
-    UINT16  = 2
-    INT16   = 3
-    UINT32  = 4
-    INT32   = 5
-    FLOAT32 = 6
-    BOOL    = 7
-    STRING  = 8
-    ARRAY   = 9
-
-    @staticmethod
-    def get_type(val):
-        if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray):
-            return GGUFValueType.STRING
-        elif isinstance(val, list):
-            return GGUFValueType.ARRAY
-        elif isinstance(val, float):
-            return GGUFValueType.FLOAT32
-        elif isinstance(val, bool):
-            return GGUFValueType.BOOL
-        elif isinstance(val, int):
-            return GGUFValueType.INT32
-        else:
-            print("Unknown type: "+str(type(val)))
-            sys.exit()
-
-
-class GGUFWriter:
-    def __init__(self, path: str, arch: str, use_temp_file = True):
-        self.fout = open(path, "wb")
-        self.arch = arch
-        self.offset_tensor = 0
-        self.data_alignment = GGUF_DEFAULT_ALIGNMENT
-        self.kv_data = b""
-        self.kv_data_count = 0
-        self.ti_data = b""
-        self.ti_data_count = 0
-        self.add_architecture()
-        self.use_temp_file = use_temp_file
-        self.tensors = []
-
-    def write_header_to_file(self):
-        self.fout.write(struct.pack("<I", GGUF_MAGIC))
-        self.fout.write(struct.pack("<I", GGUF_VERSION))
-        self.fout.write(struct.pack("<I", self.ti_data_count))
-        self.fout.write(struct.pack("<I", self.kv_data_count))
-        self.flush()
-#        print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
-
-    def write_kv_data_to_file(self):
-        self.fout.write(self.kv_data)
-        self.flush()
-
-    def write_ti_data_to_file(self):
-        self.fout.write(self.ti_data)
-        self.flush()
-
-    def add_key(self, key: str):
-        self.add_val(key, GGUFValueType.STRING, add_vtype=False)
-
-    def add_uint8(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.UINT8)
-
-    def add_int8(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.INT8)
-
-    def add_uint16(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.UINT16)
-
-    def add_int16(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.INT16)
-
-    def add_uint32(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.UINT32)
-
-    def add_int32(self, key: str, val: int):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.INT32)
-
-    def add_float32(self, key: str, val: float):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.FLOAT32)
-
-    def add_bool(self, key: str, val: bool):
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.BOOL)
-
-    def add_string(self, key: str, val: str):
-        if len(val) == 0:
-            return
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.STRING)
-
-    def add_array(self, key: str, val: list):
-        if not isinstance(val, list):
-            raise ValueError("Value must be a list for array type")
-
-        self.add_key(key)
-        self.add_val(val, GGUFValueType.ARRAY)
-
-    def add_val(self: str, val: Any, vtype: GGUFValueType = None, add_vtype: bool = True):
-        if vtype is None:
-            vtype = GGUFValueType.get_type(val)
-
-        if add_vtype:
-            self.kv_data += struct.pack("<I", vtype)
-            self.kv_data_count += 1
-
-        if vtype == GGUFValueType.UINT8:
-            self.kv_data += struct.pack("<B", val)
-        elif vtype == GGUFValueType.INT8:
-            self.kv_data += struct.pack("<b", val)
-        elif vtype == GGUFValueType.UINT16:
-            self.kv_data += struct.pack("<H", val)
-        elif vtype == GGUFValueType.INT16:
-            self.kv_data += struct.pack("<h", val)
-        elif vtype == GGUFValueType.UINT32:
-            self.kv_data += struct.pack("<I", val)
-        elif vtype == GGUFValueType.INT32:
-            self.kv_data += struct.pack("<i", val)
-        elif vtype == GGUFValueType.FLOAT32:
-            self.kv_data += struct.pack("<f", val)
-        elif vtype == GGUFValueType.BOOL:
-            self.kv_data += struct.pack("?", val)
-        elif vtype == GGUFValueType.STRING:
-            encoded_val = val.encode("utf8") if isinstance(val, str) else val
-            self.kv_data += struct.pack("<I", len(encoded_val))
-            self.kv_data += encoded_val
-        elif vtype == GGUFValueType.ARRAY:
-            ltype = set([GGUFValueType.get_type(item) for item in val])
-            assert len(ltype) == 1, "All items in a GGUF array should be of the same type"
-            self.kv_data += struct.pack("<I", list(ltype)[0])
-            self.kv_data += struct.pack("<I", len(val))
-            for item in val:
-                self.add_val(item, add_vtype=False)
-        else:
-            raise ValueError("Invalid GGUF metadata value type")
-
-    @staticmethod
-    def ggml_pad(x: int, n: int) -> int:
-        return ((x + n - 1) // n) * n
-
-    def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None):
-        assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
-
-        encoded_name = name.encode("utf8")
-        self.ti_data += struct.pack("<I", len(encoded_name))
-        self.ti_data += encoded_name
-        n_dims = len(tensor_shape)
-        self.ti_data += struct.pack("<I", n_dims)
-        for i in range(n_dims):
-            self.ti_data += struct.pack("<I", tensor_shape[n_dims - 1 - i])
-        if raw_dtype is None:
-            dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
-        else:
-            dtype = raw_dtype
-        self.ti_data += struct.pack("<I", dtype)
-        self.ti_data += struct.pack("<Q", self.offset_tensor)
-        self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
-        self.ti_data_count += 1
-
-    def add_tensor(self, name: str, tensor: np.ndarray, raw_shape: Optional[np.ndarray] = None, raw_dtype: Optional[GGMLQuantizationType] = None):
-        if self.use_temp_file and not hasattr(self, "temp_file"):
-            self.temp_file = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
-            self.temp_file.seek(0)
-
-        self.add_tensor_info(name, raw_shape if raw_shape is not None else tensor.shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype)
-
-        pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
-
-        if not self.use_temp_file:
-            self.tensors.append((tensor, pad))
-            return
-
-        tensor.tofile(self.temp_file)
-
-        if pad != 0:
-            self.temp_file.write(bytes([0] * pad))
-
-    def write_tensor_data(self, tensor: np.ndarray):
-        pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
-        if pad != 0:
-            self.fout.write(bytes([0] * pad))
-
-        tensor.tofile(self.fout)
-
-        pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes
-        if pad != 0:
-            self.fout.write(bytes([0] * pad))
-
-    def write_tensors_to_file(self):
-        self.write_ti_data_to_file()
-
-        pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell()
-        if pad != 0:
-            self.fout.write(bytes([0] * pad))
-
-        if not self.use_temp_file:
-            for (currtensor, currpad) in self.tensors:
-                currtensor.tofile(self.fout)
-                if currpad != 0:
-                    self.fout.write(bytes([0] * currpad))
-            return
-
-        self.temp_file.seek(0)
-
-        shutil.copyfileobj(self.temp_file, self.fout)
-        self.flush()
-        self.temp_file.close()
-
-    def flush(self):
-        self.fout.flush()
-
-    def close(self):
-        self.fout.close()
-
-    def add_architecture(self):
-        self.add_string(KEY_GENERAL_ARCHITECTURE, self.arch)
-
-    def add_author(self, author: str):
-        self.add_string(KEY_GENERAL_AUTHOR, author)
-
-    def add_tensor_data_layout(self, layout: str):
-        self.add_string(KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
-
-    def add_url(self, url: str):
-        self.add_string(KEY_GENERAL_URL, url)
-
-    def add_description(self, description: str):
-        self.add_string(KEY_GENERAL_DESCRIPTION, description)
-
-    def add_source_url(self, url: str):
-        self.add_string(KEY_GENERAL_SOURCE_URL, url)
-
-    def add_source_hf_repo(self, repo: str):
-        self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
-
-    def add_file_type(self, ftype: int):
-        self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
-
-    def add_name(self, name: str):
-        self.add_string(KEY_GENERAL_NAME, name)
-
-    def add_quantization_version(self, quantization_version: GGMLQuantizationType):
-        self.add_uint32(
-            KEY_GENERAL_QUANTIZATION_VERSION, quantization_version)
-
-    def add_custom_alignment(self, alignment: int):
-        self.data_alignment = alignment
-        self.add_uint32(KEY_GENERAL_ALIGNMENT, alignment)
-
-    def add_context_length(self, length: int):
-        self.add_uint32(
-            KEY_LLM_CONTEXT_LENGTH.format(arch=self.arch), length)
-
-    def add_embedding_length(self, length: int):
-        self.add_uint32(
-            KEY_LLM_EMBEDDING_LENGTH.format(arch=self.arch), length)
-
-    def add_block_count(self, length: int):
-        self.add_uint32(
-            KEY_LLM_BLOCK_COUNT.format(arch=self.arch), length)
-
-    def add_feed_forward_length(self, length: int):
-        self.add_uint32(
-            KEY_LLM_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
-
-    def add_parallel_residual(self, use: bool):
-        self.add_bool(
-            KEY_LLM_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
-
-    def add_tensor_data_layout(self, layout: str):
-        self.add_string(
-            KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
-
-    def add_head_count(self, count: int):
-        self.add_uint32(
-            KEY_ATTENTION_HEAD_COUNT.format(arch=self.arch), count)
-
-    def add_head_count_kv(self, count: int):
-        self.add_uint32(
-            KEY_ATTENTION_HEAD_COUNT_KV.format(arch=self.arch), count)
-
-    def add_max_alibi_bias(self, bias: float):
-        self.add_float32(
-            KEY_ATTENTION_MAX_ALIBI_BIAS.format(arch=self.arch), bias)
-
-    def add_clamp_kqv(self, value: float):
-        self.add_float32(
-            KEY_ATTENTION_CLAMP_KQV.format(arch=self.arch), value)
-
-    def add_layer_norm_eps(self, value: float):
-        self.add_float32(
-            KEY_ATTENTION_LAYERNORM_EPS.format(arch=self.arch), value)
-
-    def add_layer_norm_rms_eps(self, value: float):
-        self.add_float32(
-            KEY_ATTENTION_LAYERNORM_RMS_EPS.format(arch=self.arch), value)
-
-    def add_rope_dimension_count(self, count: int):
-        self.add_uint32(
-            KEY_ROPE_DIMENSION_COUNT.format(arch=self.arch), count)
-
-    def add_rope_scale_linear(self, value:  float):
-        self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value)
-
-    def add_tokenizer_model(self, model: str):
-        self.add_string(KEY_TOKENIZER_MODEL, model)
-
-    def add_token_list(self, tokens: List):
-        self.add_array(KEY_TOKENIZER_LIST, tokens)
-
-    def add_token_merges(self, merges: List):
-        self.add_array(KEY_TOKENIZER_MERGES, merges)
-
-    def add_token_types(self, types: List[int]):
-        self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types)
-
-    def add_token_scores(self, scores: List[float]):
-        self.add_array(KEY_TOKENIZER_SCORES, scores)
-
-    def add_bos_token_id(self, id: int):
-        self.add_uint32(KEY_TOKENIZER_BOS_ID, id)
-
-    def add_eos_token_id(self, id: int):
-        self.add_uint32(KEY_TOKENIZER_EOS_ID, id)
-
-    def add_unk_token_id(self, id: int):
-        self.add_uint32(KEY_TOKENIZER_UNK_ID, id)
-
-    def add_sep_token_id(self, id: int):
-        self.add_uint32(KEY_TOKENIZER_SEP_ID, id)
-
-    def add_pad_token_id(self, id: int):
-        self.add_uint32(KEY_TOKENIZER_PAD_ID, id)
-
-
-# Example usage:
-if __name__ == "__main__":
-    # Example usage with a file
-    gguf_writer = GGUFWriter("example.gguf", "llama")
-
-    gguf_writer.add_architecture()
-    gguf_writer.add_block_count(12)
-    gguf_writer.add_uint32("answer", 42)  # Write a 32-bit integer
-    gguf_writer.add_float32("answer_in_float", 42.0)  # Write a 32-bit float
-    gguf_writer.add_custom_alignment(64)
-
-    tensor1 = np.ones((32,), dtype=np.float32) * 100.0
-    tensor2 = np.ones((64,), dtype=np.float32) * 101.0
-    tensor3 = np.ones((96,), dtype=np.float32) * 102.0
-
-    gguf_writer.add_tensor("tensor1", tensor1)
-    gguf_writer.add_tensor("tensor2", tensor2)
-    gguf_writer.add_tensor("tensor3", tensor3)
-
-    gguf_writer.write_header_to_file()
-    gguf_writer.write_kv_data_to_file()
-    gguf_writer.write_tensors_to_file()
-
-    gguf_writer.close()
diff --git a/grammars/c.gbnf b/grammars/c.gbnf
new file mode 100644
index 000000000..4a0331dd2
--- /dev/null
+++ b/grammars/c.gbnf
@@ -0,0 +1,42 @@
+root ::= (declaration)*
+
+declaration ::= dataType identifier "(" parameter? ")" "{" statement* "}"
+
+dataType  ::= "int" ws | "float" ws | "char" ws
+identifier ::= [a-zA-Z_] [a-zA-Z_0-9]*
+
+parameter ::= dataType identifier
+
+statement ::=
+    ( dataType identifier ws "=" ws expression ";" ) |
+    ( identifier ws "=" ws expression ";" ) |
+    ( identifier ws "(" argList? ")" ";" ) |
+    ( "return" ws expression ";" ) |
+    ( "while" "(" condition ")" "{" statement* "}" ) |
+    ( "for" "(" forInit ";" ws condition ";" ws forUpdate ")" "{" statement* "}" ) |
+    ( "if" "(" condition ")" "{" statement* "}" ("else" "{" statement* "}")? ) |
+    ( singleLineComment ) |
+    ( multiLineComment )
+
+forInit ::= dataType identifier ws "=" ws expression | identifier ws "=" ws expression
+forUpdate ::= identifier ws "=" ws expression
+
+condition ::= expression relationOperator expression
+relationOperator ::= ("<=" | "<" | "==" | "!=" | ">=" | ">")
+
+expression ::= term (("+" | "-") term)*
+term ::= factor(("*" | "/") factor)*
+
+factor ::= identifier | number | unaryTerm | funcCall | parenExpression
+unaryTerm ::= "-" factor
+funcCall ::= identifier "(" argList? ")"
+parenExpression ::= "(" ws expression ws ")"
+
+argList ::= expression ("," ws expression)*
+
+number ::= [0-9]+
+
+singleLineComment ::= "//" [^\n]* "\n"
+multiLineComment ::= "/*" ( [^*] | ("*" [^/]) )* "*/"
+
+ws ::= ([ \t\n]+)
diff --git a/grammars/json_arr.gbnf b/grammars/json_arr.gbnf
new file mode 100644
index 000000000..ef53e77a0
--- /dev/null
+++ b/grammars/json_arr.gbnf
@@ -0,0 +1,34 @@
+# This is the same as json.gbnf but we restrict whitespaces at the end of the root array
+# Useful for generating JSON arrays
+
+root   ::= arr
+value  ::= object | array | string | number | ("true" | "false" | "null") ws
+
+arr  ::=
+  "[\n" ws (
+            value
+    (",\n" ws value)*
+  )? "]"
+
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}" ws
+
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]" ws
+
+string ::=
+  "\"" (
+    [^"\\] |
+    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
+  )* "\"" ws
+
+number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
+
+# Optional space: by convention, applied in this grammar after literal chars when allowed
+ws ::= ([ \t\n] ws)?
diff --git a/k_quants.c b/k_quants.c
index 82bf81697..558f5fda8 100644
--- a/k_quants.c
+++ b/k_quants.c
@@ -13,6 +13,26 @@
 //
 #include <arm_neon.h>
 
+#if !defined(__aarch64__)
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+#endif
+
 #else
 
 #ifdef __wasm_simd128__
@@ -34,6 +54,10 @@
 #endif
 #endif
 
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
 #undef MIN
 #undef MAX
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -45,7 +69,6 @@
 // 2-6 bit quantization in super-blocks
 //
 
-
 //
 // ===================== Helper functions
 //
@@ -63,7 +86,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
         float ax = fabsf(x[i]);
         if (ax > amax) { amax = ax; max = x[i]; }
     }
-    if (!amax) { // all zero
+    if (amax < 1e-30f) { // all zero
         for (int i = 0; i < n; ++i) {
             L[i] = 0;
         }
@@ -183,13 +206,9 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
         int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
-    float sum_x = 0;
-    float sum_x2 = 0;
     for (int i = 1; i < n; ++i) {
         if (x[i] < min) min = x[i];
         if (x[i] > max) max = x[i];
-        sum_x += x[i];
-        sum_x2 += x[i]*x[i];
     }
     if (max == min) {
         for (int i = 0; i < n; ++i) L[i] = 0;
@@ -328,7 +347,6 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
     const float q4scale = 15.f;
 
     for (int i = 0; i < nb; i++) {
-
         float max_scale = 0; // as we are deducting the min, scales are always positive
         float max_min = 0;
         for (int j = 0; j < QK_K/16; ++j) {
@@ -1070,6 +1088,13 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict
 
         }
 
+        if (!max_abs_scale) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = ggml_fp32_to_fp16(0.f);
+            x += QK_K;
+            continue;
+        }
+
         float iscale = -128.f/max_scale;
         y[i].d = ggml_fp32_to_fp16(1/iscale);
         for (int ib = 0; ib < QK_K/16; ++ib) {
@@ -1306,7 +1331,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
     const uint8x16_t m4 = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x2_t q2bytes;
     uint8_t aux[16];
@@ -1557,6 +1584,90 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        size_t vl = 16;
+
+        vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
+        vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
+
+        vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
+
+        vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
+        vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
+        vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
+        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
+        vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+
+        sumf  += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
+
+        vl = 32;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
+
+        uint8_t is=0;
+        int isum=0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load Q2
+            vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
+
+            vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
+            vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
+
+            // duplicate scale elements for product
+            vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
+            vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
+            vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
+            vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
+
+            vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
+            vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
+            vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
+            vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
+
+            // load Q8
+            vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
+            vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
+
+            vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
+            vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
+            vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
+            vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
+
+            isum += __riscv_vmv_x_s_i32m1_i32(isum1);
+
+            q2+=32;  q8+=128;  is=8;
+
+        }
+
+        sumf += dall * isum;
+
+    }
+
+    *s = sumf;
+
 #else
 
     float sumf = 0;
@@ -1612,7 +1723,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m3 = vdupq_n_u8(0x3);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q2bytes;
 
@@ -1780,6 +1893,64 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __riscv_v_intrinsic
+
+    uint32_t aux32[2];
+    const uint8_t * scales = (const uint8_t *)aux32;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * (float)x[i].d;
+        const float dmin = -y[i].d * (float)x[i].dmin;
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+
+        aux32[0] = sc[0] & 0x0f0f0f0f;
+        aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        sumf += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
+
+        int isum1 = 0;
+        int isum2 = 0;
+
+        size_t vl = 16;
+
+        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+        // load Q2
+        vuint8mf2_t q2_x = __riscv_vle8_v_u8mf2(q2, vl);
+
+        vint8mf2_t q2_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q2_x, 0x03, vl));
+        vint8mf2_t q2_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x2, vl), 0x03 , vl));
+        vint8mf2_t q2_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x4, vl), 0x03 , vl));
+        vint8mf2_t q2_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x6, vl), 0x03 , vl));
+
+        // load Q8, and take product with Q2
+        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q2_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
+        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q2_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
+        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q2_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
+        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q2_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
+
+        vint16m1_t vs_0 = __riscv_vredsum_vs_i16m1_i16m1(p0, vzero, vl);
+        vint16m1_t vs_1 = __riscv_vredsum_vs_i16m1_i16m1(p1, vzero, vl);
+        vint16m1_t vs_2 = __riscv_vredsum_vs_i16m1_i16m1(p2, vzero, vl);
+        vint16m1_t vs_3 = __riscv_vredsum_vs_i16m1_i16m1(p3, vzero, vl);
+
+        isum1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[0];
+        isum2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[1];
+        isum1 += __riscv_vmv_x_s_i16m1_i16(vs_2) * scales[2];
+        isum2 += __riscv_vmv_x_s_i16m1_i16(vs_3) * scales[3];
+
+        sumf += d * (isum1 + isum2);
+
+    }
+
+    *s = sumf;
+
 #else
 
     float sumf = 0;
@@ -2060,7 +2231,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     __m256 acc = _mm256_setzero_ps();
 
-    uint32_t *aux;
+    const uint32_t *aux;
 
     for (int i = 0; i < nb; ++i) {
 
@@ -2070,7 +2241,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
         const int8_t  * restrict q8 = y[i].qs;
 
         // Set up scales
-        aux = (uint32_t *)x[i].scales;
+        aux = (const uint32_t *)x[i].scales;
         __m128i scales128 = _mm_set_epi32(
                 ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
                 ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
@@ -2193,6 +2364,106 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+
+        size_t vl = 32;
+        uint8_t m =  1;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
+
+        int sum_t = 0;
+
+        for (int j = 0; j < QK_K; j += 128) {
+
+            vl = 32;
+
+            // load Q3
+            vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
+
+            vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
+            vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
+            vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
+            vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
+
+            // compute mask for subtraction
+            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
+            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_m(vmask_0, q3_0, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_m(vmask_1, q3_1, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_m(vmask_2, q3_2, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
+            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_m(vmask_3, q3_3, 0x4, vl);
+            m <<= 1;
+
+            // load Q8 and take product with Q3
+            vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            // retreive lane to multiply with scale
+            vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
+            vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
+            vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
+            vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
+            vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
+            vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
+            vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
+            vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
+
+            sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q3 += 32;    q8 += 128;   scale += 8;
+
+        }
+
+        const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
+
+        sumf += d*sum_t;
+
+    }
+
+    *s = sumf;
+
 #else
     // scalar version
     // This function is written like this so the compiler can manage to vectorize most of it
@@ -2496,6 +2767,79 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    uint16_t aux16[2];
+    int8_t * scales = (int8_t *)aux16;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        for (int j = 0; j < 4; ++j) scales[j] -= 8;
+
+        int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
+
+        const float d = y[i].d * (float)x[i].d;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        // load qh
+        vuint8mf4_t qh_x1   = __riscv_vle8_v_u8mf4(x[i].hmask, 8);
+        vuint8mf2_t qh_x2   = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
+
+        size_t vl = 16;
+
+        // extend and combine both qh_x1 and qh_x2
+        vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
+
+        vuint8mf2_t qh_0 = __riscv_vand_vx_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
+        vuint8mf2_t qh_1 = __riscv_vand_vx_u8mf2(qh_x, 0x4, vl);
+        vuint8mf2_t qh_2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
+        vuint8mf2_t qh_3 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), 0x4, vl);
+
+        // load Q3
+        vuint8mf2_t q3_x  = __riscv_vle8_v_u8mf2(q3, vl);
+
+        vuint8mf2_t q3h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q3_x, 0x3, vl), qh_0, vl);
+        vuint8mf2_t q3h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 2, vl), 0x3, vl), qh_1, vl);
+        vuint8mf2_t q3h_2 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 4, vl), 0x3, vl), qh_2, vl);
+        vuint8mf2_t q3h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 0x6, vl), qh_3, vl);
+
+        vint8mf2_t q3_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_0);
+        vint8mf2_t q3_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_1);
+        vint8mf2_t q3_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_2);
+        vint8mf2_t q3_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_3);
+
+        // load Q8 and take product with Q3
+        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q3_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
+        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q3_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
+        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q3_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
+        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q3_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
+
+        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
+        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
+        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
+        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
+
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scales[0];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scales[2];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scales[1];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scales[3];
+
+        sumf += d * isum;
+
+    }
+
+    *s = sumf;
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2582,7 +2926,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
         memcpy(utmp, x[i].scales, 12);
 
-        const uint32x2_t mins8 = {utmp[1] & kmask1, ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4)};
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
         utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
         utmp[0] &= kmask1;
 
@@ -2596,8 +2943,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        //int32x4_t isum = mzero;
-
         int32_t sumi1 = 0;
         int32_t sumi2 = 0;
 
@@ -2694,13 +3039,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
             const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
             p16l = _mm256_madd_epi16(scale_l, p16l);
-            sumi = _mm256_add_epi32(sumi, p16l);
 
             const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
             p16h = _mm256_madd_epi16(scale_h, p16h);
-            sumi = _mm256_add_epi32(sumi, p16h);
+            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
 
+            sumi = _mm256_add_epi32(sumi, sumj);
         }
 
         __m256 vd = _mm256_set1_ps(d);
@@ -2795,6 +3140,78 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
 
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        size_t vl = 8;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        vl = 32;
+
+        int32_t sum_1 = 0;
+        int32_t sum_2 = 0;
+
+        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q4
+            vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
+
+            // load Q8 and multiply it with lower Q4 nibble
+            vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
+            vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
+            vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
+
+            sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
+
+            // load Q8 and multiply it with upper Q4 nibble
+            vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
+            vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
+            vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
+
+            sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
+
+            q4 += 32;    q8 += 64;
+
+        }
+
+        sumf += d*(sum_1 + sum_2);
+
+    }
+
+    *s = sumf;
+
 #else
 
 
@@ -3036,6 +3453,50 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) - summs;
 
+#elif defined __riscv_v_intrinsic
+
+    uint16_t s16[2];
+    const uint8_t * restrict scales = (const uint8_t *)s16;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const uint16_t * restrict b = (const uint16_t *)x[i].scales;
+        s16[0] = b[0] & 0x0f0f;
+        s16[1] = (b[0] >> 4) & 0x0f0f;
+
+        sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]);
+
+        size_t vl = 32;
+
+        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+        // load Q4
+        vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
+
+        // load Q8 and multiply it with lower Q4 nibble
+        vint8m1_t  q4_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
+        vint16m2_t va_0 = __riscv_vwmul_vv_i16m2(q4_a, __riscv_vle8_v_i8m1(q8, vl), vl);
+        vint16m1_t aux1 = __riscv_vredsum_vs_i16m2_i16m1(va_0, vzero, vl);
+
+        sumf += d*scales[0]*__riscv_vmv_x_s_i16m1_i16(aux1);
+
+        // load Q8 and multiply it with upper Q4 nibble
+        vint8m1_t  q4_s = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
+        vint16m2_t va_1 = __riscv_vwmul_vv_i16m2(q4_s, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+        vint16m1_t aux2 = __riscv_vredsum_vs_i16m2_i16m1(va_1, vzero, vl);
+
+        sumf += d*scales[1]*__riscv_vmv_x_s_i16m1_i16(aux2);
+
+    }
+
+    *s = sumf;
+
 #else
 
     uint8_t aux8[QK_K];
@@ -3096,9 +3557,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mone = vdupq_n_u8(1);
     const uint8x16_t mtwo = vdupq_n_u8(2);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
 
@@ -3364,6 +3827,93 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    float sums = 0.0;
+
+    size_t vl;
+
+    for (int i = 0; i < nb; ++i) {
+
+        vl = 8;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
+        const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d;
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        vl = 32;
+        int32_t aux32 = 0;
+        int is = 0;
+
+        uint8_t m = 1;
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q5 and Q8
+            vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
+            vint8m1_t  q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
+
+            // compute mask for addition
+            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_m(vmask_1, q5_a, 16, vl);
+            m <<= 1;
+
+            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_m(vmask_2, q5_l, 16, vl);
+            m <<= 1;
+
+            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
+            vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
+
+            vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
+            vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
+
+            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
+            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
+
+            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
+            q5 += 32;    q8 += 64;
+
+        }
+
+        vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
+        sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
+
+    }
+
+    *s = sumf+sums;
+
 #else
 
     const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -3441,8 +3991,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 #ifdef __ARM_NEON
 
     const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
     const uint8x16_t mh = vdupq_n_u8(16);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t mzero = vdupq_n_s32(0);
+#endif
 
     int8x16x4_t q5bytes;
     uint8x16x4_t q5h;
@@ -3607,6 +4159,76 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * (float)x[i].d;
+        const int8_t * sc = x[i].scales;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        // load qh
+        vuint8mf4_t qh_x1   = __riscv_vle8_v_u8mf4(qh, 8);
+        vuint8mf2_t qh_x2   = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
+
+        size_t vl = 16;
+
+        // combine both qh_1 and qh_2
+        vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
+
+        vuint8mf2_t qh_h0 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
+        vuint8mf2_t qh_h1 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), vl), 16, vl);
+        vuint8mf2_t qh_h2 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(qh_x, vl), 16, vl);
+        vuint8mf2_t qh_h3 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
+
+        vint8mf2_t qh_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h0);
+        vint8mf2_t qh_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h1);
+        vint8mf2_t qh_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h2);
+        vint8mf2_t qh_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h3);
+
+        // load q5
+        vuint8mf2_t q5_x1  = __riscv_vle8_v_u8mf2(q5, vl);
+        vuint8mf2_t q5_x2  = __riscv_vle8_v_u8mf2(q5+16, vl);
+
+        vint8mf2_t q5s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x1, 0xF, vl));
+        vint8mf2_t q5s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x2, 0xF, vl));
+        vint8mf2_t q5s_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x1, 0x4, vl));
+        vint8mf2_t q5s_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x2, 0x4, vl));
+
+        vint8mf2_t q5_0 = __riscv_vsub_vv_i8mf2(q5s_0, qh_0, vl);
+        vint8mf2_t q5_1 = __riscv_vsub_vv_i8mf2(q5s_1, qh_1, vl);
+        vint8mf2_t q5_2 = __riscv_vsub_vv_i8mf2(q5s_2, qh_2, vl);
+        vint8mf2_t q5_3 = __riscv_vsub_vv_i8mf2(q5s_3, qh_3, vl);
+
+        // load Q8 and multiply it with Q5
+        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q5_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
+        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q5_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
+        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q5_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
+        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q5_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
+
+        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
+        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
+        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
+        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
+
+        int32_t sumi1 = sc[0] * __riscv_vmv_x_s_i32m1_i32(vs_0);
+        int32_t sumi2 = sc[1] * __riscv_vmv_x_s_i32m1_i32(vs_1);
+        int32_t sumi3 = sc[2] * __riscv_vmv_x_s_i32m1_i32(vs_2);
+        int32_t sumi4 = sc[3] * __riscv_vmv_x_s_i32m1_i32(vs_3);
+
+        sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
+
+    }
+
+    *s = sumf;
+
 #else
 
     int8_t aux8[QK_K];
@@ -3660,7 +4282,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
+#if defined(__ARM_FEATURE_DOTPROD)
     const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
     //const int8x16_t  m32s = vdupq_n_s8(32);
 
     const uint8x16_t mone = vdupq_n_u8(3);
@@ -3989,6 +4613,91 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        size_t vl;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        int sum_t = 0;
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            vl = 32;
+
+            // load qh
+            vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
+
+            // load Q6
+            vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
+            vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
+
+            vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
+            vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
+            vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
+            vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
+
+            vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
+            vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
+
+            vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
+            vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
+            vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
+            vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
+
+            vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
+            vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
+            vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
+            vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
+
+            // load Q8 and take product
+            vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
+            vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
+            vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
+            vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
+            vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
+            vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
+            vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
+            vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
+
+            sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q6 += 64;   qh += 32;   q8 += 128;   is=8;
+
+        }
+
+        sumf += d * sum_t;
+
+    }
+
+    *s = sumf;
+
 #else
 
     int8_t  aux8[QK_K];
@@ -4049,8 +4758,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
     float sum = 0;
 
     const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
     const int8x16_t  m32s = vdupq_n_s8(32);
+#if defined(__ARM_FEATURE_DOTPROD)
+    const int32x4_t  vzero = vdupq_n_s32(0);
+#endif
 
     const uint8x16_t mone = vdupq_n_u8(3);
 
@@ -4240,6 +4951,73 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d_all = (float)x[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        int32_t isum = 0;
+
+        size_t vl = 16;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        // load Q6
+        vuint8mf2_t q6_0 = __riscv_vle8_v_u8mf2(q6, vl);
+        vuint8mf2_t q6_1 = __riscv_vle8_v_u8mf2(q6+16, vl);
+
+        // load qh
+        vuint8mf2_t qh_x = __riscv_vle8_v_u8mf2(qh, vl);
+
+        vuint8mf2_t qh0 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
+        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
+        vuint8mf2_t qh1 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
+        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
+        vuint8mf2_t qh2 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
+        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
+        vuint8mf2_t qh3 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
+
+        vuint8mf2_t q6h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_0, 0xF, vl), qh0, vl);
+        vuint8mf2_t q6h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_1, 0xF, vl), qh1, vl);
+        vuint8mf2_t q6h_2 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_0, 0x4, vl), qh2, vl);
+        vuint8mf2_t q6h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_1, 0x4, vl), qh3, vl);
+
+        vint8mf2_t q6v_0 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_0), 32, vl);
+        vint8mf2_t q6v_1 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_1), 32, vl);
+        vint8mf2_t q6v_2 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_2), 32, vl);
+        vint8mf2_t q6v_3 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_3), 32, vl);
+
+        // load Q8 and take product
+        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q6v_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
+        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q6v_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
+        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q6v_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
+        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q6v_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
+
+        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
+        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
+        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
+        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
+
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scale[0];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scale[1];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scale[2];
+        isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scale[3];
+
+        sumf += isum * d_all * y[i].d;
+
+    }
+
+    *s = sumf;
+
 #else
 
     int8_t  aux8[QK_K];
diff --git a/k_quants.h b/k_quants.h
index adc6a3913..9de089e7a 100644
--- a/k_quants.h
+++ b/k_quants.h
@@ -29,7 +29,7 @@
 
 // 2-bit quantization
 // weight is represented as x = a * q + b
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 2.5625 bits per weight
 typedef struct {
     uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
@@ -41,7 +41,7 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "w
 
 // 3-bit quantization
 // weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
 #ifdef GGML_QKK_64
 typedef struct {
@@ -62,7 +62,7 @@ static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 +
 #endif
 
 // 4-bit quantization
-// 16 blocks of 32 elements each
+// 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
 #ifdef GGML_QKK_64
@@ -83,7 +83,7 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/
 #endif
 
 // 5-bit quantization
-// 16 blocks of 32 elements each
+// 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
 #ifdef GGML_QKK_64
@@ -107,7 +107,7 @@ static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/
 
 // 6-bit quantization
 // weight is represented as x = a * q
-// 16 blocks of 16 elemenets each
+// 16 blocks of 16 elements each
 // Effectively 6.5625 bits per weight
 typedef struct {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits
diff --git a/llama.cpp b/llama.cpp
index fd8eaa180..a4312ab72 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1,13 +1,8 @@
-// Defines fileno on msys:
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#include <cstddef>
-#include <cstdint>
-#include <cstdio>
-#endif
-
+#define LLAMA_API_INTERNAL
 #include "llama.h"
 
+#include "unicode.h"
+
 #include "ggml.h"
 
 #include "ggml-alloc.h"
@@ -62,6 +57,9 @@
 #include <cinttypes>
 #include <climits>
 #include <cstdarg>
+#include <cstddef>
+#include <cstdint>
+#include <cstdio>
 #include <cstring>
 #include <ctime>
 #include <fstream>
@@ -72,28 +70,16 @@
 #include <numeric>
 #include <queue>
 #include <random>
+#include <regex>
 #include <sstream>
 #include <thread>
 #include <unordered_map>
+#include <set>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-// tensor names
-#define TN_TOKEN_EMBD  "token_embd.weight"
-#define TN_OUTPUT_NORM "output_norm.weight"
-#define TN_OUTPUT      "output.weight"
-#define TN_ATTN_NORM   "blk.%d.attn_norm.weight"
-#define TN_ATTN_Q      "blk.%d.attn_q.weight"
-#define TN_ATTN_K      "blk.%d.attn_k.weight"
-#define TN_ATTN_V      "blk.%d.attn_v.weight"
-#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
-#define TN_FFN_NORM    "blk.%d.ffn_norm.weight"
-#define TN_FFN_GATE    "blk.%d.ffn_gate.weight"
-#define TN_FFN_DOWN    "blk.%d.ffn_down.weight"
-#define TN_FFN_UP      "blk.%d.ffn_up.weight"
-
 #ifdef __GNUC__
 #ifdef __MINGW32__
 #define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
@@ -107,18 +93,63 @@
 //
 // logging
 //
-LLAMA_ATTRIBUTE_FORMAT(2, 3)
-static void llama_log_internal        (llama_log_level level, const char* format, ...);
-static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data);
 
-#define LLAMA_LOG_INFO(...)  llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
-#define LLAMA_LOG_WARN(...)  llama_log_internal(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__)
-#define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
+LLAMA_ATTRIBUTE_FORMAT(2, 3)
+static void llama_log_internal        (ggml_log_level level, const char* format, ...);
+static void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
+
+#define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
 
 //
 // helpers
 //
 
+static size_t utf8_len(char src) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
+    return lookup[highbits];
+}
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    std::string result;
+    for (size_t pos = 0; ; pos += search.length()) {
+        auto new_pos = s.find(search, pos);
+        if (new_pos == std::string::npos) {
+            result += s.substr(pos, s.size() - pos);
+            break;
+        }
+        result += s.substr(pos, new_pos - pos) + replace;
+        pos = new_pos;
+    }
+    s = std::move(result);
+}
+
+static bool is_float_close(float a, float b, float abs_tol) {
+    // Check for non-negative tolerance
+    if (abs_tol < 0.0) {
+        throw std::invalid_argument("Tolerance must be non-negative");
+    }
+
+    // Exact equality check
+    if (a == b) {
+        return true;
+    }
+
+    // Check for infinities
+    if (std::isinf(a) || std::isinf(b)) {
+        return false;
+    }
+
+    // Regular comparison using the provided absolute tolerance
+    return std::fabs(b - a) <= abs_tol;
+}
+
+#ifdef GGML_USE_CPU_HBM
+#include <hbwmalloc.h>
+#endif
+
 static void zeros(std::ofstream & file, size_t n) {
     char zero = 0;
     for (size_t i = 0; i < n; ++i) {
@@ -142,6 +173,359 @@ static std::string format(const char * fmt, ...) {
     return std::string(buf.data(), size);
 }
 
+//
+// gguf constants (sync with gguf.py)
+//
+
+enum llm_arch {
+    LLM_ARCH_LLAMA,
+    LLM_ARCH_FALCON,
+    LLM_ARCH_BAICHUAN,
+    LLM_ARCH_GPT2,
+    LLM_ARCH_GPTJ,
+    LLM_ARCH_GPTNEOX,
+    LLM_ARCH_MPT,
+    LLM_ARCH_STARCODER,
+    LLM_ARCH_PERSIMMON,
+    LLM_ARCH_REFACT,
+    LLM_ARCH_UNKNOWN,
+};
+
+static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
+    { LLM_ARCH_LLAMA,           "llama"     },
+    { LLM_ARCH_FALCON,          "falcon"    },
+    { LLM_ARCH_GPT2,            "gpt2"      },
+    { LLM_ARCH_GPTJ,            "gptj"      },
+    { LLM_ARCH_GPTNEOX,         "gptneox"   },
+    { LLM_ARCH_MPT,             "mpt"       },
+    { LLM_ARCH_BAICHUAN,        "baichuan"  },
+    { LLM_ARCH_STARCODER,       "starcoder" },
+    { LLM_ARCH_PERSIMMON,       "persimmon" },
+    { LLM_ARCH_REFACT,          "refact" },
+};
+
+enum llm_kv {
+    LLM_KV_GENERAL_ARCHITECTURE,
+    LLM_KV_GENERAL_QUANTIZATION_VERSION,
+    LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_NAME,
+    LLM_KV_GENERAL_AUTHOR,
+    LLM_KV_GENERAL_URL,
+    LLM_KV_GENERAL_DESCRIPTION,
+    LLM_KV_GENERAL_LICENSE,
+    LLM_KV_GENERAL_SOURCE_URL,
+    LLM_KV_GENERAL_SOURCE_HF_REPO,
+
+    LLM_KV_CONTEXT_LENGTH,
+    LLM_KV_EMBEDDING_LENGTH,
+    LLM_KV_BLOCK_COUNT,
+    LLM_KV_FEED_FORWARD_LENGTH,
+    LLM_KV_USE_PARALLEL_RESIDUAL,
+    LLM_KV_TENSOR_DATA_LAYOUT,
+
+    LLM_KV_ATTENTION_HEAD_COUNT,
+    LLM_KV_ATTENTION_HEAD_COUNT_KV,
+    LLM_KV_ATTENTION_MAX_ALIBI_BIAS,
+    LLM_KV_ATTENTION_CLAMP_KQV,
+    LLM_KV_ATTENTION_LAYERNORM_EPS,
+    LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
+
+    LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_SCALE_LINEAR,
+
+    LLM_KV_TOKENIZER_MODEL,
+    LLM_KV_TOKENIZER_LIST,
+    LLM_KV_TOKENIZER_TOKEN_TYPE,
+    LLM_KV_TOKENIZER_SCORES,
+    LLM_KV_TOKENIZER_MERGES,
+    LLM_KV_TOKENIZER_BOS_ID,
+    LLM_KV_TOKENIZER_EOS_ID,
+    LLM_KV_TOKENIZER_UNK_ID,
+    LLM_KV_TOKENIZER_SEP_ID,
+    LLM_KV_TOKENIZER_PAD_ID,
+    LLM_KV_TOKENIZER_HF_JSON,
+    LLM_KV_TOKENIZER_RWKV,
+};
+
+static std::map<llm_kv, std::string> LLM_KV_NAMES = {
+    { LLM_KV_GENERAL_ARCHITECTURE,          "general.architecture"                  },
+    { LLM_KV_GENERAL_QUANTIZATION_VERSION,  "general.quantization_version"          },
+    { LLM_KV_GENERAL_ALIGNMENT,             "general.alignment"                     },
+    { LLM_KV_GENERAL_NAME,                  "general.name"                          },
+    { LLM_KV_GENERAL_AUTHOR,                "general.author"                        },
+    { LLM_KV_GENERAL_URL,                   "general.url"                           },
+    { LLM_KV_GENERAL_DESCRIPTION,           "general.description"                   },
+    { LLM_KV_GENERAL_LICENSE,               "general.license"                       },
+    { LLM_KV_GENERAL_SOURCE_URL,            "general.source.url"                    },
+    { LLM_KV_GENERAL_SOURCE_HF_REPO,        "general.source.huggingface.repository" },
+
+    { LLM_KV_CONTEXT_LENGTH,                "%s.context_length"        },
+    { LLM_KV_EMBEDDING_LENGTH,              "%s.embedding_length"      },
+    { LLM_KV_BLOCK_COUNT,                   "%s.block_count"           },
+    { LLM_KV_FEED_FORWARD_LENGTH,           "%s.feed_forward_length"   },
+    { LLM_KV_USE_PARALLEL_RESIDUAL,         "%s.use_parallel_residual" },
+    { LLM_KV_TENSOR_DATA_LAYOUT,            "%s.tensor_data_layout"    },
+
+    { LLM_KV_ATTENTION_HEAD_COUNT,          "%s.attention.head_count"             },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,       "%s.attention.head_count_kv"          },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,      "%s.attention.max_alibi_bias"         },
+    { LLM_KV_ATTENTION_CLAMP_KQV,           "%s.attention.clamp_kqv"              },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,       "%s.attention.layer_norm_epsilon"     },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,   "%s.attention.layer_norm_rms_epsilon" },
+
+    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count" },
+    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"       },
+    { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"    },
+
+    { LLM_KV_TOKENIZER_MODEL,               "tokenizer.ggml.model"              },
+    { LLM_KV_TOKENIZER_LIST,                "tokenizer.ggml.tokens"             },
+    { LLM_KV_TOKENIZER_TOKEN_TYPE,          "tokenizer.ggml.token_type"         },
+    { LLM_KV_TOKENIZER_SCORES,              "tokenizer.ggml.scores"             },
+    { LLM_KV_TOKENIZER_MERGES,              "tokenizer.ggml.merges"             },
+    { LLM_KV_TOKENIZER_BOS_ID,              "tokenizer.ggml.bos_token_id"       },
+    { LLM_KV_TOKENIZER_EOS_ID,              "tokenizer.ggml.eos_token_id"       },
+    { LLM_KV_TOKENIZER_UNK_ID,              "tokenizer.ggml.unknown_token_id"   },
+    { LLM_KV_TOKENIZER_SEP_ID,              "tokenizer.ggml.seperator_token_id" },
+    { LLM_KV_TOKENIZER_PAD_ID,              "tokenizer.ggml.padding_token_id"   },
+    { LLM_KV_TOKENIZER_HF_JSON,             "tokenizer.huggingface.json"        },
+    { LLM_KV_TOKENIZER_RWKV,                "tokenizer.rwkv.world"              },
+};
+
+struct LLM_KV {
+    LLM_KV(llm_arch arch) : arch(arch) {}
+
+    llm_arch arch;
+
+    std::string operator()(llm_kv kv) const {
+        return ::format(LLM_KV_NAMES[kv].c_str(), LLM_ARCH_NAMES[arch].c_str());
+    }
+};
+
+enum llm_tensor {
+    LLM_TENSOR_TOKEN_EMBD,
+    LLM_TENSOR_POS_EMBD,
+    LLM_TENSOR_OUTPUT,
+    LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_ROPE_FREQS,
+    LLM_TENSOR_ATTN_Q,
+    LLM_TENSOR_ATTN_K,
+    LLM_TENSOR_ATTN_V,
+    LLM_TENSOR_ATTN_QKV,
+    LLM_TENSOR_ATTN_OUT,
+    LLM_TENSOR_ATTN_NORM,
+    LLM_TENSOR_ATTN_NORM_2,
+    LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_FFN_GATE,
+    LLM_TENSOR_FFN_DOWN,
+    LLM_TENSOR_FFN_UP,
+    LLM_TENSOR_FFN_NORM,
+    LLM_TENSOR_ATTN_Q_NORM,
+    LLM_TENSOR_ATTN_K_NORM,
+};
+
+static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
+    {
+        LLM_ARCH_LLAMA,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_BAICHUAN,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_FALCON,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_NORM_2,     "blk.%d.attn_norm_2" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_GPT2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_GPTJ,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_GPTNEOX,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_PERSIMMON,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd"},
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm"},
+            { LLM_TENSOR_OUTPUT,          "output"},
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm"},
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv"},
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output"},
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm"},
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm"},
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm"},
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down"},
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up"},
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd"},
+        },
+    },
+    {
+        LLM_ARCH_MPT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+    {
+        LLM_ARCH_STARCODER,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_POS_EMBD,        "position_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+        },
+    },
+    {
+        LLM_ARCH_REFACT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_UNKNOWN,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+        },
+    },
+};
+
+static llm_arch llm_arch_from_string(const std::string & name) {
+    for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT
+        if (kv.second == name) {
+            return kv.first;
+        }
+    }
+
+    return LLM_ARCH_UNKNOWN;
+}
+
+// helper to handle gguf constants
+// usage:
+//
+//   const auto tn = LLM_TN(LLM_ARCH_LLAMA);
+//
+//   std::string name = tn(LLM_TENSOR_OUTPUT);                     -> "output"
+//   std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias");         -> "token_embd.bias"
+//   std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3);     -> "blk.3.attn_norm.weight"
+//
+struct LLM_TN {
+    LLM_TN(llm_arch arch) : arch(arch) {}
+
+    llm_arch arch;
+
+    std::string operator()(llm_tensor tensor) const {
+        return LLM_TENSOR_NAMES[arch].at(tensor);
+    }
+
+    std::string operator()(llm_tensor tensor, const std::string & suffix) const {
+        return LLM_TENSOR_NAMES[arch].at(tensor) + "." + suffix;
+    }
+
+    std::string operator()(llm_tensor tensor, int bid) const {
+        return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid);
+    }
+
+    std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const {
+        return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid) + "." + suffix;
+    }
+};
+
+//
+// gguf helpers
+//
+
+#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \
+do { \
+    const std::string skey(key); \
+    const int kid = gguf_find_key(ctx, skey.c_str()); \
+    if (kid >= 0) { \
+        enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
+        if (ktype != (type)) { \
+            throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \
+        } \
+        (dst) = func(ctx, kid); \
+    } else if (req) { \
+        throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \
+    } \
+} while (0)
+
 //
 // ggml helpers
 //
@@ -167,6 +551,9 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
 #elif GGML_USE_METAL
 #   define llama_host_malloc(n)  ggml_metal_host_malloc(n)
 #   define llama_host_free(data) ggml_metal_host_free(data)
+#elif GGML_USE_CPU_HBM
+#   define llama_host_malloc(n)  hbw_malloc(n)
+#   define llama_host_free(data) if (data != NULL) hbw_free(data)
 #else
 #   define llama_host_malloc(n)  malloc(n)
 #   define llama_host_free(data) free(data)
@@ -323,16 +710,16 @@ struct llama_mmap {
 
         if (prefetch > 0) {
             // Advise the kernel to preload the mapped memory
-            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
-                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+            if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
                         strerror(errno));
             }
         }
         if (numa) {
             // advise the kernel not to use readahead
             // (because the next page might not belong on the same node)
-            if (madvise(addr, file->size, MADV_RANDOM)) {
-                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+            if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
                         strerror(errno));
             }
         }
@@ -366,20 +753,25 @@ struct llama_mmap {
             throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
         }
 
-        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
         if (prefetch) {
-            // Advise the kernel to preload the mapped memory
-            WIN32_MEMORY_RANGE_ENTRY range;
-            range.VirtualAddress = addr;
-            range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
+            // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it
+            BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
+            HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
+
+            // may fail on pre-Windows 8 systems
+            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+
+            if (pPrefetchVirtualMemory) {
+                // advise the kernel to preload the mapped memory
+                WIN32_MEMORY_RANGE_ENTRY range;
+                range.VirtualAddress = addr;
+                range.NumberOfBytes = (SIZE_T)size;
+                if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                    fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                            llama_format_win_err(GetLastError()).c_str());
+                }
             }
         }
-        #else
-        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
-        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
     }
 
     ~llama_mmap() {
@@ -556,12 +948,12 @@ static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
 
-static std::string llama_token_to_text(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_str(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -576,7 +968,7 @@ static std::string llama_token_to_text(const struct llama_context * ctx, llama_t
 
 struct llama_state {
     // We save the log callback globally
-    llama_log_callback log_callback = llama_log_callback_default;
+    ggml_log_callback log_callback = llama_log_callback_default;
     void * log_callback_user_data = nullptr;
 };
 
@@ -585,36 +977,59 @@ static llama_state g_state;
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_1B,
     MODEL_3B,
     MODEL_7B,
+    MODEL_8B,
     MODEL_13B,
+    MODEL_15B,
     MODEL_30B,
+    MODEL_34B,
+    MODEL_40B,
     MODEL_65B,
     MODEL_70B,
 };
 
 static const size_t kB = 1024;
-static const size_t MB = 1024*1024;
+static const size_t MB = kB*kB;
+static const size_t GB = kB*kB*kB;
 
-// default hparams (LLaMA 7B)
 struct llama_hparams {
-    uint32_t n_vocab     = 32000;
-    uint32_t n_ctx_train = 2048;  // the context size used during training
-    uint32_t n_ctx       = 512;   // the context size used during inference
-    uint32_t n_embd      = 4096;
-    uint32_t n_head      = 32;
-    uint32_t n_head_kv   = 32;
-    uint32_t n_layer     = 32;
-    uint32_t n_rot       = 64;
-    uint32_t n_ff        = 11008;
+    bool     vocab_only;
+    uint32_t n_vocab;
+    uint32_t n_ctx_train; // context size the model was trained on
+    uint32_t n_embd;
+    uint32_t n_head;
+    uint32_t n_head_kv;
+    uint32_t n_layer;
+    uint32_t n_rot;
+    uint32_t n_ff;
 
-    float f_norm_rms_eps = 1e-5;
+    float f_norm_eps;
+    float f_norm_rms_eps;
 
-    float rope_freq_base  = 10000.0f;
-    float rope_freq_scale = 1.0f;
+    float rope_freq_base_train;
+    float rope_freq_scale_train;
 
     bool operator!=(const llama_hparams & other) const {
-        return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams))); // NOLINT
+        if (this->vocab_only != other.vocab_only) return true;
+        if (this->n_vocab != other.n_vocab) return true;
+        if (this->n_ctx_train != other.n_ctx_train) return true;
+        if (this->n_embd != other.n_embd) return true;
+        if (this->n_head != other.n_head) return true;
+        if (this->n_head_kv != other.n_head_kv) return true;
+        if (this->n_layer != other.n_layer) return true;
+        if (this->n_rot != other.n_rot) return true;
+        if (this->n_ff != other.n_ff) return true;
+
+        const float EPSILON = 1e-9;
+
+        if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
+        if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
+        if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
+        if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
+
+        return false;
     }
 
     uint32_t n_gqa() const {
@@ -628,37 +1043,82 @@ struct llama_hparams {
     uint32_t n_embd_gqa() const {
         return n_embd/n_gqa();
     }
+};
 
-    size_t kv_size() const {
-        size_t result = 2ull;
-        result *= (size_t) n_embd_gqa();
-        result *= (size_t) n_ctx;
-        result *= (size_t) n_layer;
-        result *= sizeof(ggml_fp16_t);
-        return result;
-    }
+struct llama_cparams {
+    uint32_t n_ctx;       // context size used during inference
+    uint32_t n_batch;
+    uint32_t n_threads;       // number of threads to use for generation
+    uint32_t n_threads_batch; // number of threads to use for batch processing
+
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    bool mul_mat_q;
 };
 
 struct llama_layer {
     // normalization
-    struct ggml_tensor * attention_norm;
+    struct ggml_tensor * attn_norm;
+    struct ggml_tensor * attn_norm_b;
+    struct ggml_tensor * attn_norm_2;
+    struct ggml_tensor * attn_norm_2_b;
+    struct ggml_tensor * attn_q_norm;
+    struct ggml_tensor * attn_q_norm_b;
+    struct ggml_tensor * attn_k_norm;
+    struct ggml_tensor * attn_k_norm_b;
 
     // attention
     struct ggml_tensor * wq;
     struct ggml_tensor * wk;
     struct ggml_tensor * wv;
     struct ggml_tensor * wo;
+    struct ggml_tensor * wqkv;
+
+    // attention bias
+    struct ggml_tensor * bo;
+    struct ggml_tensor * bqkv;
 
     // normalization
     struct ggml_tensor * ffn_norm;
+    struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * w1; // ffn_gate
+    struct ggml_tensor * w2; // ffn_down
+    struct ggml_tensor * w3; // ffn_up
+
+    // ff bias
+    struct ggml_tensor * b2; // ffn_down
+    struct ggml_tensor * b3; // ffn_up
 };
 
+struct llama_kv_cell {
+    llama_pos pos   = -1;
+    llama_pos delta = 0;
+
+    std::set<llama_seq_id> seq_id;
+
+    bool has_seq_id(const llama_seq_id & id) const {
+        return seq_id.find(id) != seq_id.end();
+    }
+};
+
+// ring-buffer of cached KV data
 struct llama_kv_cache {
+    bool has_shift = false;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_internal also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    std::vector<llama_kv_cell> cells;
+
     struct ggml_tensor * k = NULL;
     struct ggml_tensor * v = NULL;
 
@@ -666,8 +1126,6 @@ struct llama_kv_cache {
 
     llama_buffer buf;
 
-    int n; // number of tokens currently in the cache
-
     ~llama_kv_cache() {
         if (ctx) {
             ggml_free(ctx);
@@ -681,10 +1139,6 @@ struct llama_kv_cache {
 };
 
 struct llama_vocab {
-    // TODO:
-    // - add a vector of merges
-    //   so that we can pass it to different types of tokenizers with a common interface
-
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -695,11 +1149,13 @@ struct llama_vocab {
         ttype type;
     };
 
-    llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM;
+    enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM;
 
     std::unordered_map<token, id> token_to_id;
     std::vector<token_data>       id_to_token;
 
+    std::map<std::pair<std::string, std::string>, int> bpe_ranks;
+
     // default LLaMA special tokens
     id special_bos_id = 1;
     id special_eos_id = 2;
@@ -708,21 +1164,45 @@ struct llama_vocab {
     id special_pad_id = -1;
 
     id linefeed_id = 13;
+    id special_prefix_id = 32007;
+    id special_middle_id = 32009;
+    id special_suffix_id = 32008;
+    id special_eot_id = 32010;
+
+    int find_bpe_rank(std::string token_left, std::string token_right) const {
+        replace_all(token_left,  " ",  "\u0120");
+        replace_all(token_left,  "\n", "\u010A");
+        replace_all(token_right, " ",  "\u0120");
+        replace_all(token_right, "\n", "\u010A");
+
+        auto it = bpe_ranks.find(std::make_pair(token_left, token_right));
+        if (it == bpe_ranks.end()) {
+            return -1;
+        }
+
+        return it->second;
+    }
 };
 
 struct llama_model {
     e_model     type  = MODEL_UNKNOWN;
+    llm_arch    arch  = LLM_ARCH_UNKNOWN;
     llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
 
-    llama_hparams hparams;
+    std::string name = "n/a";
+
+    llama_hparams hparams = {};
     llama_vocab   vocab;
 
     struct ggml_tensor * tok_embeddings;
+    struct ggml_tensor * pos_embeddings;
 
-    struct ggml_tensor * norm;
+    struct ggml_tensor * output_norm;
+    struct ggml_tensor * output_norm_b;
     struct ggml_tensor * output;
 
     std::vector<llama_layer> layers;
+
     int n_gpu_layers;
 
     // context
@@ -763,11 +1243,8 @@ struct llama_model {
 };
 
 struct llama_context {
-    llama_context(const llama_model & model) : model(model), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {}
+    llama_context(const llama_model & model) : model(model), t_start_us(model.t_start_us), t_load_us(model.t_load_us) {}
     ~llama_context() {
-        if (model_owner) {
-            delete &model;
-        }
 #ifdef GGML_USE_METAL
         if (ctx_metal) {
             ggml_metal_free(ctx_metal);
@@ -778,29 +1255,26 @@ struct llama_context {
         }
     }
 
-    std::mt19937 rng;
-
-    bool has_evaluated_once = false;
-
-    int64_t t_sample_us = 0;
-    int64_t t_eval_us   = 0;
-    int64_t t_p_eval_us = 0;
-
-    int32_t n_sample = 0; // number of tokens sampled
-    int32_t n_eval   = 0; // number of eval calls
-    int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    llama_cparams cparams;
 
     const llama_model & model;
 
-    bool model_owner = false;
-
-    int64_t t_load_us;
-    int64_t t_start_us;
-
     // key + value cache for the self attention
     struct llama_kv_cache kv_self;
 
-    size_t mem_per_token = 0;
+    std::mt19937 rng;
+
+    bool has_evaluated_once = false;
+
+    int64_t t_start_us;
+    int64_t t_load_us;
+    int64_t t_sample_us = 0;
+    int64_t t_p_eval_us = 0;
+    int64_t t_eval_us   = 0;
+
+    int32_t n_sample = 0; // number of tokens sampled
+    int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    int32_t n_eval   = 0; // number of eval calls
 
     // decode output (2-dimensional array: [n_tokens][n_vocab])
     std::vector<float> logits;
@@ -835,16 +1309,23 @@ static bool llama_kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
                          ggml_type   wtype,
-                               int   n_ctx,
+                          uint32_t   n_ctx,
                                int   n_gpu_layers) {
-    const int n_embd  = hparams.n_embd_gqa();
-    const int n_layer = hparams.n_layer;
+    const uint32_t n_embd  = hparams.n_embd_gqa();
+    const uint32_t n_layer = hparams.n_layer;
 
     const int64_t n_mem      = n_layer*n_ctx;
     const int64_t n_elements = n_embd*n_mem;
 
+    cache.has_shift = false;
+
+    cache.head = 0;
+    cache.size = n_ctx;
+
+    cache.cells.clear();
+    cache.cells.resize(n_ctx);
+
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
-    cache.n = 0;
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size;
@@ -865,38 +1346,212 @@ static bool llama_kv_cache_init(
 
     (void) n_gpu_layers;
 #ifdef GGML_USE_CUBLAS
-    if (n_gpu_layers > n_layer + 1) {
+    size_t vram_kv_cache = 0;
+
+    if (n_gpu_layers > (int)n_layer + 1) {
         ggml_cuda_assign_buffers_no_scratch(cache.v);
+        LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__);
+        vram_kv_cache += ggml_nbytes(cache.v);
     }
-    if (n_gpu_layers > n_layer + 2) {
+    if (n_gpu_layers > (int)n_layer + 2) {
         ggml_cuda_assign_buffers_no_scratch(cache.k);
+        LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__);
+        vram_kv_cache += ggml_nbytes(cache.k);
+    }
+    if (vram_kv_cache > 0) {
+        LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MB\n", __func__, vram_kv_cache / 1024.0 / 1024.0);
     }
 #endif // GGML_USE_CUBLAS
 
     return true;
 }
 
+// find an empty slot of size "n_tokens" in the cache
+// updates the cache head
+// Note: On success, it's important that cache.head points
+// to the first cell of the slot.
+static bool llama_kv_cache_find_slot(
+           struct llama_kv_cache & cache,
+        const struct llama_batch & batch) {
+    const uint32_t n_ctx    = cache.size;
+    const uint32_t n_tokens = batch.n_tokens;
+
+    if (n_tokens > n_ctx) {
+        LLAMA_LOG_ERROR("%s: n_tokens=%d > n_ctx=%d\n", __func__, n_tokens, n_ctx);
+        return false;
+    }
+
+    uint32_t n_tested = 0;
+
+    while (true) {
+        if (cache.head + n_tokens > n_ctx) {
+            n_tested += n_ctx - cache.head;
+            cache.head = 0;
+            continue;
+        }
+
+        bool found = true;
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            if (cache.cells[cache.head + i].pos >= 0) {
+                found = false;
+                cache.head += i + 1;
+                n_tested   += i + 1;
+                break;
+            }
+        }
+
+        if (found) {
+            break;
+        }
+
+        if (n_tested >= n_ctx) {
+            //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
+            return false;
+        }
+    }
+
+    for (uint32_t i = 0; i < n_tokens; i++) {
+        cache.cells[cache.head + i].pos = batch.pos[i];
+        cache.cells[cache.head + i].seq_id.insert(batch.seq_id[i]);
+    }
+
+    return true;
+}
+
+// find how many cells are currently in use
+static int32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) {
+    for (uint32_t i = cache.size - 1; i > 0; --i) {
+        if (cache.cells[i].pos >= 0 && !cache.cells[i].seq_id.empty()) {
+            return i + 1;
+        }
+    }
+
+    return 0;
+}
+
+static void llama_kv_cache_tokens_rm(struct llama_kv_cache & cache, int32_t c0, int32_t c1) {
+    if (c0 < 0) c0 = 0;
+    if (c1 < 0) c1 = cache.size;
+
+    for (int32_t i = c0; i < c1; ++i) {
+        cache.cells[i].pos = -1;
+        cache.cells[i].seq_id.clear();
+    }
+
+    // Searching for a free slot can start here since we know it will be empty.
+    cache.head = uint32_t(c0);
+}
+
+static void llama_kv_cache_seq_rm(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1) {
+    uint32_t new_head = cache.size;
+
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.cells[i].seq_id.erase(seq_id);
+            if (cache.cells[i].seq_id.empty()) {
+                cache.cells[i].pos = -1;
+                if (new_head == cache.size) new_head = i;
+            }
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    if (new_head != cache.size) cache.head = new_head;
+}
+
+static void llama_kv_cache_seq_cp(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id_src,
+                 llama_seq_id   seq_id_dst,
+                    llama_pos   p0,
+                    llama_pos   p1) {
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    cache.head = 0;
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id_src) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.cells[i].seq_id.insert(seq_id_dst);
+        }
+    }
+}
+
+static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id seq_id) {
+    uint32_t new_head = cache.size;
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (!cache.cells[i].has_seq_id(seq_id)) {
+            cache.cells[i].pos = -1;
+            cache.cells[i].seq_id.clear();
+            if (new_head == cache.size) new_head = i;
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    if (new_head != cache.size) cache.head = new_head;
+}
+
+static void llama_kv_cache_seq_shift(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                    llama_pos   delta) {
+    uint32_t new_head = cache.size;
+
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.cells[i].pos += delta;
+            if (cache.cells[i].pos < 0) {
+                cache.cells[i].pos = -1;
+                cache.cells[i].seq_id.clear();
+                if (new_head == cache.size) new_head = i;
+            } else {
+                cache.has_shift = true;
+                cache.cells[i].delta = delta;
+            }
+        }
+    }
+
+    // If we freed up a slot, set head to it so searching can start there.
+    // Otherwise we just start the next search from the beginning.
+    cache.head = new_head != cache.size ? new_head : 0;
+}
+
 //
 // model loading and saving
 //
 
-enum llama_file_version {
+enum llama_fver {
     GGUF_FILE_VERSION_V1 = 1,
+    GGUF_FILE_VERSION_V2 = 2,
 };
 
-static const char * llama_file_version_name(llama_file_version version) {
+static const char * llama_file_version_name(llama_fver version) {
     switch (version) {
-        case GGUF_FILE_VERSION_V1: return "GGUF V1 (latest)";
+        case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
+        case GGUF_FILE_VERSION_V2: return "GGUF V2 (latest)";
     }
 
     return "unknown";
 }
 
-static std::string llama_format_tensor_shape(const std::vector<uint32_t> & ne) {
+static std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
     char buf[256];
-    snprintf(buf, sizeof(buf), "%5u", ne.at(0));
+    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
     for (size_t i = 1; i < ne.size(); i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5u", ne.at(i));
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
     }
     return buf;
 }
@@ -916,12 +1571,13 @@ struct llama_model_loader {
     int n_created = 0;
 
     int64_t n_elements = 0;
+    size_t  n_bytes    = 0;
 
     bool use_mmap = false;
 
-    llama_file file;
+    llama_file  file;
     llama_ftype ftype;
-    llama_file_version fver;
+    llama_fver  fver;
 
     std::unique_ptr<llama_mmap> mapping;
 
@@ -942,12 +1598,13 @@ struct llama_model_loader {
         n_kv      = gguf_get_n_kv(ctx_gguf);
         n_tensors = gguf_get_n_tensors(ctx_gguf);
 
-        fver = (enum llama_file_version) gguf_get_version(ctx_gguf);
+        fver = (enum llama_fver ) gguf_get_version(ctx_gguf);
 
         for (int i = 0; i < n_tensors; i++) {
             const char * name = gguf_get_tensor_name(ctx_gguf, i);
             struct ggml_tensor * t = ggml_get_tensor(ctx_meta, name);
             n_elements += ggml_nelements(t);
+            n_bytes    += ggml_nbytes(t);
         }
 
         LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n",
@@ -1039,6 +1696,21 @@ struct llama_model_loader {
         }
     }
 
+    std::string get_arch_name() const {
+        const auto kv = LLM_KV(LLM_ARCH_UNKNOWN);
+
+        std::string arch_name;
+        GGUF_GET_KEY(ctx_gguf, arch_name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_ARCHITECTURE));
+
+        return arch_name;
+    }
+
+    enum llm_arch get_arch() const {
+        const std::string arch_name = get_arch_name();
+
+        return llm_arch_from_string(arch_name);
+    }
+
     const char * get_tensor_name(int i) const {
         return gguf_get_tensor_name(ctx_gguf, i);
     }
@@ -1076,7 +1748,7 @@ struct llama_model_loader {
         return tensor;
     }
 
-    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<uint32_t> & ne, ggml_backend backend) {
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, ggml_backend backend) {
         struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str());
 
         if (cur == NULL) {
@@ -1162,7 +1834,11 @@ struct llama_model_loader {
             // allocate temp buffer if not using mmap
             if (!use_mmap && cur->data == NULL) {
                 GGML_ASSERT(cur->backend != GGML_BACKEND_CPU);
-                cur->data = malloc(ggml_nbytes(cur));
+                #ifdef GGML_USE_CPU_HBM
+                cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur));
+                #else
+                cur->data = (uint8_t*)malloc(ggml_nbytes(cur));
+                #endif
             }
 
             load_data_for(cur);
@@ -1174,7 +1850,7 @@ struct llama_model_loader {
                         lmlock->grow_to(size_lock);
                     }
                     break;
-#if defined(GGML_USE_CUBLAS)
+#ifdef GGML_USE_CUBLAS
                 case GGML_BACKEND_GPU:
                 case GGML_BACKEND_GPU_SPLIT:
                     // old code:
@@ -1207,7 +1883,15 @@ struct llama_model_loader {
 // load LLaMA models
 //
 
-std::string llama_model_ftype_name(enum llama_ftype ftype) {
+static std::string llama_model_arch_name(llm_arch arch) {
+    auto it = LLM_ARCH_NAMES.find(arch);
+    if (it == LLM_ARCH_NAMES.end()) {
+        return "unknown";
+    }
+    return it->second;
+}
+
+static std::string llama_model_ftype_name(llama_ftype ftype) {
     if (ftype & LLAMA_FTYPE_GUESSED) {
         return llama_model_ftype_name((enum llama_ftype) (ftype & ~LLAMA_FTYPE_GUESSED)) + " (guessed)";
     }
@@ -1240,232 +1924,326 @@ std::string llama_model_ftype_name(enum llama_ftype ftype) {
 
 static const char * llama_model_type_name(e_model type) {
     switch (type) {
+        case MODEL_1B:  return "1B";
         case MODEL_3B:  return "3B";
         case MODEL_7B:  return "7B";
+        case MODEL_8B:  return "8B";
         case MODEL_13B: return "13B";
+        case MODEL_15B: return "15B";
         case MODEL_30B: return "30B";
+        case MODEL_34B: return "34B";
+        case MODEL_40B: return "40B";
         case MODEL_65B: return "65B";
         case MODEL_70B: return "70B";
-        default: GGML_ASSERT(false);
+        default:        return "?B";
     }
 }
 
-static void llama_model_load_internal(
-        const std::string & fname,
+static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
+    model.arch = ml.get_arch();
+    if (model.arch == LLM_ARCH_UNKNOWN) {
+        throw std::runtime_error("unknown model architecture: '" + ml.get_arch_name() + "'");
+    }
+}
+
+static void llm_load_hparams(
+        llama_model_loader & ml,
+        llama_model & model) {
+    struct gguf_context * ctx = ml.ctx_gguf;
+
+    const auto kv = LLM_KV(model.arch);
+
+    auto & hparams = model.hparams;
+
+    // get general kv
+    GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME));
+
+    // get hparams kv
+    GGUF_GET_KEY(ctx, hparams.n_vocab,        gguf_get_arr_n,   GGUF_TYPE_ARRAY,  true, kv(LLM_KV_TOKENIZER_LIST));
+    GGUF_GET_KEY(ctx, hparams.n_ctx_train,    gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_CONTEXT_LENGTH));
+    GGUF_GET_KEY(ctx, hparams.n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH));
+    GGUF_GET_KEY(ctx, hparams.n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH));
+    GGUF_GET_KEY(ctx, hparams.n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT));
+    GGUF_GET_KEY(ctx, hparams.n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT));
+
+    // n_head_kv is optional, default to n_head
+    hparams.n_head_kv = hparams.n_head;
+    GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV));
+
+    // rope_freq_base (optional)
+    hparams.rope_freq_base_train = 10000.0f;
+    GGUF_GET_KEY(ctx, hparams.rope_freq_base_train, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
+
+    // rope_freq_scale (inverse of the kv) is optional
+    float ropescale = 1.0f;
+    GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    hparams.rope_freq_scale_train = 1.0f/ropescale;
+
+    // sanity check for n_rot (optional)
+    {
+        hparams.n_rot = hparams.n_embd / hparams.n_head;
+
+        GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT));
+
+        if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
+            if (hparams.n_rot != hparams.n_embd / hparams.n_head) {
+                throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head));
+            }
+        }
+        // gpt-neox n_rot = rotary_pct * (n_embd / n_head)
+        // gpt-j n_rot = rotary_dim
+    }
+
+    // arch-specific KVs
+    switch (model.arch) {
+        case LLM_ARCH_LLAMA:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+
+                switch (hparams.n_layer) {
+                    case 26: model.type = e_model::MODEL_3B; break;
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    case 48: model.type = e_model::MODEL_34B; break;
+                    case 60: model.type = e_model::MODEL_30B; break;
+                    case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_FALCON:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 60: model.type = e_model::MODEL_40B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_BAICHUAN:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_STARCODER:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+                switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
+                    case 36: model.type = e_model::MODEL_3B; break;
+                    case 42: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_15B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_PERSIMMON:
+        {
+            GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
+            switch (hparams.n_layer) {
+                case 36: model.type = e_model::MODEL_8B; break;
+                default: model.type = e_model::MODEL_UNKNOWN;
+            }
+        } break;
+        case LLM_ARCH_REFACT:
+            {
+                GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_1B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        default: (void)0;
+    }
+
+    model.ftype = ml.ftype;
+}
+
+// TODO: This should probably be in llama.h
+static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos);
+static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch);
+
+static void llm_load_vocab(
+        llama_model_loader & ml,
+        llama_model & model) {
+    auto & vocab = model.vocab;
+
+    struct gguf_context * ctx = ml.ctx_gguf;
+
+    const auto kv = LLM_KV(model.arch);
+
+    const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str());
+    if (token_idx == -1) {
+        throw std::runtime_error("cannot find tokenizer vocab in model file\n");
+    }
+
+    const float * scores = nullptr;
+    const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str());
+    if (score_idx != -1) {
+        scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
+    }
+
+    const int * toktypes = nullptr;
+    const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str());
+    if (toktype_idx != -1) {
+        toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
+    }
+
+    // determine vocab type
+    {
+        std::string tokenizer_name;
+
+        GGUF_GET_KEY(ctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL));
+
+        if (tokenizer_name == "llama") {
+            vocab.type = LLAMA_VOCAB_TYPE_SPM;
+
+            // default special tokens
+            vocab.special_bos_id = 1;
+            vocab.special_eos_id = 2;
+            vocab.special_unk_id = 0;
+            vocab.special_sep_id = -1;
+            vocab.special_pad_id = -1;
+        } else if (tokenizer_name == "gpt2") {
+            vocab.type = LLAMA_VOCAB_TYPE_BPE;
+
+            // read bpe merges and populate bpe ranks
+            const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str());
+            if (merges_keyidx == -1) {
+                throw std::runtime_error("cannot find tokenizer merges in model file\n");
+            }
+
+            const int n_merges = gguf_get_arr_n(ctx, merges_keyidx);
+
+            for (int i = 0; i < n_merges; i++) {
+                const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
+                GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
+
+                std::string first;
+                std::string second;
+
+                const size_t pos = word.find(' ', 1);
+
+                if (pos != std::string::npos) {
+                    first  = word.substr(0, pos);
+                    second = word.substr(pos + 1);
+                }
+
+                vocab.bpe_ranks.emplace(std::make_pair(first, second), i);
+            }
+
+            // default special tokens
+            vocab.special_bos_id = 11;
+            vocab.special_eos_id = 11;
+            vocab.special_unk_id = -1;
+            vocab.special_sep_id = -1;
+            vocab.special_pad_id = -1;
+        } else {
+            LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str());
+            LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__);
+
+            vocab.type = LLAMA_VOCAB_TYPE_SPM;
+        }
+    }
+
+    const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+
+    vocab.id_to_token.resize(n_vocab);
+
+    for (uint32_t i = 0; i < n_vocab; i++) {
+        std::string word = gguf_get_arr_str(ctx, token_idx, i);
+        GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
+
+        vocab.token_to_id[word] = i;
+
+        auto & token_data = vocab.id_to_token[i];
+        token_data.text  = std::move(word);
+        token_data.score = scores ? scores[i] : 0.0f;
+        token_data.type  = toktypes ? (llama_token_type) toktypes[i] : LLAMA_TOKEN_TYPE_NORMAL;
+    }
+    GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
+
+    // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
+    if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
+        vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
+    } else {
+        vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0];
+    }
+
+    // special tokens
+    GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
+    GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
+    GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
+    GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
+    GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
+}
+
+static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
+    const auto & hparams = model.hparams;
+    const auto & vocab   = model.vocab;
+
+    // hparams
+    LLAMA_LOG_INFO("%s: format           = %s\n",     __func__, llama_file_version_name(ml.fver));
+    LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, LLM_ARCH_NAMES.at(model.arch).c_str());
+    LLAMA_LOG_INFO("%s: vocab type       = %s\n",     __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix
+    LLAMA_LOG_INFO("%s: n_vocab          = %u\n",     __func__, hparams.n_vocab);
+    LLAMA_LOG_INFO("%s: n_merges         = %u\n",     __func__, (int) vocab.bpe_ranks.size());
+    LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
+    LLAMA_LOG_INFO("%s: n_embd           = %u\n",     __func__, hparams.n_embd);
+    LLAMA_LOG_INFO("%s: n_head           = %u\n",     __func__, hparams.n_head);
+    LLAMA_LOG_INFO("%s: n_head_kv        = %u\n",     __func__, hparams.n_head_kv);
+    LLAMA_LOG_INFO("%s: n_layer          = %u\n",     __func__, hparams.n_layer);
+    LLAMA_LOG_INFO("%s: n_rot            = %u\n",     __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
+    LLAMA_LOG_INFO("%s: n_gqa            = %u\n",     __func__, hparams.n_gqa());
+    LLAMA_LOG_INFO("%s: f_norm_eps       = %.1e\n",   __func__, hparams.f_norm_eps);
+    LLAMA_LOG_INFO("%s: f_norm_rms_eps   = %.1e\n",   __func__, hparams.f_norm_rms_eps);
+    LLAMA_LOG_INFO("%s: n_ff             = %u\n",     __func__, hparams.n_ff);
+    LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
+    LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
+    LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, llama_model_type_name(model.type));
+    LLAMA_LOG_INFO("%s: model ftype      = %s\n",     __func__, llama_model_ftype_name(model.ftype).c_str());
+    LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    if (ml.n_bytes < GB) {
+        LLAMA_LOG_INFO("%s: model size       = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements);
+    } else {
+        LLAMA_LOG_INFO("%s: model size       = %.2f GiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements);
+    }
+
+    // general kv
+    LLAMA_LOG_INFO("%s: general.name   = %s\n",    __func__, model.name.c_str());
+
+    // special tokens
+    if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); }
+    if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); }
+    if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); }
+    if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); }
+    if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); }
+    if (vocab.linefeed_id    != -1) { LLAMA_LOG_INFO( "%s: LF token  = %d '%s'\n", __func__, vocab.linefeed_id,    vocab.id_to_token[vocab.linefeed_id].text.c_str() );    }
+}
+
+static void llm_load_tensors(
+        llama_model_loader & ml,
         llama_model & model,
-        llama_vocab & vocab,
-        int n_ctx,
-        int n_batch,
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
-        const bool mul_mat_q,
-        float rope_freq_base,
-        float rope_freq_scale,
-        bool low_vram,
-        ggml_type memory_type,
-        bool use_mmap,
         bool use_mlock,
-        bool vocab_only,
         llama_progress_callback progress_callback,
         void * progress_callback_user_data) {
     model.t_start_us = ggml_time_us();
 
-    std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname, use_mmap));
-
-    model.n_gpu_layers = n_gpu_layers;
-
+    auto & ctx     = model.ctx;
     auto & hparams = model.hparams;
 
-    std::string general_name = "n/a";
-    std::string general_arch = "n/a";
-
-    // read hparams
-    {
-        struct gguf_context * ctx = ml->ctx_gguf;
-
-#define GGUF_GET(dst, func, type, req, key) \
-        { \
-            const int kid = gguf_find_key(ctx, key); \
-            if (kid >= 0) { \
-                enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \
-                if (ktype != (type)) { \
-                    throw std::runtime_error(format("key %s has wrong type: %s", key, gguf_type_name(ktype))); \
-                } \
-                (dst) = func(ctx, kid); \
-            } else if (req) { \
-                throw std::runtime_error(format("key not found in model: %s", key)); \
-            } \
-        }
-
-        std::string tokenizer_name;
-        GGUF_GET(tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, "tokenizer.ggml.model");
-
-        if (tokenizer_name == "llama") {
-            vocab.type = LLAMA_VOCAB_TYPE_SPM;
-        } else if (tokenizer_name == "gpt2") {
-            vocab.type = LLAMA_VOCAB_TYPE_BPE;
-        } else {
-            LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str());
-            LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__);
-            vocab.type = LLAMA_VOCAB_TYPE_SPM;
-        }
-
-        // get hparams kv
-        GGUF_GET(hparams.n_vocab,        gguf_get_arr_n,   GGUF_TYPE_ARRAY,   true, "tokenizer.ggml.tokens");
-        GGUF_GET(hparams.n_ctx_train,    gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.context_length");
-        GGUF_GET(hparams.n_embd,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.embedding_length");
-        GGUF_GET(hparams.n_ff,           gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.feed_forward_length");
-        GGUF_GET(hparams.n_head,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.attention.head_count");
-        GGUF_GET(hparams.n_layer,        gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.block_count");
-        GGUF_GET(hparams.n_rot,          gguf_get_val_u32, GGUF_TYPE_UINT32,  true, "llama.rope.dimension_count");
-        GGUF_GET(hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, "llama.attention.layer_norm_rms_epsilon");
-
-        // n_head_kv is optional, default to n_head
-        hparams.n_head_kv = hparams.n_head;
-        GGUF_GET(hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "llama.attention.head_count_kv");
-
-        // TODO: manually setting rope scale should override this
-        // rope_freq_scale (inverse of the kv) is optional
-        float ropescale = 1.0f;
-        GGUF_GET(ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, "llama.rope.scale_linear");
-        if (ropescale != 1.0f) {
-            rope_freq_scale = 1.0f/ropescale;
-        }
-
-        // get general kv
-        GGUF_GET(general_name, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.name");
-        GGUF_GET(general_arch, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.architecture");
-
-        // special tokens
-        GGUF_GET(vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.bos_token_id");
-        GGUF_GET(vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.eos_token_id");
-        GGUF_GET(vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.unknown_token_id");
-        GGUF_GET(vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.separator_token_id");
-        GGUF_GET(vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.padding_token_id");
-
-#undef GGUF_GET
-
-        switch (hparams.n_layer) {
-            case 26: model.type = e_model::MODEL_3B; break;
-            case 32: model.type = e_model::MODEL_7B; break;
-            case 40: model.type = e_model::MODEL_13B; break;
-            case 60: model.type = e_model::MODEL_30B; break;
-            case 80: model.type = e_model::MODEL_65B; break;
-            default:
-                {
-                    if (hparams.n_layer < 32) {
-                        model.type = e_model::MODEL_7B;
-                    }
-                } break;
-        }
-
-        model.ftype = ml->ftype;
-
-        hparams.n_ctx = n_ctx;
-
-        // LLaMAv2
-        // TODO: probably not needed
-        {
-            const auto n_gqa = hparams.n_gqa();
-
-            if (model.type == e_model::MODEL_65B && n_gqa == 8) {
-                LLAMA_LOG_WARN("%s: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
-                model.type = e_model::MODEL_70B;
-            }
-        }
-
-        hparams.rope_freq_base  = rope_freq_base;
-        hparams.rope_freq_scale = rope_freq_scale;
-    }
-
-    // read vocab
-    {
-        struct gguf_context * ctx = ml->ctx_gguf;
-
-        vocab.id_to_token.resize(hparams.n_vocab);
-
-        const int token_idx = gguf_find_key(ctx, "tokenizer.ggml.tokens");
-        if (token_idx == -1) {
-            throw std::runtime_error("cannot find tokenizer vocab in model file\n");
-        }
-
-        const int score_idx = gguf_find_key(ctx, "tokenizer.ggml.scores");
-        if (score_idx == -1) {
-            throw std::runtime_error("cannot find tokenizer scores in model file\n");
-        }
-
-        const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
-
-        const int toktype_idx = gguf_find_key(ctx, "tokenizer.ggml.token_type");
-        if (toktype_idx == -1) {
-            throw std::runtime_error("cannot find token type list in GGUF file\n");
-        }
-
-        const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
-
-        for (uint32_t i = 0; i < hparams.n_vocab; i++) {
-            std::string word = gguf_get_arr_str(ctx, token_idx, i);
-
-            vocab.token_to_id[word] = i;
-
-            auto & token_data = vocab.id_to_token[i];
-            token_data.text  = std::move(word);
-            token_data.score = scores[i];
-            token_data.type  = (llama_token_type) toktypes[i];
-
-            // determine the newline token: 0x0A == 10 == '\n'
-            if (token_data.text == "<0x0A>") {
-                vocab.linefeed_id = i;
-            }
-        }
-    }
-
-    {
-        // hparams
-        LLAMA_LOG_INFO("%s: format       = %s\n",     __func__, llama_file_version_name(ml->fver));
-        LLAMA_LOG_INFO("%s: arch         = %s\n",     __func__, general_arch.c_str());
-        LLAMA_LOG_INFO("%s: vocab type   = %s\n",     __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix
-        LLAMA_LOG_INFO("%s: n_vocab      = %u\n",     __func__, hparams.n_vocab);
-        LLAMA_LOG_INFO("%s: n_ctx_train  = %u\n",     __func__, hparams.n_ctx_train);
-        LLAMA_LOG_INFO("%s: n_ctx        = %u\n",     __func__, hparams.n_ctx);
-        LLAMA_LOG_INFO("%s: n_embd       = %u\n",     __func__, hparams.n_embd);
-        LLAMA_LOG_INFO("%s: n_head       = %u\n",     __func__, hparams.n_head);
-        LLAMA_LOG_INFO("%s: n_head_kv    = %u\n",     __func__, hparams.n_head_kv);
-        LLAMA_LOG_INFO("%s: n_layer      = %u\n",     __func__, hparams.n_layer);
-        LLAMA_LOG_INFO("%s: n_rot        = %u\n",     __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
-        LLAMA_LOG_INFO("%s: n_gqa        = %u\n",     __func__, hparams.n_gqa());
-        LLAMA_LOG_INFO("%s: f_norm_eps   = %.1e\n",   __func__, hparams.f_norm_rms_eps);
-        LLAMA_LOG_INFO("%s: n_ff         = %u\n",     __func__, hparams.n_ff);
-        LLAMA_LOG_INFO("%s: freq_base    = %.1f\n",   __func__, hparams.rope_freq_base);
-        LLAMA_LOG_INFO("%s: freq_scale   = %g\n",     __func__, hparams.rope_freq_scale);
-        LLAMA_LOG_INFO("%s: model type   = %s\n",     __func__, llama_model_type_name(model.type));
-        LLAMA_LOG_INFO("%s: model ftype  = %s\n",     __func__, llama_model_ftype_name(model.ftype).c_str());
-        LLAMA_LOG_INFO("%s: model size   = %.2f B\n", __func__, ml->n_elements*1e-9);
-
-        // general kv
-        LLAMA_LOG_INFO("%s: general.name = %s\n",    __func__, general_name.c_str());
-
-        // special tokens
-        if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); }
-        if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); }
-        if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); }
-        if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); }
-        if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); }
-        if (vocab.linefeed_id    != -1) { LLAMA_LOG_INFO( "%s: LF token  = %d '%s'\n", __func__, vocab.linefeed_id,    vocab.id_to_token[vocab.linefeed_id].text.c_str() );    }
-    }
-
-    if (vocab_only) {
-        LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
-        return;
-    }
-
-    auto & ctx = model.ctx;
+    model.n_gpu_layers = n_gpu_layers;
 
     size_t ctx_size;
     size_t mmapped_size;
 
-    ml->calc_sizes(ctx_size, mmapped_size);
+    ml.calc_sizes(ctx_size, mmapped_size);
 
     LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);
 
@@ -1480,7 +2258,7 @@ static void llama_model_load_internal(
         struct ggml_init_params params = {
             /*.mem_size   =*/ model.buf.size,
             /*.mem_buffer =*/ model.buf.data,
-            /*.no_alloc   =*/ ml->use_mmap,
+            /*.no_alloc   =*/ ml.use_mmap,
         };
 
         model.ctx = ggml_init(params);
@@ -1490,11 +2268,9 @@ static void llama_model_load_internal(
     }
 
     (void) main_gpu;
-    (void) mul_mat_q;
-#if defined(GGML_USE_CUBLAS)
-    LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__);
+#ifdef GGML_USE_CUBLAS
+    LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__);
     ggml_cuda_set_main_device(main_gpu);
-    ggml_cuda_set_mul_mat_q(mul_mat_q);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
 #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT
 #elif defined(GGML_USE_CLBLAST)
@@ -1509,93 +2285,381 @@ static void llama_model_load_internal(
     // prepare memory for the weights
     size_t vram_weights = 0;
     {
-        const uint32_t n_embd     = hparams.n_embd;
-        const uint32_t n_embd_gqa = hparams.n_embd_gqa();
-        const uint32_t n_layer    = hparams.n_layer;
-        const uint32_t n_vocab    = hparams.n_vocab;
+        const int64_t n_embd     = hparams.n_embd;
+        const int64_t n_embd_gqa = hparams.n_embd_gqa();
+        const int64_t n_layer    = hparams.n_layer;
+        const int64_t n_vocab    = hparams.n_vocab;
 
-        model.tok_embeddings = ml->create_tensor(ctx, TN_TOKEN_EMBD, {n_embd, n_vocab}, GGML_BACKEND_CPU);
+        const auto tn = LLM_TN(model.arch);
+        switch (model.arch) {
+            case LLM_ARCH_LLAMA:
+            case LLM_ARCH_REFACT:
+                {
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
 
-        // "output" tensor
-        {
-            ggml_backend backend_norm;
-            ggml_backend backend_output;
-            if (n_gpu_layers > int(n_layer)) { // NOLINT
-                // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
-                // on Windows however this is detrimental unless everything is on the GPU
+                    // output
+                    {
+                        ggml_backend backend_norm;
+                        ggml_backend backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
 #ifndef _WIN32
-                backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
 #else
-                backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
 #endif // _WIN32
 
-                backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
-            } else {
-                backend_norm = GGML_BACKEND_CPU;
-                backend_output = GGML_BACKEND_CPU;
-            }
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
 
-            model.norm   = ml->create_tensor(ctx, TN_OUTPUT_NORM, {n_embd},          backend_norm);
-            model.output = ml->create_tensor(ctx, TN_OUTPUT,      {n_embd, n_vocab}, backend_output);
-            if (backend_norm == GGML_BACKEND_GPU) {
-                vram_weights += ggml_nbytes(model.norm);
-            }
-            if (backend_output == GGML_BACKEND_GPU_SPLIT) {
-                vram_weights += ggml_nbytes(model.output);
-            }
-        }
+                        model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
 
-        const uint32_t n_ff = hparams.n_ff;
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
 
-        const int i_gpu_start = n_layer - n_gpu_layers;
+                    const uint32_t n_ff = hparams.n_ff;
 
-        model.layers.resize(n_layer);
-        for (uint32_t i = 0; i < n_layer; ++i) {
-            const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
-            const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+                    const int i_gpu_start = n_layer - n_gpu_layers;
 
-            auto & layer = model.layers[i];
-            layer.attention_norm = ml->create_tensor(ctx, format(TN_ATTN_NORM, i), {n_embd}, backend);
+                    model.layers.resize(n_layer);
 
-            layer.wq = ml->create_tensor(ctx, format(TN_ATTN_Q, i),      {n_embd, n_embd},     backend_split);
-            layer.wk = ml->create_tensor(ctx, format(TN_ATTN_K, i),      {n_embd, n_embd_gqa}, backend_split);
-            layer.wv = ml->create_tensor(ctx, format(TN_ATTN_V, i),      {n_embd, n_embd_gqa}, backend_split);
-            layer.wo = ml->create_tensor(ctx, format(TN_ATTN_OUTPUT, i), {n_embd, n_embd},     backend_split);
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
 
-            layer.ffn_norm = ml->create_tensor(ctx, format(TN_FFN_NORM, i), {n_embd}, backend);
+                        auto & layer = model.layers[i];
 
-            layer.w1 = ml->create_tensor(ctx, format(TN_FFN_GATE, i), {n_embd,   n_ff}, backend_split);
-            layer.w2 = ml->create_tensor(ctx, format(TN_FFN_DOWN, i), {  n_ff, n_embd}, backend_split);
-            layer.w3 = ml->create_tensor(ctx, format(TN_FFN_UP, i),   {n_embd,   n_ff}, backend_split);
+                        layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
 
-            if (backend == GGML_BACKEND_GPU) {
-                vram_weights +=
-                    ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)             +
-                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
-                    ggml_nbytes(layer.w1)             + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
-            }
+                        layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd},     backend_split);
+                        layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, backend_split);
+                        layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, backend_split);
+                        layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},     backend_split);
+
+                        layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+
+                        layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)       +
+                                ggml_nbytes(layer.wv)        + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
+                                ggml_nbytes(layer.w1)        + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
+                        }
+                    }
+                } break;
+            case LLM_ARCH_BAICHUAN:
+                {
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+                    {
+                        ggml_backend backend_norm;
+                        ggml_backend backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output      = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
+
+                        layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd},     backend_split);
+                        layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, backend_split);
+                        layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, backend_split);
+                        layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},     backend_split);
+
+                        layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+
+                        layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)       +
+                                ggml_nbytes(layer.wv)        + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) +
+                                ggml_nbytes(layer.w1)        + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
+                        }
+                    }
+                } break;
+            case LLM_ARCH_FALCON:
+                {
+                    // TODO: CPU-only for now
+
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+
+                    // output
+                    {
+                        ggml_backend backend_norm;
+                        ggml_backend backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd},          backend_norm);
+                        model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                            vram_weights += ggml_nbytes(model.output_norm_b);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, backend);
+                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, backend);
+
+                        if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) {
+                            layer.attn_norm_2   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend);
+                            layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, backend);
+
+                            if (backend == GGML_BACKEND_GPU) {
+                                vram_weights += ggml_nbytes(layer.attn_norm_2);
+                                vram_weights += ggml_nbytes(layer.attn_norm_2_b);
+                            }
+                        }
+
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},                backend_split);
+
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, backend_split);
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) +
+                                ggml_nbytes(layer.wqkv)      + ggml_nbytes(layer.wo)          +
+                                ggml_nbytes(layer.w2)        + ggml_nbytes(layer.w3);
+                        }
+                    }
+                } break;
+            case LLM_ARCH_STARCODER:
+                {
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+                    model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU);
+
+                    // output
+                    {
+                        ggml_backend backend_norm;
+                        ggml_backend backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd},          backend_norm);
+                        model.output        = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                            vram_weights += ggml_nbytes(model.output_norm_b);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+
+                    model.layers.resize(n_layer);
+
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend backend       = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT
+                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, backend);
+                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, backend);
+
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend_split);
+
+                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},   backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},           backend_split);
+
+                        layer.ffn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+                        layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, backend);
+
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
+                        layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend_split);
+
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend_split);
+
+                        if (backend == GGML_BACKEND_GPU) {
+                            vram_weights +=
+                                ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) +
+                                ggml_nbytes(layer.wqkv)      + ggml_nbytes(layer.bqkv)        +
+                                ggml_nbytes(layer.wo)        + ggml_nbytes(layer.bo)          +
+                                ggml_nbytes(layer.ffn_norm)  + ggml_nbytes(layer.ffn_norm_b)  +
+                                ggml_nbytes(layer.w2)        + ggml_nbytes(layer.b2)          +
+                                ggml_nbytes(layer.w3)        + ggml_nbytes(layer.b3);
+                        }
+                    }
+                } break;
+            case LLM_ARCH_PERSIMMON:
+                {
+                    model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"),  {n_embd, n_vocab}, GGML_BACKEND_CPU);
+
+                    {
+                        ggml_backend backend_norm;
+                        ggml_backend backend_output;
+
+                        if (n_gpu_layers > int(n_layer)) {
+                            // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                            // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                            backend_norm = LLAMA_BACKEND_OFFLOAD;
+#else
+                            backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
+                            backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        } else {
+                            backend_norm   = GGML_BACKEND_CPU;
+                            backend_output = GGML_BACKEND_CPU;
+                        }
+
+                        model.output_norm    = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd},          backend_norm);
+                        model.output_norm_b  = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd},          backend_norm);
+                        model.output         = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, backend_output);
+
+                        if (backend_norm == GGML_BACKEND_GPU) {
+                            vram_weights += ggml_nbytes(model.output_norm);
+                            vram_weights += ggml_nbytes(model.output_norm_b);
+                        }
+                        if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                            vram_weights += ggml_nbytes(model.output);
+                        }
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff;
+                    const int i_gpu_start = n_layer - n_gpu_layers;
+                    model.layers.resize(n_layer);
+                    for (uint32_t i = 0; i < n_layer; ++i) {
+                        const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+                        const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT;
+                        auto & layer = model.layers[i];
+                        layer.attn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
+                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, backend);
+                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend_split);
+                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},   backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},           backend_split);
+                        layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
+                        layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend_split);
+                        layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend_split);
+                        layer.ffn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+                        layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, backend);
+                        layer.attn_q_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i),   {64}, backend);
+                        layer.attn_k_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64}, backend);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i),   {64}, backend);
+                    }
+                } break;
+            default:
+                throw std::runtime_error("unknown architecture");
         }
     }
 
-    ml->done_getting_tensors();
+    ml.done_getting_tensors();
 
     // print memory requirements
     {
-        const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1;
-
         // this is the total memory required to run the inference
         size_t mem_required =
             ctx_size +
             mmapped_size - vram_weights; // weights in VRAM not in memory
 
-        // this is the memory required by one llama_state
-        const size_t mem_required_state =
-            scale*hparams.kv_size();
-
-        LLAMA_LOG_INFO("%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
-                mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
-
-        (void) n_batch;
+        LLAMA_LOG_INFO("%s: mem required  = %7.2f MB\n", __func__, mem_required / 1024.0 / 1024.0);
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
@@ -1604,61 +2668,42 @@ static void llama_model_load_internal(
         if (n_gpu_layers > (int) hparams.n_layer) {
             LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__);
         }
-        size_t vram_kv_cache = 0;
 
 #ifdef GGML_USE_CUBLAS
         const int max_backend_supported_layers = hparams.n_layer + 3;
-        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
-        if (n_gpu_layers > (int) hparams.n_layer + 1) {
-            if (low_vram) {
-                LLAMA_LOG_INFO("%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__);
-                vram_kv_cache += hparams.kv_size() / 2;
-            }
-        }
-        if (n_gpu_layers > (int) hparams.n_layer + 2) {
-            if (low_vram) {
-                LLAMA_LOG_WARN("%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
-            } else {
-                LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__);
-                vram_kv_cache += hparams.kv_size() / 2;
-            }
-        }
+        const int max_offloadable_layers = hparams.n_layer + 3;
 #elif defined(GGML_USE_CLBLAST)
         const int max_backend_supported_layers = hparams.n_layer + 1;
         const int max_offloadable_layers = hparams.n_layer + 1;
 #endif // GGML_USE_CUBLAS
 
-        LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n",
-                __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
-        LLAMA_LOG_INFO("%s: VRAM used: %zu MB\n",
-                __func__, (vram_weights + vram_kv_cache + MB - 1) / MB); // round up
+        LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
+        LLAMA_LOG_INFO("%s: VRAM used: %.2f MB\n", __func__, vram_weights / 1024.0 / 1024.0);
 #else
         (void) n_gpu_layers;
 #endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
     }
 
     // populate `tensors_by_name`
-    for (int i = 0; i < ml->n_tensors; ++i) {
-        struct ggml_tensor * cur = ggml_get_tensor(ctx, ml->get_tensor_name(i));
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        struct ggml_tensor * cur = ggml_get_tensor(ctx, ml.get_tensor_name(i));
         model.tensors_by_name.emplace_back(ggml_get_name(cur), cur);
     }
 
     (void) tensor_split;
-#if defined(GGML_USE_CUBLAS)
+#ifdef GGML_USE_CUBLAS
     {
         ggml_cuda_set_tensor_split(tensor_split);
     }
 #endif
 
-    ml->load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
+    ml.load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL);
 
     if (progress_callback) {
         progress_callback(1.0f, progress_callback_user_data);
     }
 
-    model.mapping = std::move(ml->mapping);
+    model.mapping = std::move(ml.mapping);
 
     // loading time will be recalculate after the first eval, so
     // we take page faults deferred by mmap() into consideration
@@ -1668,46 +2713,52 @@ static void llama_model_load_internal(
 static bool llama_model_load(
         const std::string & fname,
         llama_model & model,
-        llama_vocab & vocab,
-        int n_ctx,
-        int n_batch,
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
-        const bool mul_mat_q,
-        float rope_freq_base,
-        float rope_freq_scale,
-        bool low_vram,
-        ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
         bool vocab_only,
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers,
-                                  main_gpu, tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, low_vram, memory_type,
-                                  use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
-        return true;
+        llama_model_loader ml(fname, use_mmap);
+
+        model.hparams.vocab_only = vocab_only;
+
+        llm_load_arch   (ml, model);
+        llm_load_hparams(ml, model);
+        llm_load_vocab  (ml, model);
+
+        llm_load_print_meta(ml, model);
+
+        if (model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+            throw std::runtime_error("vocab size mismatch");
+        }
+
+        if (vocab_only) {
+            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
+            return true;
+        }
+
+        llm_load_tensors(
+                ml, model, n_gpu_layers,
+                main_gpu, tensor_split,
+                use_mlock, progress_callback, progress_callback_user_data);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
         return false;
     }
+
+    return true;
 }
 
-static struct ggml_cgraph * llama_build_graph(
-         llama_context & lctx,
-     const llama_token * tokens,
-           const float * embd,
-                   int   n_tokens,
-                   int   n_past) {
-
-    GGML_ASSERT((!tokens && embd) || (tokens && !embd)); // NOLINT
-
-    const int N = n_tokens;
-
+static struct ggml_cgraph * llm_build_llama(
+    llama_context & lctx,
+    const llama_batch & batch) {
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
 
     const auto & kv_self = lctx.kv_self;
 
@@ -1715,7 +2766,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     const int64_t n_embd      = hparams.n_embd;
     const int64_t n_layer     = hparams.n_layer;
-    const int64_t n_ctx       = hparams.n_ctx;
+    const int64_t n_ctx       = cparams.n_ctx;
     const int64_t n_head      = hparams.n_head;
     const int64_t n_head_kv   = hparams.n_head_kv;
     const int64_t n_embd_head = hparams.n_embd_head();
@@ -1723,23 +2774,28 @@ static struct ggml_cgraph * llama_build_graph(
 
     GGML_ASSERT(n_embd_head == hparams.n_rot);
 
-    const float freq_base    = hparams.rope_freq_base;
-    const float freq_scale   = hparams.rope_freq_scale;
+    const float freq_base    = cparams.rope_freq_base;
+    const float freq_scale   = cparams.rope_freq_scale;
     const float norm_rms_eps = hparams.f_norm_rms_eps;
 
     const int n_gpu_layers = model.n_gpu_layers;
 
-    auto & mem_per_token = lctx.mem_per_token;
-    auto & buf_compute   = lctx.buf_compute;
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift;
+
+    //printf("n_kv = %d\n", n_kv);
+
+    auto & buf_compute = lctx.buf_compute;
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_compute.size,
         /*.mem_buffer =*/ buf_compute.data,
-        /*.no_alloc   =*/ false,
+        /*.no_alloc   =*/ true,
     };
 
-    params.no_alloc = true;
-
     struct ggml_context * ctx0 = ggml_init(params);
 
     ggml_cgraph * gf = ggml_new_graph(ctx0);
@@ -1747,12 +2803,12 @@ static struct ggml_cgraph * llama_build_graph(
     struct ggml_tensor * cur;
     struct ggml_tensor * inpL;
 
-    if (tokens) {
-        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
 
         ggml_allocr_alloc(lctx.alloc, inp_tokens);
         if (!ggml_allocr_is_measure(lctx.alloc)) {
-            memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens));
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
         }
         ggml_set_name(inp_tokens, "inp_tokens");
 
@@ -1762,11 +2818,11 @@ static struct ggml_cgraph * llama_build_graph(
         GGML_ASSERT(false && "not implemented");
 #endif
 
-        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N);
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
 
         ggml_allocr_alloc(lctx.alloc, inpL);
         if (!ggml_allocr_is_measure(lctx.alloc)) {
-            memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL));
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
         }
     }
 
@@ -1775,9 +2831,6 @@ static struct ggml_cgraph * llama_build_graph(
 
     // offload functions set the tensor output backend to GPU
     // tensors are GPU-accelerated if any input or the output has been offloaded
-    //
-    // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal
-    // in that case ggml_cuda_assign_buffers has no effect
     offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
     offload_func_t offload_func_kq = llama_nop;
     offload_func_t offload_func_v  = llama_nop;
@@ -1794,12 +2847,75 @@ static struct ggml_cgraph * llama_build_graph(
     }
 #endif // GGML_USE_CUBLAS
 
+    // KQ_scale
     struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
     ggml_allocr_alloc(lctx.alloc, KQ_scale);
     if (!ggml_allocr_is_measure(lctx.alloc)) {
-        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head)));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    offload_func_kq(KQ_pos);
+    ggml_set_name(KQ_pos, "KQ_pos");
+    ggml_allocr_alloc(lctx.alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < n_tokens; ++i) {
+            data[i] = batch.pos[i];
+        }
+    }
+
+    // shift the entire K-cache if needed
+    if (do_rope_shift) {
+        struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
+        offload_func_kq(K_shift);
+        ggml_set_name(K_shift, "K_shift");
+        ggml_allocr_alloc(lctx.alloc, K_shift);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            int * data = (int *) K_shift->data;
+            for (int i = 0; i < n_ctx; ++i) {
+                data[i] = kv_self.cells[i].delta;
+            }
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * tmp =
+                    ggml_rope_custom_inplace(ctx0,
+                        ggml_view_3d(ctx0, kv_self.k,
+                            n_embd_head, n_head_kv, n_ctx,
+                            ggml_element_size(kv_self.k)*n_embd_head,
+                            ggml_element_size(kv_self.k)*n_embd_gqa,
+                            ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il),
+                        K_shift, n_embd_head, 0, 0, freq_base, freq_scale);
+            offload_func_kq(tmp);
+            ggml_build_forward_expand(gf, tmp);
+        }
     }
-    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
 
     for (int il = 0; il < n_layer; ++il) {
         ggml_format_name(inpL, "layer_inp_%d", il);
@@ -1820,8 +2936,8 @@ static struct ggml_cgraph * llama_build_graph(
             offload_func(cur);
             ggml_set_name(cur, "rms_norm_0");
 
-            // cur = cur*attention_norm(broadcasted)
-            cur = ggml_mul(ctx0, cur, model.layers[il].attention_norm);
+            // cur = cur*attn_norm(broadcasted)
+            cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm);
             offload_func(cur);
             ggml_set_name(cur, "attention_norm_0");
         }
@@ -1837,33 +2953,33 @@ static struct ggml_cgraph * llama_build_graph(
             offload_func_kq(tmpq);
             ggml_set_name(tmpq, "tmpq");
 
-            struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), n_past, n_embd_head, 0, 0, freq_base, freq_scale);
+            struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
             offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
-            struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N),    n_past, n_embd_head, 0, 0, freq_base, freq_scale);
+            struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
             offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
             // store key and value to memory
             {
-                // compute the transposed [N, n_embd] V matrix
+                // compute the transposed [n_tokens, n_embd] V matrix
 
                 struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 offload_func_v(tmpv);
                 ggml_set_name(tmpv, "tmpv");
 
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, N));
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens));
                 offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
-                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past));
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
                 offload_func_kq(k);
                 ggml_set_name(k, "k");
 
-                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa,
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
                         (   n_ctx)*ggml_element_size(kv_self.v),
-                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v));
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
                 offload_func_v(v);
                 ggml_set_name(v, "v");
 
@@ -1872,16 +2988,13 @@ static struct ggml_cgraph * llama_build_graph(
                 ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
             }
 
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        Qcur,
-                        0, 2, 1, 3);
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
             offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
                 ggml_view_3d(ctx0, kv_self.k,
-                        n_embd_head, n_past + N, n_head_kv,
+                        n_embd_head, n_kv, n_head_kv,
                         ggml_element_size(kv_self.k)*n_embd_gqa,
                         ggml_element_size(kv_self.k)*n_embd_head,
                         ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
@@ -1894,25 +3007,25 @@ static struct ggml_cgraph * llama_build_graph(
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd_head)
-            // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            // KQ_scaled shape [n_kv, n_tokens, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
             offload_func_kq(KQ_scaled);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask);
             offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
             offload_func_v(KQ_soft_max);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
             // split cached V into n_head heads
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
-                        n_past + N, n_embd_head, n_head_kv,
+                        n_kv, n_embd_head, n_head_kv,
                         ggml_element_size(kv_self.v)*n_ctx,
                         ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
                         ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
@@ -1927,7 +3040,7 @@ static struct ggml_cgraph * llama_build_graph(
             // make V contiguous in memory to speed up the matmul, however we waste time on the copy
             // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation
             // is there a better way?
-            struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head));
+            struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head));
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
 #endif
 
@@ -1936,10 +3049,8 @@ static struct ggml_cgraph * llama_build_graph(
             offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
             offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
 
@@ -2005,14 +3116,16 @@ static struct ggml_cgraph * llama_build_graph(
         inpL = cur;
     }
 
+    cur = inpL;
+
     // norm
     {
-        cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps);
+        cur = ggml_rms_norm(ctx0, cur, norm_rms_eps);
         offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_2");
 
         // cur = cur*norm(broadcasted)
-        cur = ggml_mul(ctx0, cur, model.norm);
+        cur = ggml_mul(ctx0, cur, model.output_norm);
         // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
         ggml_set_name(cur, "result_norm");
     }
@@ -2021,68 +3134,1896 @@ static struct ggml_cgraph * llama_build_graph(
     cur = ggml_mul_mat(ctx0, model.output, cur);
     ggml_set_name(cur, "result_output");
 
-    // logits -> probs
-    //cur = ggml_soft_max_inplace(ctx0, cur);
-
     ggml_build_forward_expand(gf, cur);
 
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-
     ggml_free(ctx0);
 
     return gf;
 }
 
-// evaluate the transformer
-//
-//   - lctx:      llama context
-//   - tokens:    new batch of tokens to process
-//   - embd       embeddings input
-//   - n_tokens   number of tokens
-//   - n_past:    the context size so far
-//   - n_threads: number of threads to use
-//
-static bool llama_eval_internal(
+static struct ggml_cgraph * llm_build_baichaun(
          llama_context & lctx,
-     const llama_token * tokens,
-           const float * embd,
-                   int   n_tokens,
-                   int   n_past,
-                   int   n_threads,
-            const char * cgraph_fname) {
-
-    GGML_ASSERT((!tokens && embd) || (tokens && !embd)); // NOLINT
-
-    GGML_ASSERT(n_tokens > 0);
-    GGML_ASSERT(n_past >= 0);
-    GGML_ASSERT(n_threads > 0);
-    // TODO: keep the values of n_batch and n_ctx
-    // GGML_ASSERT(n_tokens <= n_batch);
-    // GGML_ASSERT(n_past + n_tokens <= n_ctx);
-
-    const int64_t t_start_us = ggml_time_us();
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
-#endif
-
-    const int N = n_tokens;
-
+     const llama_batch & batch) {
     const auto & model   = lctx.model;
     const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
 
     const auto & kv_self = lctx.kv_self;
 
     GGML_ASSERT(!!kv_self.ctx);
 
     const int64_t n_embd      = hparams.n_embd;
-    const int64_t n_vocab     = hparams.n_vocab;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    const float freq_base    = cparams.rope_freq_base;
+    const float freq_scale   = cparams.rope_freq_scale;
+    const float norm_rms_eps = hparams.f_norm_rms_eps;
+
+    const int n_gpu_layers = model.n_gpu_layers;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift;
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
+        }
+    }
+
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
+    }
+#endif // GGML_USE_CUBLAS
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    offload_func_kq(KQ_pos);
+    ggml_set_name(KQ_pos, "KQ_pos");
+    ggml_allocr_alloc(lctx.alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < n_tokens; ++i) {
+            data[i] = batch.pos[i];
+        }
+    }
+
+    // shift the entire K-cache if needed
+    if (do_rope_shift) {
+        struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
+        offload_func_kq(K_shift);
+        ggml_set_name(K_shift, "K_shift");
+        ggml_allocr_alloc(lctx.alloc, K_shift);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            int * data = (int *) K_shift->data;
+            for (int i = 0; i < n_ctx; ++i) {
+                data[i] = kv_self.cells[i].delta;
+            }
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * tmp =
+                    ggml_rope_custom_inplace(ctx0,
+                        ggml_view_3d(ctx0, kv_self.k,
+                            n_embd_head, n_head_kv, n_ctx,
+                            ggml_element_size(kv_self.k)*n_embd_head,
+                            ggml_element_size(kv_self.k)*n_embd_gqa,
+                            ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il),
+                        K_shift, n_embd_head, 0, 0, freq_base, freq_scale);
+            offload_func_kq(tmp);
+            ggml_build_forward_expand(gf, tmp);
+        }
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_format_name(inpL, "layer_inp_%d", il);
+
+        offload_func_t offload_func = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (il >= i_gpu_start) {
+            offload_func = ggml_cuda_assign_buffers_no_alloc;
+        }
+#endif // GGML_USE_CUBLAS
+
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps);
+            offload_func(cur);
+            ggml_set_name(cur, "rms_norm_0");
+
+            // cur = cur*attn_norm(broadcasted)
+            cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm);
+            offload_func(cur);
+            ggml_set_name(cur, "attention_norm_0");
+        }
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+            offload_func_kq(tmpk);
+            ggml_set_name(tmpk, "tmpk");
+
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
+            struct ggml_tensor * Kcur;
+            struct ggml_tensor * Qcur;
+            switch (model.type) {
+                case MODEL_7B:
+                    Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
+                    Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens),    KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale);
+                    break;
+                case MODEL_13B:
+                    Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, n_tokens);
+                    Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, n_tokens);
+                    break;
+                default:
+                    GGML_ASSERT(false);
+            }
+
+            offload_func_kq(Kcur);
+            ggml_set_name(Kcur, "Kcur");
+
+            offload_func_kq(Qcur);
+            ggml_set_name(Qcur, "Qcur");
+
+            // store key and value to memory
+            {
+                // compute the transposed [n_tokens, n_embd] V matrix
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens));
+                offload_func_v(Vcur);
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                offload_func_kq(k);
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+                ggml_set_name(v, "v");
+
+                // important: storing RoPE-ed version of K in the KV cache!
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            offload_func_kq(Q);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            offload_func_kq(K);
+            ggml_set_name(K, "K");
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
+            ggml_set_name(KQ, "KQ");
+
+            // KQ_scaled = KQ / sqrt(n_embd_head)
+            // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            struct ggml_tensor * KQ_masked;
+            struct ggml_tensor * KQ_scaled_alibi;
+
+            switch (model.type) {
+                case MODEL_7B:
+                    KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask);
+                    break;
+                case MODEL_13B:
+                    // TODO: replace with ggml_add()
+                    KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8);
+                    ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");
+                    KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask);
+                    break;
+                default:
+                    GGML_ASSERT(false);
+            }
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            // split cached V into n_head heads
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            offload_func_v(V);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
+            ggml_set_name(KQV, "KQV");
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].wo,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_wo");
+        }
+
+        struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
+        offload_func(inpFF);
+        ggml_set_name(inpFF, "inpFF");
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps);
+                offload_func(cur);
+                ggml_set_name(cur, "rms_norm_1");
+
+                // cur = cur*ffn_norm(broadcasted)
+                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm);
+                offload_func(cur);
+                ggml_set_name(cur, "ffn_norm");
+            }
+
+            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
+                    model.layers[il].w3,
+                    cur);
+            offload_func(tmp);
+            ggml_set_name(tmp, "result_w3");
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w1,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_w1");
+
+            // SILU activation
+            cur = ggml_silu(ctx0, cur);
+            offload_func(cur);
+            ggml_set_name(cur, "silu");
+
+            cur = ggml_mul(ctx0, cur, tmp);
+            offload_func(cur);
+            ggml_set_name(cur, "silu_x_result_w3");
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w2,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_w2");
+        }
+
+        cur = ggml_add(ctx0, cur, inpFF);
+        offload_func(cur);
+        ggml_set_name(cur, "inpFF_+_result_w2");
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    {
+        cur = ggml_rms_norm(ctx0, cur, norm_rms_eps);
+        offload_func_nr(cur);
+        ggml_set_name(cur, "rms_norm_2");
+
+        // cur = cur*norm(broadcasted)
+        cur = ggml_mul(ctx0, cur, model.output_norm);
+        // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
+        ggml_set_name(cur, "result_norm");
+    }
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_cgraph * llm_build_refact(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    const float norm_rms_eps = hparams.f_norm_rms_eps;
+
+    const int n_gpu_layers = model.n_gpu_layers;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    // printf("n_kv = %d\n", n_kv);
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
+        }
+    }
+
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
+    }
+#endif // GGML_USE_CUBLAS
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head)));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_format_name(inpL, "layer_inp_%d", il);
+
+        offload_func_t offload_func = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (il >= i_gpu_start) {
+            offload_func = ggml_cuda_assign_buffers_no_alloc;
+        }
+#endif // GGML_USE_CUBLAS
+
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps);
+            offload_func(cur);
+            ggml_set_name(cur, "rms_norm_0");
+
+            // cur = cur*attn_norm(broadcasted)
+            cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm);
+            offload_func(cur);
+            ggml_set_name(cur, "attention_norm_0");
+        }
+
+        // self-attention
+        {
+            // compute Q and K
+            struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+            offload_func_kq(tmpk);
+            ggml_set_name(tmpk, "tmpk");
+
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
+            struct ggml_tensor * Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens);
+            offload_func_kq(Kcur);
+            ggml_set_name(Kcur, "Kcur");
+
+            struct ggml_tensor * Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens);
+            offload_func_kq(Qcur);
+            ggml_set_name(Qcur, "Qcur");
+
+            // store key and value to memory
+            {
+                // compute the transposed [n_tokens, n_embd] V matrix
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens));
+                offload_func_v(Vcur);
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                offload_func_kq(k);
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+                ggml_set_name(v, "v");
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            offload_func_kq(Q);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            offload_func_kq(K);
+            ggml_set_name(K, "K");
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
+            ggml_set_name(KQ, "KQ");
+
+            // KQ_scaled = KQ / sqrt(n_embd_head)
+            // KQ_scaled shape [n_kv, n_tokens, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8);
+            ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");
+
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask);
+            offload_func_kq(KQ_masked);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            // split cached V into n_head heads
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            offload_func_v(V);
+            ggml_set_name(V, "V");
+
+#if 1
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
+            ggml_set_name(KQV, "KQV");
+#else
+            // make V contiguous in memory to speed up the matmul, however we waste time on the copy
+            // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation
+            // is there a better way?
+            struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head));
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
+#endif
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].wo,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_wo");
+        }
+
+        struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
+        offload_func(inpFF);
+        ggml_set_name(inpFF, "inpFF");
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps);
+                offload_func(cur);
+                ggml_set_name(cur, "rms_norm_1");
+
+                // cur = cur*ffn_norm(broadcasted)
+                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm);
+                offload_func(cur);
+                ggml_set_name(cur, "ffn_norm");
+            }
+
+            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
+                    model.layers[il].w3,
+                    cur);
+            offload_func(tmp);
+            ggml_set_name(tmp, "result_w3");
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w1,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_w1");
+
+            // SILU activation
+            cur = ggml_silu(ctx0, cur);
+            offload_func(cur);
+            ggml_set_name(cur, "silu");
+
+            cur = ggml_mul(ctx0, cur, tmp);
+            offload_func(cur);
+            ggml_set_name(cur, "silu_x_result_w3");
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w2,
+                    cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_w2");
+        }
+
+        cur = ggml_add(ctx0, cur, inpFF);
+        offload_func(cur);
+        ggml_set_name(cur, "inpFF_+_result_w2");
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    {
+        cur = ggml_rms_norm(ctx0, cur, norm_rms_eps);
+        offload_func_nr(cur);
+        ggml_set_name(cur, "rms_norm_2");
+
+        // cur = cur*norm(broadcasted)
+        cur = ggml_mul(ctx0, cur, model.output_norm);
+        // offload_func_nr(cur); // TODO CPU + GPU mirrored backend
+        ggml_set_name(cur, "result_norm");
+    }
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_cgraph * llm_build_falcon(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    const float freq_base  = cparams.rope_freq_base;
+    const float freq_scale = cparams.rope_freq_scale;
+    const float norm_eps   = hparams.f_norm_eps;
+
+    const int n_gpu_layers = model.n_gpu_layers;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift;
+
+    //printf("kv_head = %d, n_kv = %d, n_tokens = %d, n_ctx = %d, is_measure = %d, has_shift = %d\n",
+    //        kv_head, n_kv, n_tokens, n_ctx, ggml_allocr_is_measure(lctx.alloc), kv_self.has_shift);
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
+        }
+    }
+
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer) {
+        offload_func_nr = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 1) {
+        offload_func_v  = ggml_cuda_assign_buffers_no_alloc;
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
+    }
+#endif // GGML_USE_CUBLAS
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    // KQ_pos - contains the positions
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    offload_func_kq(KQ_pos);
+    ggml_set_name(KQ_pos, "KQ_pos");
+    ggml_allocr_alloc(lctx.alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < n_tokens; ++i) {
+            data[i] = batch.pos[i];
+        }
+    }
+
+    // shift the entire K-cache if needed
+    if (do_rope_shift) {
+        struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
+        offload_func_kq(K_shift);
+        ggml_set_name(K_shift, "K_shift");
+        ggml_allocr_alloc(lctx.alloc, K_shift);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            int * data = (int *) K_shift->data;
+            for (int i = 0; i < n_ctx; ++i) {
+                data[i] = kv_self.cells[i].delta;
+            }
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * tmp =
+                    ggml_rope_custom_inplace(ctx0,
+                        ggml_view_3d(ctx0, kv_self.k,
+                            n_embd_head, n_head_kv, n_ctx,
+                            ggml_element_size(kv_self.k)*n_embd_head,
+                            ggml_element_size(kv_self.k)*n_embd_gqa,
+                            ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il),
+                        K_shift, n_embd_head, 2, 0, freq_base, freq_scale);
+            offload_func_kq(tmp);
+            ggml_build_forward_expand(gf, tmp);
+        }
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * attn_norm;
+
+        offload_func_t offload_func = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (il >= i_gpu_start) {
+            offload_func = ggml_cuda_assign_buffers_no_alloc;
+        }
+#endif // GGML_USE_CUBLAS
+
+        // self-attention
+        // TODO: refactor into common function (shared with LLaMA)
+        {
+            attn_norm = ggml_norm(ctx0, inpL, norm_eps);
+            offload_func(attn_norm);
+
+            attn_norm = ggml_add(ctx0,
+                    ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm),
+                    model.layers[il].attn_norm_b);
+            offload_func(attn_norm->src[0]);
+            offload_func(attn_norm);
+
+            if (model.layers[il].attn_norm_2) { // Falcon-40B
+                cur = ggml_norm(ctx0, inpL, norm_eps);
+                offload_func(cur);
+
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0, cur, model.layers[il].attn_norm_2),
+                        model.layers[il].attn_norm_2_b);
+                offload_func(cur->src[0]);
+                offload_func(cur);
+            } else { // Falcon 7B
+                cur = attn_norm;
+            }
+
+            // compute QKV
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+            offload_func_kq(cur);
+
+            // Note that the strides for Kcur, Vcur are set up so that the
+            // resulting views are misaligned with the tensor's storage
+            // (by applying the K/V offset we shift the tensor's original
+            // view to stick out behind the viewed QKV tensor's allocated
+            // memory, so to say). This is ok because no actual accesses
+            // happen to that out-of-range memory, but it can require some
+            // trickery when trying to accurately dump these views for
+            // debugging.
+
+            const size_t wsize = ggml_type_size(cur->type);
+
+            // TODO: these 2 ggml_conts are technically not needed, but we add them until CUDA support for
+            //       non-contiguous views is added for the rope operator
+            struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                0));
+            offload_func_kq(tmpq);
+
+            struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head_kv, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                wsize * n_embd_head *  n_head));
+            offload_func_kq(tmpk);
+
+            struct ggml_tensor * tmpv = ggml_view_3d(
+                ctx0, cur, n_embd_head, n_head_kv, n_tokens,
+                wsize * n_embd_head,
+                wsize * n_embd_head * (n_head + 2 * n_head_kv),
+                wsize * n_embd_head * (n_head +     n_head_kv));
+            offload_func_v(tmpv);
+
+            // using mode = 2 for neox mode
+            struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, tmpq, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale);
+            offload_func_kq(Qcur);
+            struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, tmpk, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale);
+            offload_func_kq(Kcur);
+
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens));
+                offload_func_v(Vcur);
+                offload_func_v(Vcur->src[0]->src[0]);
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                offload_func_kq(k);
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            offload_func_kq(Q);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            offload_func_kq(K);
+            ggml_set_name(K, "K");
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
+            ggml_set_name(KQ, "KQ");
+
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask);
+            offload_func_kq(KQ_masked);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            offload_func_v(V);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
+            ggml_set_name(KQV, "KQV");
+
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
+            offload_func(cur);
+            ggml_set_name(cur, "result_wo");
+        }
+
+        struct ggml_tensor * attn_out = cur;
+
+        // feed forward
+        {
+            struct ggml_tensor * inpFF = attn_norm;
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF);
+            offload_func(cur);
+
+            cur = ggml_gelu(ctx0, cur);
+            offload_func(cur);
+            cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur);
+            offload_func(cur);
+        }
+
+        cur = ggml_add(ctx0, cur, attn_out);
+        offload_func(cur);
+        cur = ggml_add(ctx0, cur, inpL);
+        offload_func(cur);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    {
+        cur = ggml_norm(ctx0, cur, norm_eps);
+        offload_func_nr(cur);
+
+        cur = ggml_add(ctx0,
+                ggml_mul(ctx0, cur, model.output_norm),
+                model.output_norm_b);
+        ggml_set_name(cur, "result_norm");
+    }
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+static struct ggml_cgraph * llm_build_starcoder(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    const float norm_eps = hparams.f_norm_eps;
+
+    const int32_t n_tokens = batch.n_tokens;
+    const int32_t n_kv     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head  = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    auto & buf_compute = lctx.buf_compute;
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * token;
+    struct ggml_tensor * position;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+
+        token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+#ifdef GGML_USE_MPI
+        GGML_ASSERT(false && "not implemented");
+#endif
+
+        token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, token);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token));
+        }
+    }
+
+    {
+        // Compute position embeddings.
+        struct ggml_tensor * inp_positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        ggml_allocr_alloc(lctx.alloc, inp_positions);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            for (int i = 0; i < n_tokens; ++i) {
+                ((int32_t *) inp_positions->data)[i] = batch.pos[i];
+            }
+        }
+        ggml_set_name(inp_positions, "inp_positions");
+
+        position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions);
+    }
+
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    inpL = ggml_add(ctx0, token, position);
+    ggml_set_name(inpL, "inpL");
+
+    for (int il = 0; il < n_layer; ++il) {
+        {
+            // Norm
+            cur = ggml_norm(ctx0, inpL, norm_eps);
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b);
+        }
+
+        {
+            // Self Attention
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv);
+
+            struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd);
+            struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa));
+
+            struct ggml_tensor * Qcur = tmpq;
+            struct ggml_tensor * Kcur = tmpk;
+
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens));
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head));
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)),
+                        0, 2, 1, 3);
+            ggml_set_name(Q, "Q");
+
+            struct ggml_tensor * K =
+                ggml_view_3d(ctx0, kv_self.k,
+                        n_embd_head, n_kv, n_head_kv,
+                        ggml_element_size(kv_self.k)*n_embd_gqa,
+                        ggml_element_size(kv_self.k)*n_embd_head,
+                        ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+            ggml_set_name(K, "K");
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            ggml_set_name(KQ, "KQ");
+
+            // KQ_scaled = KQ / sqrt(n_embd_head)
+            // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            // split cached V into n_head heads
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            ggml_set_name(KQV, "KQV");
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            // cur = KQV_merged.contiguous().view(n_embd, n_tokens)
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+        }
+
+        // Projection
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo);
+
+        // Add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        // FF
+        {
+            // Norm
+            {
+                cur = ggml_norm(ctx0, inpFF, norm_eps);
+                cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b);
+            }
+
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // Projection
+            cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2);
+        }
+
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    // Output Norm
+    {
+        cur = ggml_norm(ctx0, inpL, norm_eps);
+        cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b);
+    }
+    ggml_set_name(cur, "result_norm");
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+
+    ggml_build_forward_expand(gf, cur);
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+
+static struct ggml_cgraph * llm_build_persimmon(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model = lctx.model;
+    const auto & hparams = model.hparams;
+
+    const auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const auto & cparams = lctx.cparams;
+    const int64_t n_embd      = hparams.n_embd;
+    const int64_t n_layer     = hparams.n_layer;
+    const int64_t n_ctx       = cparams.n_ctx;
+    const int64_t n_head_kv   = hparams.n_head_kv;
+    const int64_t n_head      = hparams.n_head;
+    const int64_t n_embd_head = hparams.n_embd_head();
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+    const size_t n_rot        = n_embd_head / 2;
+
+    const float freq_base  = cparams.rope_freq_base;
+    const float freq_scale = cparams.rope_freq_scale;
+    const float norm_eps = hparams.f_norm_eps;
+
+    const int n_gpu_layers = model.n_gpu_layers;
+
+
+    const int32_t n_tokens    = batch.n_tokens;
+    const int32_t n_kv        = ggml_allocr_is_measure(lctx.alloc) ? n_ctx            : kv_self.n;
+    const int32_t kv_head     = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head;
+
+    const bool do_rope_shift  = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift;
+
+    auto & buf_compute = lctx.buf_compute;
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute.size,
+        /*.mem_buffer =*/ buf_compute.data,
+        /*.no_alloc   =*/ true,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * cur;
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+
+        ggml_allocr_alloc(lctx.alloc, inp_tokens);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens));
+        }
+        ggml_set_name(inp_tokens, "inp_tokens");
+        inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens);
+    } else {
+        inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
+        ggml_allocr_alloc(lctx.alloc, inpL);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL));
+        }
+    }
+    const int i_gpu_start = n_layer - n_gpu_layers;
+    (void) i_gpu_start;
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+    // KQ_scale
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_allocr_alloc(lctx.alloc, KQ_scale);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head)));
+    }
+    ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+    offload_func_kq(KQ_mask);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_allocr_alloc(lctx.alloc, KQ_mask);
+
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        float * data = (float *) KQ_mask->data;
+        memset(data, 0, ggml_nbytes(KQ_mask));
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                const llama_pos    pos    = batch.pos[j];
+                const llama_seq_id seq_id = batch.seq_id[j];
+                for (int i = 0; i < n_kv; ++i) {
+                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    offload_func_kq(KQ_pos);
+    ggml_set_name(KQ_pos, "KQ_pos");
+    ggml_allocr_alloc(lctx.alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(lctx.alloc)) {
+        int * data = (int *) KQ_pos->data;
+        for (int i = 0; i < n_tokens; ++i) {
+            data[i] = batch.pos[i];
+        }
+    }
+    if (do_rope_shift) {
+        struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
+        offload_func_kq(K_shift);
+        ggml_set_name(K_shift, "K_shift");
+        ggml_allocr_alloc(lctx.alloc, K_shift);
+        if (!ggml_allocr_is_measure(lctx.alloc)) {
+            int * data = (int *) K_shift->data;
+            for (int i = 0; i < n_ctx; ++i) {
+                data[i] = kv_self.cells[i].delta;
+            }
+        }
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * tmp =
+                    // we rotate only the first n_rot dimensions.
+                    ggml_rope_custom_inplace(ctx0,
+                        ggml_view_3d(ctx0, kv_self.k,
+                            n_rot, n_head, n_ctx,
+                            ggml_element_size(kv_self.k)*n_embd_gqa,
+                            ggml_element_size(kv_self.k)*n_embd_head,
+                            ggml_element_size(kv_self.k)*(n_embd_head*n_ctx*il)
+                        ),
+                        K_shift, n_rot, 2, 0, freq_base, freq_scale);
+            offload_func_kq(tmp);
+            ggml_build_forward_expand(gf, tmp);
+        }
+    }
+    for (int il=0; il < n_layer; ++il) {
+        struct ggml_tensor * residual = inpL;
+        offload_func_t offload_func = llama_nop;
+        {
+            cur = ggml_norm(ctx0, inpL, norm_eps);
+            offload_func(cur);
+            cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm);
+            offload_func(cur);
+            cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b);
+            offload_func(cur);
+            ggml_format_name(cur, "input_layernorm_%d", il);
+        }
+        // self attention
+        {
+            cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+            offload_func_kq(cur);
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            offload_func_kq(cur);
+
+            // split qkv
+            GGML_ASSERT(n_head_kv == n_head);
+            ggml_set_name(cur, format("qkv_%d", il).c_str());
+            struct ggml_tensor * tmpqkv = ggml_reshape_4d(ctx0, cur, n_embd_head, 3, n_head, n_tokens);
+            offload_func_kq(tmpqkv);
+            struct ggml_tensor * tmpqkv_perm = ggml_cont(ctx0, ggml_permute(ctx0, tmpqkv, 0, 3, 1, 2));
+            offload_func_kq(tmpqkv_perm);
+            ggml_format_name(tmpqkv_perm, "tmpqkv_perm_%d", il);
+            struct ggml_tensor * tmpq = ggml_view_3d(
+                    ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
+                    ggml_element_size(tmpqkv_perm) * n_embd_head,
+                    ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
+                    0
+                );
+            offload_func_kq(tmpq);
+            struct ggml_tensor * tmpk = ggml_view_3d(
+                    ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
+                    ggml_element_size(tmpqkv_perm) * n_embd_head,
+                    ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
+                    ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens
+                );
+            offload_func_kq(tmpk);
+            // Q/K Layernorm
+            tmpq = ggml_norm(ctx0, tmpq, norm_eps);
+            offload_func_kq(tmpq);
+            tmpq =  ggml_mul(ctx0, tmpq, model.layers[il].attn_q_norm);
+            offload_func_kq(tmpq);
+            tmpq =  ggml_add(ctx0, tmpq, model.layers[il].attn_q_norm_b);
+            offload_func_kq(tmpq);
+
+            tmpk = ggml_norm(ctx0, tmpk, norm_eps);
+            offload_func_v(tmpk);
+            tmpk =  ggml_mul(ctx0, tmpk, model.layers[il].attn_k_norm);
+            offload_func_v(tmpk);
+            tmpk =  ggml_add(ctx0, tmpk, model.layers[il].attn_k_norm_b);
+            offload_func_v(tmpk);
+
+            // RoPE the first n_rot of q/k, pass the other half, and concat.
+            struct ggml_tensor * qrot = ggml_view_3d(
+                ctx0, tmpq, n_rot, n_head, n_tokens,
+                ggml_element_size(tmpq) * n_embd_head,
+                ggml_element_size(tmpq) * n_embd_head * n_head,
+                0
+            );
+            offload_func_kq(qrot);
+            ggml_format_name(qrot, "qrot_%d", il);
+            struct ggml_tensor * krot = ggml_view_3d(
+                ctx0, tmpk, n_rot, n_head, n_tokens,
+                ggml_element_size(tmpk) * n_embd_head,
+                ggml_element_size(tmpk) * n_embd_head * n_head,
+                0
+            );
+            offload_func_kq(krot);
+            ggml_format_name(krot, "krot_%d", il);
+
+            // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
+            struct ggml_tensor * qpass = ggml_view_3d(
+                ctx0, tmpq, n_rot, n_head, n_tokens,
+                ggml_element_size(tmpq) * n_embd_head,
+                ggml_element_size(tmpq) * n_embd_head * n_head,
+                ggml_element_size(tmpq) * n_rot
+            );
+            offload_func_kq(qpass);
+            ggml_format_name(qpass, "qpass_%d", il);
+            struct ggml_tensor * kpass = ggml_view_3d(
+                ctx0, tmpk, n_rot, n_head, n_tokens,
+                ggml_element_size(tmpk) * n_embd_head,
+                ggml_element_size(tmpk) * n_embd_head * n_head,
+                ggml_element_size(tmpk) * n_rot
+            );
+            offload_func_kq(kpass);
+            ggml_format_name(kpass, "kpass_%d", il);
+
+            struct ggml_tensor * qrotated =  ggml_rope_custom(
+                    ctx0, qrot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale
+            );
+            offload_func_kq(qrotated);
+            struct ggml_tensor * krotated = ggml_rope_custom(
+                    ctx0, krot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale
+            );
+            offload_func_kq(krotated);
+            // ggml currently only supports concatenation on dim=2
+            // so we need to permute qrot, qpass, concat, then permute back.
+            qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3));
+            offload_func_kq(qrotated);
+            krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3));
+            offload_func_kq(krotated);
+
+            qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3));
+            offload_func_kq(qpass);
+            kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3));
+            offload_func_kq(kpass);
+
+            struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass);
+            offload_func_kq(Qcur);
+            struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass);
+            offload_func_kq(Kcur);
+
+            struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 1, 2, 0, 3));
+            offload_func_kq(Q);
+
+            Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3));
+            offload_func_kq(Kcur);
+            {
+                struct ggml_tensor * tmpv = ggml_view_3d(
+                        ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
+                        ggml_element_size(tmpqkv_perm) * n_embd_head,
+                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
+                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2
+                    );
+                offload_func_v(tmpv);
+                // store K, V in cache
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens));
+                offload_func_v(Vcur);
+                ggml_set_name(Vcur, "Vcur");
+
+                struct ggml_tensor * k = ggml_view_1d(
+                    ctx0, kv_self.k, n_tokens*n_embd_gqa,
+                    (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)
+                );
+                offload_func_kq(k);
+                ggml_set_name(k, "k");
+
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v));
+                offload_func_v(v);
+                ggml_set_name(v, "v");
+
+                // important: storing RoPE-ed version of K in the KV cache!
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+            struct ggml_tensor * K = ggml_view_3d(ctx0, kv_self.k,
+                    n_embd_head, n_kv, n_head_kv,
+                    ggml_element_size(kv_self.k)*n_embd_gqa,
+                    ggml_element_size(kv_self.k)*n_embd_head,
+                    ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il);
+
+            offload_func_kq(K);
+            ggml_format_name(K, "K_%d", il);
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
+            ggml_set_name(KQ, "KQ");
+
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
+            ggml_set_name(KQ_scaled, "KQ_scaled");
+
+            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask);
+            offload_func_kq(KQ_masked);
+            ggml_set_name(KQ_masked, "KQ_masked");
+
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            offload_func_kq(KQ_soft_max);
+            ggml_set_name(KQ_soft_max, "KQ_soft_max");
+
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_kv, n_embd_head, n_head_kv,
+                        ggml_element_size(kv_self.v)*n_ctx,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
+                        ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
+            offload_func_v(V);
+            ggml_set_name(V, "V");
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
+            ggml_set_name(KQV, "KQV");
+
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
+            ggml_set_name(KQV_merged, "KQV_merged");
+
+            cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens);
+            offload_func_v(cur);
+            ggml_set_name(cur, "KQV_merged_contiguous");
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
+            offload_func(cur);
+            cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            offload_func(cur);
+            ggml_set_name(cur, "result_wo");
+        }
+
+        struct ggml_tensor * inpFF = ggml_add(ctx0, residual, cur);
+        offload_func(inpFF);
+        ggml_set_name(inpFF, "inpFF");
+        {
+            // MLP
+            {
+                // Norm
+                cur = ggml_norm(ctx0, inpFF, norm_eps);
+                offload_func(cur);
+                cur = ggml_add(ctx0,
+                    ggml_mul(ctx0, cur, model.layers[il].ffn_norm),
+                    model.layers[il].ffn_norm_b
+                );
+                ggml_set_name(cur, "ffn_norm");
+                offload_func(cur);
+            }
+            cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur);
+            offload_func(cur);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].b3);
+            offload_func(cur);
+            ggml_set_name(cur, "result_ffn_up");
+
+            cur = ggml_sqr(ctx0, ggml_relu(ctx0, cur));
+            ggml_set_name(cur, "result_ffn_act");
+            offload_func(cur);
+            offload_func(cur->src[0]);
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur);
+            offload_func(cur);
+            cur = ggml_add(ctx0,
+                cur,
+                model.layers[il].b2);
+            offload_func(cur);
+            ggml_set_name(cur, "outFF");
+        }
+        cur = ggml_add(ctx0, cur, inpFF);
+        offload_func(cur);
+        ggml_set_name(cur, "inpFF_+_outFF");
+        inpL = cur;
+    }
+    cur = inpL;
+    {
+        cur = ggml_norm(ctx0, cur, norm_eps);
+        offload_func_nr(cur);
+        cur = ggml_mul(ctx0, cur, model.output_norm);
+        offload_func_nr(cur);
+
+        cur = ggml_add(ctx0, cur, model.output_norm_b);
+        // offload_func_nr(cur);
+
+        ggml_set_name(cur, "result_norm");
+    }
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    ggml_set_name(cur, "result_output");
+    ggml_build_forward_expand(gf, cur);
+    ggml_free(ctx0);
+    return gf;
+}
+
+static struct ggml_cgraph * llama_build_graph(
+         llama_context & lctx,
+     const llama_batch & batch) {
+    const auto & model = lctx.model;
+
+    struct ggml_cgraph * result = NULL;
+
+    switch (model.arch) {
+        case LLM_ARCH_LLAMA:
+            {
+                result = llm_build_llama(lctx, batch);
+            } break;
+        case LLM_ARCH_BAICHUAN:
+            {
+                result = llm_build_baichaun(lctx, batch);
+            } break;
+        case LLM_ARCH_FALCON:
+            {
+                result = llm_build_falcon(lctx, batch);
+            } break;
+        case LLM_ARCH_STARCODER:
+            {
+                result = llm_build_starcoder(lctx, batch);
+            } break;
+        case LLM_ARCH_PERSIMMON:
+            {
+                result = llm_build_persimmon(lctx, batch);
+            } break;
+        case LLM_ARCH_REFACT:
+            {
+                result = llm_build_refact(lctx, batch);
+            } break;
+        default:
+            GGML_ASSERT(false);
+    }
+
+    return result;
+}
+
+// decode a batch of tokens by evaluating the transformer
+//
+//   - lctx:      llama context
+//   - batch:     batch to evaluate
+//   - n_threads: number of threads to use
+//
+// return 0 on success
+// return positive int on warning
+// return negative int on error
+//
+static int llama_decode_internal(
+         llama_context & lctx,
+           llama_batch   batch) {
+    const uint32_t n_tokens = batch.n_tokens;
+
+    if (n_tokens == 0) {
+        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        return -1;
+    }
+
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    const auto n_batch = cparams.n_batch;
+
+    GGML_ASSERT(n_tokens <= n_batch);
+
+    int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
+
+    const int64_t t_start_us = ggml_time_us();
+
+#ifdef GGML_USE_MPI
+    // TODO: needs fix after #3228
+    GGML_ASSERT(false && "not implemented");
+    //ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
+#endif
+
+    GGML_ASSERT(n_threads > 0);
+
+    auto & kv_self = lctx.kv_self;
+
+    GGML_ASSERT(!!kv_self.ctx);
+
+    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_vocab = hparams.n_vocab;
+
+    // helpers for smoother batch API transistion
+    // after deprecating the llama_eval calls, these will be removed
+    std::vector<llama_pos>    pos;
+    std::vector<llama_seq_id> seq_id;
+
+    if (batch.pos == nullptr) {
+        pos.resize(n_tokens);
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            pos[i] = batch.all_pos_0 + i*batch.all_pos_1;
+        }
+
+        batch.pos = pos.data();
+    }
+
+    if (batch.seq_id == nullptr) {
+        seq_id.resize(n_tokens);
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            seq_id[i] = batch.all_seq_id;
+        }
+
+        batch.seq_id = seq_id.data();
+    }
+
+    if (!llama_kv_cache_find_slot(kv_self, batch)) {
+        return 1;
+    }
+
+    // a heuristic, to avoid attending the full cache if it is not yet utilized
+    // after enough generations, the benefit from this heuristic disappears
+    // if we start defragmenting the cache, the benefit from this will be more important
+    //kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32));   // TODO: this might be better for CUDA?
+    kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self)));
+
+    //printf("kv_self.n = %d\n", kv_self.n);
 
     ggml_allocr_reset(lctx.alloc);
 
-    ggml_cgraph * gf = llama_build_graph(lctx, tokens, embd, n_tokens, n_past);
+    ggml_cgraph * gf = llama_build_graph(lctx, batch);
 
     ggml_allocr_alloc_graph(lctx.alloc, gf);
 
@@ -2091,6 +5032,7 @@ static bool llama_eval_internal(
         ggml_tensor * node = gf->leafs[i];
         if (node->backend == GGML_BACKEND_GPU && node->extra == NULL) {
             ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data);
+            ggml_cuda_copy_to_device(node);
         }
     }
 
@@ -2100,19 +5042,36 @@ static bool llama_eval_internal(
             ggml_cuda_assign_scratch_offset(node, (char*)node->data - (char *) lctx.buf_alloc.data);
         }
     }
+
+    ggml_cuda_set_mul_mat_q(cparams.mul_mat_q);
 #endif
 
     // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
 
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+    // TODO: this is mostly important for Apple Silicon where CBLAS is still performing very well
+    //       we still need some threads to process all non-mul_mat ops, but not too much to avoid interfering
+    //       with the BLAS calls. need a better solution
+    if (n_tokens >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
+        n_threads = std::min(4, n_threads);
+    }
 
-    struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
+    // If all tensors can be run on the GPU then using more than 1 thread is detrimental.
+    const bool full_offload_supported = model.arch == LLM_ARCH_LLAMA ||
+        model.arch == LLM_ARCH_BAICHUAN ||
+        model.arch == LLM_ARCH_FALCON ||
+        model.arch == LLM_ARCH_REFACT;
+    const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
+    if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {
+        n_threads = 1;
+    }
+
+    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
     struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
 
-    GGML_ASSERT(strcmp(res->name, "result_output") == 0);
-    GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
+    GGML_ASSERT(strcmp(res->name,        "result_output") == 0);
+    GGML_ASSERT(strcmp(embeddings->name, "result_norm")   == 0);
 
 #if GGML_USE_MPI
     const int64_t n_layer = hparams.n_layer;
@@ -2123,10 +5082,6 @@ static bool llama_eval_internal(
     if (lctx.ctx_metal) {
         ggml_metal_set_n_cb     (lctx.ctx_metal, n_threads);
         ggml_metal_graph_compute(lctx.ctx_metal, gf);
-        ggml_metal_get_tensor   (lctx.ctx_metal, res);
-        if (!lctx.embedding.empty()) {
-            ggml_metal_get_tensor(lctx.ctx_metal, embeddings);
-        }
     } else {
         ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads);
     }
@@ -2138,11 +5093,12 @@ static bool llama_eval_internal(
     ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
 #endif
 
-    // update kv token count
-    lctx.kv_self.n = n_past + N;
-
-    if (cgraph_fname) {
-        ggml_graph_export(gf, cgraph_fname);
+    // update the kv ring buffer
+    lctx.kv_self.has_shift  = false;
+    lctx.kv_self.head      += n_tokens;
+    // Ensure kv cache head points to a valid index.
+    if (lctx.kv_self.head >= lctx.kv_self.size) {
+        lctx.kv_self.head = 0;
     }
 
 #ifdef GGML_PERF
@@ -2160,13 +5116,20 @@ static bool llama_eval_internal(
     {
         auto & logits_out = lctx.logits;
 
-        if (lctx.logits_all) {
-            logits_out.resize(n_vocab * N);
-            memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*n_vocab*N);
+        if (batch.logits) {
+            logits_out.resize(n_vocab * n_tokens);
+            for (uint32_t i = 0; i < n_tokens; i++) {
+                if (batch.logits[i] == 0) {
+                    continue;
+                }
+                memcpy(logits_out.data() + (n_vocab*i), (float *) ggml_get_data(res) + (n_vocab*i), sizeof(float)*n_vocab);
+            }
+        } else if (lctx.logits_all) {
+            logits_out.resize(n_vocab * n_tokens);
+            memcpy(logits_out.data(), (float *) ggml_get_data(res), sizeof(float)*n_vocab*n_tokens);
         } else {
-            // return result for just the last token
             logits_out.resize(n_vocab);
-            memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+            memcpy(logits_out.data(), (float *) ggml_get_data(res) + (n_vocab*(n_tokens - 1)), sizeof(float)*n_vocab);
         }
     }
 
@@ -2175,20 +5138,27 @@ static bool llama_eval_internal(
         auto & embedding_out = lctx.embedding;
 
         embedding_out.resize(n_embd);
-        memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd);
+        memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(n_tokens - 1)), sizeof(float)*n_embd);
     }
 
     // measure the performance only for the single-token evals
-    if (N == 1) {
+    if (n_tokens == 1) {
         lctx.t_eval_us += ggml_time_us() - t_start_us;
         lctx.n_eval++;
     }
-    else if (N > 1) {
+    else if (n_tokens > 1) {
         lctx.t_p_eval_us += ggml_time_us() - t_start_us;
-        lctx.n_p_eval += N;
+        lctx.n_p_eval += n_tokens;
     }
 
-    return true;
+    // get a more accurate load time, upon first eval
+    // TODO: fix this
+    if (!lctx.has_evaluated_once) {
+        lctx.t_load_us = ggml_time_us() - lctx.t_start_us;
+        lctx.has_evaluated_once = true;
+    }
+
+    return 0;
 }
 
 //
@@ -2211,73 +5181,56 @@ static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL;
 }
 
-static bool llama_is_user_defined_token(const llama_vocab & vocab, llama_token id) {
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
-}
-
-static bool llama_is_unused_token(const llama_vocab & vocab, llama_token id) {
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNUSED;
-}
-
 static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
     return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
 }
 
-static bool llama_is_bos_token(const llama_vocab & vocab, llama_token id) {
-    GGML_ASSERT(llama_is_control_token(vocab, id));
-    return id == vocab.special_bos_id;
+static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) {
+    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
 }
 
-static bool llama_is_eos_token(const llama_vocab & vocab, llama_token id ) {
-    GGML_ASSERT(llama_is_control_token(vocab, id));
-    return id == vocab.special_eos_id;
-}
-
-static bool llama_is_pad_token(const llama_vocab & vocab, llama_token id ) {
-    GGML_ASSERT(id < 0 || llama_is_control_token(vocab, id));
-    return id == vocab.special_pad_id;
-}
-
-static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
+static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
     GGML_ASSERT(llama_is_byte_token(vocab, id));
     const auto& token_data = vocab.id_to_token.at(id);
-    auto buf = token_data.text.substr(3, 2);
-    return strtol(buf.c_str(), NULL, 16);
+    switch (llama_vocab_get_type(vocab)) {
+    case LLAMA_VOCAB_TYPE_SPM: {
+        auto buf = token_data.text.substr(3, 2);
+        return strtol(buf.c_str(), NULL, 16);
+    }
+    case LLAMA_VOCAB_TYPE_BPE: {
+        GGML_ASSERT(false);
+        return unicode_to_bytes_bpe(token_data.text);
+    }
+    default:
+        GGML_ASSERT(false);
+    }
 }
 
 static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
-    char buf[7];
-    int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
-    GGML_ASSERT(0 <= result && result < 7);
-    return vocab.token_to_id.at(buf);
-}
-
-static std::string llama_escape_whitespace(const std::string& text) {
-    std::string result = "\xe2\x96\x81";
-    for (size_t offs = 0; offs < text.length(); ++offs) {
-        if (text[offs] == ' ') {
-            result += "\xe2\x96\x81";
-        } else {
-            result += text[offs];
-        }
+    switch (llama_vocab_get_type(vocab)) {
+    case LLAMA_VOCAB_TYPE_SPM: {
+        char buf[7];
+        int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
+        GGML_ASSERT(0 <= result && result < 7);
+        return vocab.token_to_id.at(buf);
     }
-    return result;
-}
-
-static std::string llama_unescape_whitespace(const std::string& word) {
-    if (word.length() >= 3 && word.substr(0, 3) == "\xe2\x96\x81") {
-        return std::string(" ") + word.substr(3);
+    case LLAMA_VOCAB_TYPE_BPE: {
+        return vocab.token_to_id.at(bytes_to_unicode_bpe(ch));
+    }
+    default:
+        GGML_ASSERT(false);
     }
-    return word;
 }
 
-static size_t utf8_len(char src) {
-    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
-    uint8_t highbits = static_cast<uint8_t>(src) >> 4;
-    return lookup[highbits];
+static void llama_escape_whitespace(std::string & text) {
+    replace_all(text, " ", "\xe2\x96\x81");
 }
 
-struct llama_sp_symbol {
+static void llama_unescape_whitespace(std::string & word) {
+    replace_all(word, "\xe2\x96\x81", " ");
+}
+
+struct llm_symbol {
     using index = int;
     index prev;
     index next;
@@ -2285,56 +5238,57 @@ struct llama_sp_symbol {
     size_t n;
 };
 
-static_assert(std::is_trivially_copyable<llama_sp_symbol>::value, "llama_sp_symbol is not trivially copyable");
+static_assert(std::is_trivially_copyable<llm_symbol>::value, "llm_symbol is not trivially copyable");
 
-struct llama_sp_bigram {
+// SPM tokenizer
+// original implementation:
+// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
+
+struct llm_bigram_spm {
     struct comparator {
-        bool operator()(llama_sp_bigram & l, llama_sp_bigram & r) {
+        bool operator()(llm_bigram_spm & l, llm_bigram_spm & r) {
             return (l.score < r.score) || (l.score == r.score && l.left > r.left);
         }
     };
-    using queue_storage = std::vector<llama_sp_bigram>;
-    using queue = std::priority_queue<llama_sp_bigram, queue_storage, comparator>;
-    llama_sp_symbol::index left;
-    llama_sp_symbol::index right;
+    using queue_storage = std::vector<llm_bigram_spm>;
+    using queue = std::priority_queue<llm_bigram_spm, queue_storage, comparator>;
+    llm_symbol::index left;
+    llm_symbol::index right;
     float score;
     size_t size;
 };
 
-// original implementation:
-// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4
-struct llama_tokenizer {
-    llama_tokenizer(const llama_vocab & vocab): vocab_(vocab) {}
+struct llm_tokenizer_spm {
+    llm_tokenizer_spm(const llama_vocab & vocab): vocab(vocab) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         // split string into utf8 chars
         int index = 0;
         size_t offs = 0;
         while (offs < text.size()) {
-            llama_sp_symbol sym;
+            llm_symbol sym;
             size_t len = utf8_len(text[offs]);
-            GGML_ASSERT(offs + len <= text.size());
             sym.text = text.c_str() + offs;
-            sym.n = len;
-            offs += len;
+            sym.n = std::min(len, text.size() - offs);
+            offs += sym.n;
             sym.prev = index - 1;
             sym.next = offs == text.size() ? -1 : index + 1;
             index++;
-            symbols_.emplace_back(sym);
+            symbols.emplace_back(sym);
         }
 
         // seed the work queue with all possible 2-character tokens.
-        for (size_t i = 1; i < symbols_.size(); ++i) {
+        for (size_t i = 1; i < symbols.size(); ++i) {
             try_add_bigram(i - 1, i);
         }
 
         // keep substituting the highest frequency pairs for as long as we can.
-        while (!work_queue_.empty()) {
-            auto bigram = work_queue_.top();
-            work_queue_.pop();
+        while (!work_queue.empty()) {
+            auto bigram = work_queue.top();
+            work_queue.pop();
 
-            auto & left_sym = symbols_[bigram.left];
-            auto & right_sym = symbols_[bigram.right];
+            auto & left_sym = symbols[bigram.left];
+            auto & right_sym = symbols[bigram.right];
 
             // if one of the symbols already got merged, skip it.
             if (left_sym.n == 0 || right_sym.n == 0 ||
@@ -2351,7 +5305,7 @@ struct llama_tokenizer {
             // remove the right sym from the chain
             left_sym.next = right_sym.next;
             if (right_sym.next >= 0) {
-                symbols_[right_sym.next].prev = bigram.left;
+                symbols[right_sym.next].prev = bigram.left;
             }
 
             // find more substitutions
@@ -2359,19 +5313,19 @@ struct llama_tokenizer {
             try_add_bigram(bigram.left, left_sym.next);
         }
 
-        for (int i = 0; i != -1; i = symbols_[i].next) {
-            auto & symbol = symbols_[i];
+        for (int i = 0; i != -1; i = symbols[i].next) {
+            auto & symbol = symbols[i];
             resegment(symbol, output);
         }
     }
 
 private:
-    void resegment(llama_sp_symbol &symbol, std::vector<llama_vocab::id> &output) {
+    void resegment(llm_symbol & symbol, std::vector<llama_vocab::id> & output) {
         auto text = std::string(symbol.text, symbol.n);
-        auto token = vocab_.token_to_id.find(text);
+        auto token = vocab.token_to_id.find(text);
 
         // Do we need to support is_unused?
-        if (token != vocab_.token_to_id.end()) {
+        if (token != vocab.token_to_id.end()) {
             output.push_back((*token).second);
             return;
         }
@@ -2381,14 +5335,14 @@ private:
         if (p == rev_merge.end()) {
             // output any symbols that did not form tokens as bytes.
             for (int j = 0; j < (int)symbol.n; ++j) {
-                llama_vocab::id token_id = llama_byte_to_token(vocab_, symbol.text[j]);
+                llama_vocab::id token_id = llama_byte_to_token(vocab, symbol.text[j]);
                 output.push_back(token_id);
             }
             return;
         }
 
-        resegment(symbols_[p->second.first], output);
-        resegment(symbols_[p->second.second], output);
+        resegment(symbols[p->second.first],  output);
+        resegment(symbols[p->second.second], output);
     }
 
     void try_add_bigram(int left, int right) {
@@ -2396,56 +5350,375 @@ private:
             return;
         }
 
-        const std::string text = std::string(symbols_[left].text, symbols_[left].n + symbols_[right].n);
-        auto token = vocab_.token_to_id.find(text);
+        const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n);
+        auto token = vocab.token_to_id.find(text);
 
-        if (token == vocab_.token_to_id.end()) {
+        if (token == vocab.token_to_id.end()) {
             return;
         }
 
-        if (static_cast<size_t>((*token).second) >= vocab_.id_to_token.size()) {
+        if (static_cast<size_t>((*token).second) >= vocab.id_to_token.size()) {
             return;
         }
 
-        const auto &tok_data = vocab_.id_to_token[(*token).second];
+        const auto & tok_data = vocab.id_to_token[(*token).second];
 
-        llama_sp_bigram bigram;
-        bigram.left = left;
+        llm_bigram_spm bigram;
+        bigram.left  = left;
         bigram.right = right;
         bigram.score = tok_data.score;
-        bigram.size = text.size();
-        work_queue_.push(bigram);
+        bigram.size  = text.size();
+
+        work_queue.push(bigram);
 
         // Do we need to support is_unused?
         rev_merge[text] = std::make_pair(left, right);
     }
 
-    const llama_vocab & vocab_;
-    std::vector<llama_sp_symbol> symbols_;
-    llama_sp_bigram::queue work_queue_;
-    std::map<std::string, std::pair<int, int> > rev_merge;
+    const llama_vocab & vocab;
+
+    std::vector<llm_symbol> symbols;
+    llm_bigram_spm::queue work_queue;
+
+    std::map<std::string, std::pair<int, int>> rev_merge;
 };
 
-static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape) {
-    llama_tokenizer tokenizer(vocab);
+// BPE tokenizer
+// adapted from https://github.com/cmp-nct/ggllm.cpp [MIT License]
+// tried to simplify unicode stuff, so most likely does not work 100% correctly!
+
+// TODO: there are a lot of common parts between spm and bpe tokenizers, should be refactored and reused
+
+struct llm_bigram_bpe {
+    struct comparator {
+        bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const {
+            return l.rank > r.rank || (l.rank == r.rank && l.left > r.left);
+        }
+    };
+
+    using queue_storage = std::vector<llm_bigram_bpe>;
+    using queue = std::priority_queue<llm_bigram_bpe, queue_storage, comparator>;
+    llm_symbol::index left;
+    llm_symbol::index right;
+    std::string text;
+    int rank;
+    size_t size;
+};
+
+struct llm_tokenizer_bpe {
+    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {}
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+        int final_prev_index = -1;
+        auto word_collection = bpe_gpt2_preprocess(text);
+
+        symbols_final.clear();
+
+        for (auto & word : word_collection) {
+            work_queue = llm_bigram_bpe::queue();
+            symbols.clear();
+
+            int index = 0;
+            size_t offset = 0;
+
+            while (offset < word.size()) {
+                llm_symbol sym;
+                size_t char_len = std::min(word.size() - offset, (size_t) ::utf8_len(word[offset]));
+                sym.text = word.c_str() + offset;
+                sym.n = 1;
+                sym.n = char_len;
+                offset += sym.n;
+                sym.prev = index - 1;
+                sym.next = offset == word.size() ? -1 : index + 1;
+                index++;
+                symbols.emplace_back(sym);
+            }
+            for (size_t i = 1; i < symbols.size(); ++i) {
+                add_new_bigram(i - 1, i);
+            }
+
+            // build token(s)
+            while (!work_queue.empty()) {
+                auto bigram = work_queue.top();
+                work_queue.pop();
+
+                auto & left_symbol = symbols[bigram.left];
+                auto & right_symbol = symbols[bigram.right];
+
+                if (left_symbol.n == 0 || right_symbol.n == 0) {
+                    continue;
+                }
+                std::string left_token = std::string(left_symbol.text, left_symbol.n);
+                std::string right_token = std::string(right_symbol.text, right_symbol.n);
+                if (left_token + right_token != bigram.text) {
+                    continue;  // Skip this bigram if it's outdated
+                }
+
+                // merge the right sym into the left one
+                left_symbol.n += right_symbol.n;
+                right_symbol.n = 0;
+
+                // remove the right sym from the chain
+                left_symbol.next = right_symbol.next;
+                if (right_symbol.next >= 0) {
+                    symbols[right_symbol.next].prev = bigram.left;
+                }
+
+                add_new_bigram(left_symbol.prev, bigram.left);  // left side of current symbol
+                add_new_bigram(bigram.left, left_symbol.next);  // right side of current symbol
+            }
+
+            // add the fnished tokens to the final list keeping correct order for next and prev
+            for (auto & sym : symbols) {
+                if (sym.n > 0) {
+                    sym.prev = final_prev_index;
+                    sym.next = -1;
+                    if (final_prev_index != -1) {
+                        symbols_final[final_prev_index].next = symbols_final.size();
+                    }
+                    symbols_final.emplace_back(sym);
+                    final_prev_index = symbols_final.size() - 1;
+                }
+            }
+        }
+
+        symbols = symbols_final;
+
+        if (!symbols.empty()) {
+            for (int i = 0; i != -1; i = symbols[i].next) {
+                auto & symbol = symbols[i];
+                if (symbol.n == 0) {
+                    continue;
+                }
+
+                const std::string str = std::string(symbol.text, symbol.n);
+                const auto token = vocab.token_to_id.find(str);
+
+                if (token == vocab.token_to_id.end()) {
+                    for (auto j = str.begin(); j != str.end(); ++j) {
+                        std::string byte_str(1, *j);
+                        auto token_multibyte = vocab.token_to_id.find(byte_str);
+                        if (token_multibyte == vocab.token_to_id.end()) {
+                            throw std::runtime_error("ERROR: byte not found in vocab");
+                        }
+                        output.push_back((*token_multibyte).second);
+                    }
+                } else {
+                    output.push_back((*token).second);
+                }
+            }
+        }
+    }
+
+private:
+    void add_new_bigram(int left, int right) {
+        if (left == -1 || right == -1) {
+            return;
+        }
+
+        std::string left_token  = std::string(symbols[left].text,  symbols[left].n);
+        std::string right_token = std::string(symbols[right].text, symbols[right].n);
+
+        int rank_found = -1;
+
+        rank_found = vocab.find_bpe_rank(left_token, right_token);
+
+        if (rank_found < 0) {
+            return;
+        }
+
+        llm_bigram_bpe bigram;
+
+        bigram.left  = left;
+        bigram.right = right;
+        bigram.text  = left_token + right_token;
+        bigram.size  = left_token.size() + right_token.size();
+        bigram.rank  = rank_found;
+
+        work_queue.push(bigram);
+    }
+
+    std::vector<std::string> bpe_gpt2_preprocess(const std::string & text) {
+        std::vector<std::string> bpe_words;
+        std::vector<std::string> bpe_encoded_words;
+
+        std::string token = "";
+        // GPT2 system regex:  's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
+        bool collecting_numeric = false;
+        bool collecting_letter = false;
+        bool collecting_special = false;
+        bool collecting_whitespace_lookahead = false;
+        bool collecting = false;
+
+        std::vector<std::string> text_utf;
+        text_utf.reserve(text.size());
+        bpe_words.reserve(text.size());
+        bpe_encoded_words.reserve(text.size());
+
+        auto cps = codepoints_from_utf8(text);
+        for (size_t i = 0; i < cps.size(); ++i)
+            text_utf.emplace_back(codepoint_to_utf8(cps[i]));
+
+        for (int i = 0; i < (int)text_utf.size(); i++) {
+            const std::string & utf_char = text_utf[i];
+            bool split_condition = false;
+            // const char* text_pos = raw_text_p + utf_char.seq_offset_bytes;
+            int bytes_remain = text_utf.size() - i;
+            // forward backward lookups
+            const std::string & utf_char_next = (i + 1 < (int)text_utf.size()) ? text_utf[i + 1] : "";
+            const std::string & utf_char_next_next = (i + 2 < (int)text_utf.size()) ? text_utf[i + 2] : "";
+
+            // handling contractions
+            if (!split_condition && bytes_remain >= 2) {
+                // 's|'t|'m|'d
+                if (utf_char == "\'" && (utf_char_next == "s" || utf_char_next == "t" || utf_char_next == "m" || utf_char_next == "d")) {
+                    split_condition = true;
+                }
+                if (split_condition) {
+                    if (token.size()) {
+                        bpe_words.emplace_back(token); // push previous content as token
+                    }
+                    token = utf_char + utf_char_next;
+                    bpe_words.emplace_back(token);
+                    token = "";
+                    i++;
+                    continue;
+                }
+            }
+            if (!split_condition && bytes_remain >= 3) {
+                // 're|'ve|'ll
+                if (utf_char == "\'" && (
+                    (utf_char_next == "r" || utf_char_next_next == "e") ||
+                    (utf_char_next == "v" || utf_char_next_next == "e") ||
+                    (utf_char_next == "l" || utf_char_next_next == "l"))
+                    ) {
+                    split_condition = true;
+                }
+                if (split_condition) {
+                    // current token + next token can be defined
+                    if (token.size()) {
+                        bpe_words.emplace_back(token); // push previous content as token
+                    }
+                    token = utf_char + utf_char_next + utf_char_next_next;
+                    bpe_words.emplace_back(token); // the contraction
+                    token = "";
+                    i += 2;
+                    continue;
+                }
+            }
+
+            if (!split_condition && !collecting) {
+                if (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) {
+                    collecting_letter = true;
+                    collecting = true;
+                }
+                else if (codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
+                    collecting_numeric = true;
+                    collecting = true;
+                }
+                else if (
+                    ((codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (codepoint_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
+                    (!token.size() && utf_char == " " && codepoint_type(utf_char_next) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
+                    ) {
+                    collecting_special = true;
+                    collecting = true;
+                }
+                else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && codepoint_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) {
+                    collecting_whitespace_lookahead = true;
+                    collecting = true;
+                }
+                else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) {
+                    split_condition = true;
+                }
+            }
+            else if (!split_condition && collecting) {
+                if (collecting_letter && codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER) {
+                    split_condition = true;
+                }
+                else if (collecting_numeric && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) {
+                    split_condition = true;
+                }
+                else if (collecting_special && (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
+                    split_condition = true;
+                }
+                else if (collecting_whitespace_lookahead && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) {
+                    split_condition = true;
+                }
+            }
+
+            if (utf_char_next == "") {
+                split_condition = true; // final
+                token += utf_char;
+            }
+
+            if (split_condition) {
+                if (token.size()) {
+                    bpe_words.emplace_back(token);
+                }
+                token = utf_char;
+                collecting = false;
+                collecting_letter = false;
+                collecting_numeric = false;
+                collecting_special = false;
+                collecting_whitespace_lookahead = false;
+            }
+            else {
+                token += utf_char;
+            }
+        }
+
+        for (std::string & word : bpe_words) {
+            std::string encoded_token = "";
+            for (char & c : word) {
+                encoded_token += bytes_to_unicode_bpe(c);
+            }
+            bpe_encoded_words.emplace_back(encoded_token);
+        }
+
+        return bpe_encoded_words;
+    }
+
+    const llama_vocab & vocab;
+
+    std::vector<llm_symbol> symbols;
+    std::vector<llm_symbol> symbols_final;
+
+    llm_bigram_bpe::queue work_queue;
+};
+
+static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos) {
     std::vector<llama_vocab::id> output;
 
+    // OG tokenizer behavior:
+    //
+    // tokenizer.encode('', add_bos=True)  returns [1]
+    // tokenizer.encode('', add_bos=False) returns []
+
+    if (bos && vocab.special_bos_id != -1) {
+        output.push_back(vocab.special_bos_id);
+    }
+
     if (raw_text.empty()) {
         return output;
     }
 
-    if (bos) {
-        output.push_back(vocab.special_bos_id);
+    switch (vocab.type) {
+        case LLAMA_VOCAB_TYPE_SPM:
+            {
+                // without adding this leading whitespace, we do not get the same results as the original tokenizer
+                raw_text = " " + raw_text;
+
+                llm_tokenizer_spm tokenizer(vocab);
+                llama_escape_whitespace(raw_text);
+                tokenizer.tokenize(raw_text, output);
+            } break;
+        case LLAMA_VOCAB_TYPE_BPE:
+            {
+                llm_tokenizer_bpe tokenizer(vocab);
+                tokenizer.tokenize(raw_text, output);
+            } break;
     }
 
-    std::string text;
-    if (escape) {
-        text = llama_escape_whitespace(raw_text);
-    } else {
-        text = raw_text;
-    }
-
-    tokenizer.tokenize(text, output);
     return output;
 }
 
@@ -2474,7 +5747,7 @@ struct llama_grammar_candidate {
 
 // Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as
 // pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`.
-std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
+static std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
         const char         * src,
         llama_partial_utf8   partial_start) {
     static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 };
@@ -2623,7 +5896,7 @@ static void llama_grammar_advance_stack(
         std::vector<std::vector<const llama_grammar_element *>> & new_stacks) {
 
     if (stack.empty()) {
-        new_stacks.push_back(stack);
+        new_stacks.emplace_back(stack);
         return;
     }
 
@@ -2660,7 +5933,7 @@ static void llama_grammar_advance_stack(
         }
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
-            new_stacks.push_back(stack);
+            new_stacks.emplace_back(stack);
             break;
         default:
             // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
@@ -2825,10 +6098,36 @@ void llama_grammar_free(struct llama_grammar * grammar) {
     delete grammar;
 }
 
+struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar) {
+    llama_grammar * result = new llama_grammar{ grammar->rules, grammar->stacks, grammar->partial_utf8 };
+
+    // redirect elements in stacks to point to new rules
+    for (size_t is = 0; is < result->stacks.size(); is++) {
+        for (size_t ie = 0; ie < result->stacks[is].size(); ie++) {
+            for (size_t ir0 = 0; ir0 < grammar->rules.size(); ir0++) {
+                for (size_t ir1 = 0; ir1 < grammar->rules[ir0].size(); ir1++) {
+                    if (grammar->stacks[is][ie] == &grammar->rules[ir0][ir1]) {
+                         result->stacks[is][ie]  =  &result->rules[ir0][ir1];
+                    }
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
 //
 // sampling
 //
 
+void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
+    if (seed == LLAMA_DEFAULT_SEED) {
+        seed = time(NULL);
+    }
+    ctx->rng.seed(seed);
+}
+
 void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates) {
     GGML_ASSERT(candidates->size > 0);
 
@@ -2936,7 +6235,7 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
 
     // Calculate absolute value of second derivatives
     for (size_t i = 0; i < second_derivatives.size(); ++i) {
-        second_derivatives[i] = abs(second_derivatives[i]);
+        second_derivatives[i] = std::abs(second_derivatives[i]);
     }
 
     // Normalize the second derivatives
@@ -3037,7 +6336,7 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
     }
 }
 
-void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
+void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
     const int64_t t_start_sample_us = ggml_time_us();
 
     for (size_t i = 0; i < candidates_p->size; ++i) {
@@ -3049,6 +6348,10 @@ void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array
     }
 }
 
+void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
+    llama_sample_temp(ctx, candidates_p, temp);
+}
+
 void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty) {
     if (last_tokens_size == 0 || penalty == 1.0f) {
         return;
@@ -3127,16 +6430,16 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
     std::vector<llama_grammar_candidate>                              candidates_grammar;
 
     for (size_t i = 0; i < candidates->size; ++i) {
-        const llama_token id   = candidates->data[i].id;
-        const std::string text = llama_token_to_text(ctx, id);
+        const llama_token id    = candidates->data[i].id;
+        const std::string piece = llama_token_to_str(ctx, id);
         if (id == eos) {
             if (!allow_eos) {
                 candidates->data[i].logit = -INFINITY;
             }
-        } else if (text.empty()) {
+        } else if (piece.empty() || piece[0] == 0) {
             candidates->data[i].logit = -INFINITY;
         } else {
-            candidates_decoded.push_back(decode_utf8(text.c_str(), grammar->partial_utf8));
+            candidates_decoded.push_back(decode_utf8(piece.c_str(), grammar->partial_utf8));
             candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
         }
     }
@@ -3172,7 +6475,7 @@ void llama_sample_classifier_free_guidance(
 
     GGML_ASSERT(ctx);
 
-    auto n_vocab = llama_n_vocab(ctx);
+    auto n_vocab = llama_n_vocab(llama_get_model(ctx));
 
     GGML_ASSERT(n_vocab == (int)candidates->size);
     GGML_ASSERT(!candidates->sorted);
@@ -3201,7 +6504,7 @@ void llama_sample_classifier_free_guidance(
 llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu) {
     GGML_ASSERT(ctx);
 
-    auto N = float(llama_n_vocab(ctx));
+    auto N = float(llama_n_vocab(llama_get_model(ctx)));
     int64_t t_start_sample_us;
     t_start_sample_us = ggml_time_us();
 
@@ -3340,10 +6643,10 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
         GGML_ASSERT(false);
     }
 
-    const std::string text = llama_token_to_text(ctx, token);
+    const std::string piece = llama_token_to_str(ctx, token);
 
     // Note terminating 0 in decoded string
-    const auto   decoded     = decode_utf8(text.c_str(), grammar->partial_utf8);
+    const auto   decoded     = decode_utf8(piece.c_str(), grammar->partial_utf8);
     const auto & code_points = decoded.first;
     for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
         grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
@@ -3354,11 +6657,269 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
     ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
 }
 
+//
+// Beam search
+//
+
+struct llama_beam {
+    std::vector<llama_token> tokens;
+    float p;  // Cumulative beam probability (renormalized relative to all beams)
+    bool eob; // Initialize end-of-beam to false. Callback sets this to true.
+    // Sort beams by probability. In case of ties, prefer beams at eob.
+    bool operator<(const llama_beam & rhs) const {
+        return std::make_pair(p, eob) < std::make_pair(rhs.p, rhs.eob);
+    }
+    // Shift off first n tokens and discard them.
+    void shift_tokens(const size_t n) {
+        if (n) {
+            std::copy(tokens.begin() + n, tokens.end(), tokens.begin());
+            tokens.resize(tokens.size() - n);
+        }
+    }
+    llama_beam_view view() const { return {tokens.data(), tokens.size(), p, eob}; }
+};
+
+// A struct for calculating logit-related info.
+struct llama_logit_info {
+    const float * const logits;
+    const int n_vocab;
+    const float max_l;
+    const float normalizer;
+    struct sum_exp {
+        float max_l;
+        float operator()(float sum, float l) const { return sum + std::exp(l - max_l); }
+    };
+    llama_logit_info(llama_context * ctx)
+      : logits(llama_get_logits(ctx))
+      , n_vocab(llama_n_vocab(llama_get_model(ctx)))
+      , max_l(*std::max_element(logits, logits + n_vocab))
+      , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l}))
+      { }
+    llama_token_data get_token_data(const llama_token token_id) const {
+        constexpr auto p = std::numeric_limits<float>::quiet_NaN();  // never used
+        return {token_id, logits[token_id], p};
+    }
+    // Return top k token_data by logit.
+    std::vector<llama_token_data> top_k(size_t k) {
+        std::vector<llama_token_data> min_heap;  // min-heap by logit
+        const llama_token k_min = std::min(static_cast<llama_token>(k), n_vocab);
+        min_heap.reserve(k_min);
+        for (llama_token token_id = 0 ; token_id < k_min ; ++token_id) {
+            min_heap.push_back(get_token_data(token_id));
+        }
+        auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; };
+        std::make_heap(min_heap.begin(), min_heap.end(), comp);
+        for (llama_token token_id = k_min ; token_id < n_vocab ; ++token_id) {
+            if (min_heap.front().logit < logits[token_id]) {
+                std::pop_heap(min_heap.begin(), min_heap.end(), comp);
+                min_heap.back().id = token_id;
+                min_heap.back().logit = logits[token_id];
+                std::push_heap(min_heap.begin(), min_heap.end(), comp);
+            }
+        }
+        return min_heap;
+    }
+    float probability_from_logit(float logit) const {
+        return normalizer * std::exp(logit - max_l);
+    }
+};
+
+struct llama_beam_search_data {
+    llama_context * ctx;
+    size_t n_beams;
+    int n_past;
+    int n_predict;
+    std::vector<llama_beam> beams;
+    std::vector<llama_beam> next_beams;
+
+    // Re-calculated on each loop iteration
+    size_t common_prefix_length;
+
+    // Used to communicate to/from callback on beams state.
+    std::vector<llama_beam_view> beam_views;
+
+    llama_beam_search_data(llama_context * ctx, size_t n_beams, int n_past, int n_predict)
+      : ctx(ctx)
+      , n_beams(n_beams)
+      , n_past(n_past)
+      , n_predict(n_predict)
+      , beam_views(n_beams) {
+        beams.reserve(n_beams);
+        next_beams.reserve(n_beams);
+    }
+
+    // Collapse beams to a single beam given by index.
+    void collapse_beams(const size_t beam_idx) {
+        if (0u < beam_idx) {
+            std::swap(beams[0], beams[beam_idx]);
+        }
+        beams.resize(1);
+    }
+
+    // Min-heaps are used to efficiently collect the top-k elements (k=n_beams).
+    // The repetative patterns below reflect the 2 stages of heaps:
+    //  * Gather elements until the vector is full, then call std::make_heap() on it.
+    //  * If the heap is full and a new element is found that should be included, pop the
+    //    least element to the back(), replace it with the new, then push it into the heap.
+    void fill_next_beams_by_top_probabilities(llama_beam & beam) {
+        // Min-heaps use a greater-than comparator.
+        const auto comp = [](const llama_beam & a, const llama_beam & b) { return a.p > b.p; };
+        if (beam.eob) {
+            // beam is at end-of-sentence, so just copy it to next_beams if its probability is high enough.
+            if (next_beams.size() < n_beams) {
+                next_beams.push_back(std::move(beam));
+                if (next_beams.size() == n_beams) {
+                    std::make_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            } else if (next_beams.front().p < beam.p) {
+                std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                next_beams.back() = std::move(beam);
+                std::push_heap(next_beams.begin(), next_beams.end(), comp);
+            }
+        } else {
+            // beam is not at end-of-sentence, so branch with next top_k tokens.
+            if (!beam.tokens.empty()) {
+                llama_decode(ctx, llama_batch_get_one(beam.tokens.data(), beam.tokens.size(), n_past, 0));
+            }
+            llama_logit_info logit_info(ctx);
+            std::vector<llama_token_data> next_tokens = logit_info.top_k(n_beams);
+            size_t i=0;
+            if (next_beams.size() < n_beams) {
+                for (; next_beams.size() < n_beams ; ++i) {
+                    llama_beam next_beam = beam;
+                    next_beam.tokens.push_back(next_tokens[i].id);
+                    next_beam.p *= logit_info.probability_from_logit(next_tokens[i].logit);
+                    next_beams.push_back(std::move(next_beam));
+                }
+                std::make_heap(next_beams.begin(), next_beams.end(), comp);
+            } else {
+                for (; next_beams.front().p == 0.0f ; ++i) {
+                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                    next_beams.back() = beam;
+                    next_beams.back().tokens.push_back(next_tokens[i].id);
+                    next_beams.back().p *= logit_info.probability_from_logit(next_tokens[i].logit);
+                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            }
+            for (; i < n_beams ; ++i) {
+                const float next_p = beam.p * logit_info.probability_from_logit(next_tokens[i].logit);
+                if (next_beams.front().p < next_p) {
+                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
+                    next_beams.back() = beam;
+                    next_beams.back().tokens.push_back(next_tokens[i].id);
+                    next_beams.back().p = next_p;
+                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
+                }
+            }
+        }
+    }
+
+    // Find common_prefix_length based on beams.
+    // Requires beams is not empty.
+    size_t find_common_prefix_length() {
+        size_t common_prefix_length = beams[0].tokens.size();
+        for (size_t i = 1 ; i < beams.size() ; ++i) {
+            common_prefix_length = std::min(common_prefix_length, beams[i].tokens.size());
+            for (size_t j = 0 ; j < common_prefix_length ; ++j) {
+                if (beams[0].tokens[j] != beams[i].tokens[j]) {
+                    common_prefix_length = j;
+                    break;
+                }
+            }
+        }
+        return common_prefix_length;
+    }
+
+    // Construct beams_state to send back to caller via the callback function.
+    // Side effect: set common_prefix_length = find_common_prefix_length();
+    llama_beams_state get_beams_state(const bool last_call) {
+        for (size_t i = 0 ; i < beams.size() ; ++i) {
+            beam_views[i] = beams[i].view();
+        }
+        common_prefix_length = find_common_prefix_length();
+        return {beam_views.data(), beams.size(), common_prefix_length, last_call};
+    }
+
+    // Loop:
+    //  * while i < n_predict, AND
+    //  * any of the beams have not yet reached end-of-beam (eob), AND
+    //  * the highest probability beam(s) (plural in case of ties) are not at end-of-sentence
+    //    (since all other beam probabilities can only decrease)
+    void loop(const llama_beam_search_callback_fn_t callback, void * const callback_data) {
+        beams.push_back({{}, 1.0f, false});  // Start with one empty beam w/ probability = 1.0 and !eob.
+        const auto not_eob = [](const llama_beam & beam) { return !beam.eob; };
+        for (int i = 0 ; i < n_predict && std::any_of(beams.begin(),beams.end(),not_eob) &&
+                       !beams[top_beam_index()].eob ; ++i) {
+            callback(callback_data, get_beams_state(false));  // Sets common_prefix_length
+            update_beams_from_beam_views();   // Update values (p,eob) that callback may have changed.
+            if (common_prefix_length) {
+                llama_decode(ctx, llama_batch_get_one(beams[0].tokens.data(), common_prefix_length, n_past, 0));
+                n_past += common_prefix_length;
+            }
+            // Zero-out next_beam probabilities to place them last in following min-heap.
+            std::for_each(next_beams.begin(), next_beams.end(), [](llama_beam & beam) { beam.p = 0.0f; });
+            for (llama_beam & beam : beams) {
+                beam.shift_tokens(common_prefix_length);
+                fill_next_beams_by_top_probabilities(beam);
+            }
+            // next_beams become the beams of next/final iteration. Swap them to re-use memory.
+            beams.swap(next_beams);
+            renormalize_beam_probabilities(beams);
+        }
+        collapse_beams(top_beam_index());
+        callback(callback_data, get_beams_state(true));
+    }
+
+    // As beams grow, the cumulative probabilities decrease.
+    // Renormalize them to avoid floating point underflow.
+    static void renormalize_beam_probabilities(std::vector<llama_beam> & beams) {
+        const auto sum_p = [](float sum, llama_beam & beam) { return sum + beam.p; };
+        const float inv_sum = 1.0f / std::accumulate(beams.begin(), beams.end(), 0.0f, sum_p);
+        std::for_each(beams.begin(), beams.end(), [=](llama_beam & beam) { beam.p *= inv_sum; });
+    }
+
+    // Assumes beams is non-empty.  Uses llama_beam::operator<() for ordering.
+    size_t top_beam_index() {
+        return std::max_element(beams.begin(), beams.end()) - beams.begin();
+    }
+
+    // Copy (p,eob) for each beam which may have been changed by the callback.
+    void update_beams_from_beam_views() {
+        for (size_t i = 0 ; i < beams.size() ; ++i) {
+            beams[i].p = beam_views[i].p;
+            beams[i].eob = beam_views[i].eob;
+        }
+    }
+};
+
+void llama_beam_search(llama_context * ctx,
+                       llama_beam_search_callback_fn_t callback, void * callback_data,
+                       size_t n_beams, int n_past, int n_predict) {
+    assert(ctx);
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    llama_beam_search_data beam_search_data(ctx, n_beams, n_past, n_predict);
+
+    beam_search_data.loop(callback, callback_data);
+
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    ctx->n_sample++;
+}
+
 //
 // quantization
 //
 
-static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vector<float> & output, const size_t nelements, const int nthread) {
+template <typename T>
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
+};
+
+static void llama_convert_tensor_internal(
+    struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
+    const size_t nelements, const int nthread
+) {
     if (output.size() < nelements) {
         output.resize(nelements);
     }
@@ -3393,7 +6954,6 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect
     auto blocks_per_thread = nblocks / nthread;
     auto spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count
 
-    std::vector<std::thread> workers;
     for (auto tnum = 0, in_buff_offs = 0, out_buff_offs = 0; tnum < nthread; tnum++) {
         auto thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
         auto thr_elems = thr_blocks * block_size; // number of elements for this thread
@@ -3406,15 +6966,124 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect
                 qtype.to_float(inbuf, outbuf, nels);
             }
         };
-        workers.push_back(std::thread(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems));
+        workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
         in_buff_offs += thr_block_bytes;
         out_buff_offs += thr_elems;
     }
-    for (auto & worker : workers) {
-        worker.join();
-    }
+    for (auto & w : workers) { w.join(); }
+    workers.clear();
 }
 
+#ifdef GGML_USE_K_QUANTS
+static ggml_type get_k_quant_type(
+    ggml_type new_type, const ggml_tensor * tensor, const llama_model & model, llama_ftype ftype, int * i_attention_wv,
+    int n_attention_wv, int * i_feed_forward_w2, int n_feed_forward_w2
+) {
+    const std::string name = ggml_get_name(tensor);
+    // TODO: avoid hardcoded tensor names - use the TN_* constants
+    const auto tn = LLM_TN(model.arch);
+
+    auto use_more_bits = [](int i_layer, int num_layers) -> bool {
+        return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
+    };
+
+    if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
+        int nx = tensor->ne[0];
+        if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (new_type != GGML_TYPE_Q8_0) {
+            new_type = GGML_TYPE_Q6_K;
+        }
+    } else if (name.find("attn_v.weight") != std::string::npos) {
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = *i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
+                use_more_bits(*i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && *i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
+        else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
+                (*i_attention_wv < n_attention_wv/8 || *i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
+        if (model.type == MODEL_70B) {
+            // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
+            // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
+            // nearly negligible increase in model size by quantizing this tensor with more bits:
+            if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
+        }
+        ++*i_attention_wv;
+    } else if (name.find("ffn_down.weight") != std::string::npos) {
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
+                     : model.arch != LLM_ARCH_FALCON || use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K
+                     : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+            new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+            if (model.arch == LLM_ARCH_FALCON) {
+                new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
+                           use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+            } else {
+                if (use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+            }
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && *i_feed_forward_w2 < 4) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        ++*i_feed_forward_w2;
+    } else if (name.find("attn_output.weight") != std::string::npos) {
+        if (model.arch != LLM_ARCH_FALCON) {
+            if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
+            else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
+            else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+        }
+    }
+    else if (name.find("attn_qkv.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
+    }
+    else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    }
+    // This can be used to reduce the size of the Q5_K_S model.
+    // The associated PPL increase is fully in line with the size reduction
+    //else {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
+    //}
+    bool convert_incompatible_tensor = false;
+    if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
+        new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
+        int nx = tensor->ne[0];
+        int ny = tensor->ne[1];
+        if (nx % QK_K != 0) {
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
+            convert_incompatible_tensor = true;
+        }
+    }
+    if (convert_incompatible_tensor) {
+        if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
+            new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
+            LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
+        } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+            new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
+            LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
+        } else {
+            throw std::runtime_error("Unsupported tensor size encountered\n");
+        }
+    }
+
+    return new_type;
+}
+#endif
+
 static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
     ggml_type quantized_type;
     llama_ftype ftype = params->ftype;
@@ -3449,13 +7118,32 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         nthread = std::thread::hardware_concurrency();
     }
 
-    std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false));
+    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
+#if defined(__linux__) || defined(_WIN32)
+    constexpr bool use_mmap = true;
+#else
+    constexpr bool use_mmap = false;
+#endif
+
+    llama_model_loader ml(fname_inp, use_mmap);
+    if (ml.use_mmap) {
+        ml.mapping.reset(new llama_mmap(&ml.file, /* prefetch */ 0, ggml_is_numa()));
+    }
+
+    llama_model model;
+    llm_load_arch(ml, model);
+    llm_load_hparams(ml, model);
+
+    if (params->only_copy) {
+        ftype = model.ftype;
+    }
 
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     struct gguf_context * ctx_out = gguf_init_empty();
 
     // copy the KV pairs from the input file
-    gguf_set_kv     (ctx_out, model_loader->ctx_gguf);
+    gguf_set_kv     (ctx_out, ml.ctx_gguf);
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", ftype);
 
@@ -3463,8 +7151,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     int n_attention_wv    = 0;
     int n_feed_forward_w2 = 0;
 
-    for (int i = 0; i < model_loader->n_tensors; ++i) {
-        struct ggml_tensor * meta = model_loader->get_tensor_meta(i);
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        struct ggml_tensor * meta = ml.get_tensor_meta(i);
 
         const std::string name = ggml_get_name(meta);
 
@@ -3476,6 +7164,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             ++n_feed_forward_w2;
         }
     }
+    if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) {
+        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
+                __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer);
+    }
 
     int i_attention_wv = 0;
     int i_feed_forward_w2 = 0;
@@ -3486,24 +7178,23 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     std::vector<int64_t> hist_all(1 << 4, 0);
 
     std::vector<std::thread> workers;
+    workers.reserve(nthread);
     std::mutex mutex;
 
-    auto use_more_bits = [] (int i_layer, int num_layers) -> bool {
-        return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
-    };
-
     int idx = 0;
 
-    std::vector<uint8_t> read_data;
-    std::vector<uint8_t> work;
+    std::vector<no_init<uint8_t>> read_data;
+    std::vector<no_init<uint8_t>> work;
+    std::vector<no_init<float>> f32_conv_buf;
 
     // populate the original tensors so we get an initial meta data
-    for (int i = 0; i < model_loader->n_tensors; ++i) {
-        struct ggml_tensor * meta = model_loader->get_tensor_meta(i);
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        struct ggml_tensor * meta = ml.get_tensor_meta(i);
         gguf_add_tensor(ctx_out, meta);
     }
 
     std::ofstream fout(fname_out, std::ios::binary);
+    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
 
     const size_t meta_size = gguf_get_meta_size(ctx_out);
 
@@ -3512,17 +7203,21 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     // placeholder for the meta data
     ::zeros(fout, meta_size);
 
-    for (int i = 0; i < model_loader->n_tensors; ++i) {
-        struct ggml_tensor * tensor = model_loader->get_tensor_meta(i);
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        struct ggml_tensor * tensor = ml.get_tensor_meta(i);
 
         const std::string name = ggml_get_name(tensor);
 
-        read_data.resize(ggml_nbytes(tensor));
-        tensor->data = read_data.data();
-        model_loader->load_data_for(tensor);
+        if (!ml.use_mmap) {
+            if (read_data.size() < ggml_nbytes(tensor)) {
+                read_data.resize(ggml_nbytes(tensor));
+            }
+            tensor->data = read_data.data();
+        }
+        ml.load_data_for(tensor);
 
         LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
-               ++idx, model_loader->n_tensors,
+               ++idx, ml.n_tensors,
                ggml_get_name(tensor),
                llama_format_tensor_shape(tensor).c_str(),
                ggml_type_name(tensor->type));
@@ -3533,105 +7228,50 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         // quantize only 2D tensors
         quantize &= (tensor->n_dims == 2);
         quantize &= params->quantize_output_tensor || name != "output.weight";
-        quantize &= quantized_type != tensor->type;
+        quantize &= !params->only_copy;
 
         enum ggml_type new_type;
         void * new_data;
         size_t new_size;
 
+        if (quantize) {
+            new_type = quantized_type;
+#ifdef GGML_USE_K_QUANTS
+            new_type = get_k_quant_type(
+                new_type, tensor, model, ftype, &i_attention_wv, n_attention_wv, &i_feed_forward_w2, n_feed_forward_w2
+            );
+#endif
+            // If we've decided to quantize to the same type the tensor is already
+            // in then there's nothing to do.
+            quantize = tensor->type != new_type;
+        }
         if (!quantize) {
             new_type = tensor->type;
             new_data = tensor->data;
             new_size = ggml_nbytes(tensor);
             LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
         } else {
-            new_type = quantized_type;
-#ifdef GGML_USE_K_QUANTS
-            // TODO: avoid hardcoded tensor names - use the TN_* constants
-            if (name == TN_OUTPUT) {
-                int nx = tensor->ne[0];
-                int ny = tensor->ne[1];
-                if (nx % QK_K == 0 && ny % QK_K == 0) {
-                    new_type = GGML_TYPE_Q6_K;
-                }
-            } else if (name.find("attn_v.weight") != std::string::npos) {
-                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-                    new_type = i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
-                }
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-                else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                        use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
-                else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
-                        (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
-                ++i_attention_wv;
-            } else if (name.find("ffn_down.weight") != std::string::npos) {
-                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-                    new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
-                }
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-                else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
-                         use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < 4) new_type = GGML_TYPE_Q5_K;
-                ++i_feed_forward_w2;
-            } else if (name.find("attn_output.weight") != std::string::npos) {
-                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
-                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
-            }
-            else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
-                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-            }
-            // This can be used to reduce the size of the Q5_K_S model.
-            // The associated PPL increase is fully in line with the size reduction
-            //else {
-            //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
-            //}
-            bool convert_incompatible_tensor = false;
-            if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
-                new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
-                int nx = tensor->ne[0];
-                int ny = tensor->ne[1];
-                if (nx % QK_K != 0 || ny % QK_K != 0) {
-                    LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
-                    convert_incompatible_tensor = true;
-                }
-            }
-            if (convert_incompatible_tensor) {
-                if (name == TN_OUTPUT) {
-                    new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
-                    LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
-                } else if (name == TN_TOKEN_EMBD) {
-                    new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
-                    LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
-                } else {
-                    throw std::runtime_error("Unsupported tensor size encountered\n");
-                }
-            }
-#endif
-
             const size_t nelements = ggml_nelements(tensor);
 
             float * f32_data;
-            std::vector<float> f32_conv_buf;
 
             if (tensor->type == GGML_TYPE_F32) {
                 f32_data = (float *) tensor->data;
             } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
                 throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
             } else {
-                llama_convert_tensor_internal(tensor, f32_conv_buf, nelements, nthread);
+                llama_convert_tensor_internal(tensor, f32_conv_buf, workers, nelements, nthread);
                 f32_data = (float *) f32_conv_buf.data();
             }
 
             LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
-            work.resize(nelements * 4); // upper bound on size
+            if (work.size() < nelements * 4) {
+                work.resize(nelements * 4); // upper bound on size
+            }
             new_data = work.data();
-            std::vector<int64_t> hist_cur(1 << 4, 0);
+            std::array<int64_t, 1 << 4> hist_cur = {};
 
             static const int chunk_size = 32 * 512;
             const int nchunk = (nelements + chunk_size - 1)/chunk_size;
@@ -3642,13 +7282,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 size_t counter = 0;
                 new_size = 0;
                 auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements]() {
-                    std::vector<int64_t> local_hist;
+                    std::array<int64_t, 1 << 4> local_hist = {};
                     size_t local_size = 0;
                     while (true) {
                         std::unique_lock<std::mutex> lock(mutex);
                         size_t first = counter; counter += chunk_size;
                         if (first >= nelements) {
-                            if (!local_hist.empty()) {
+                            if (local_size > 0) {
                                 for (int j=0; j<int(local_hist.size()); ++j) {
                                     hist_cur[j] += local_hist[j];
                                 }
@@ -3658,22 +7298,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                         }
                         lock.unlock();
                         size_t last = std::min(nelements, first + chunk_size);
-                        if (local_hist.empty()) {
-                            local_hist.resize(hist_cur.size(), 0);
-                        }
                         local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data());
                     }
                 };
-                if ((int) workers.size() < nthread_use - 1) {
-                    workers.resize(nthread_use - 1);
-                }
                 for (int it = 0; it < nthread_use - 1; ++it) {
-                    workers[it] = std::thread(compute);
+                    workers.emplace_back(compute);
                 }
                 compute();
-                for (int it = 0; it < nthread_use - 1; ++it) {
-                    workers[it].join();
-                }
+                for (auto & w : workers) { w.join(); }
+                workers.clear();
             }
 
             LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
@@ -3734,8 +7367,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     }
 }
 
-// TODO: after the GGUF PR, this likely won't work and needs to be updated
-int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) {
+static int llama_apply_lora_from_file_internal(
+    const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
+) {
     LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
 
     const int64_t t_start_lora_us = ggml_time_us();
@@ -3763,7 +7397,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     int32_t lora_alpha;
     fin.read((char *) &lora_r, sizeof(lora_r));
     fin.read((char *) &lora_alpha, sizeof(lora_alpha));
-    float scaling = (float)lora_alpha / (float)lora_r;
+    float scaling = scale * (float)lora_alpha / (float)lora_r;
 
     LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
 
@@ -3785,28 +7419,28 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     }
 
     // load base model
-    std::unique_ptr<llama_model_loader> model_loader;
+    std::unique_ptr<llama_model_loader> ml;
     ggml_context * base_ctx = NULL;
     std::vector<uint8_t> base_buf;
     if (path_base_model) {
         LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
-        model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
+        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true));
 
         size_t ctx_size;
         size_t mmapped_size;
-        model_loader->calc_sizes(ctx_size, mmapped_size);
+        ml->calc_sizes(ctx_size, mmapped_size);
         base_buf.resize(ctx_size);
 
         ggml_init_params base_params;
         base_params.mem_size   = base_buf.size();
         base_params.mem_buffer = base_buf.data();
-        base_params.no_alloc   = model_loader->use_mmap;
+        base_params.no_alloc   = ml->use_mmap;
 
         base_ctx = ggml_init(base_params);
 
         // maybe this should in llama_model_loader
-        if (model_loader->use_mmap) {
-            model_loader->mapping.reset(new llama_mmap(&model_loader->file, /* prefetch */ 0, ggml_is_numa()));
+        if (ml->use_mmap) {
+            ml->mapping.reset(new llama_mmap(&ml->file, /* prefetch */ 0, ggml_is_numa()));
         }
     }
 
@@ -3910,18 +7544,19 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 #endif // GGML_USE_CUBLAS
 
             ggml_tensor * base_t;
-            if (model_loader) {
-                struct gguf_context * ctx_gguf = model_loader->ctx_gguf;
+            if (ml) {
+                struct gguf_context * ctx_gguf = ml->ctx_gguf;
 
                 // load from base model
                 if (gguf_find_tensor(ctx_gguf, base_name.c_str()) < 0) {
+                    // TODO: throw
                     LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
                     return 1;
                 }
 
                 // TODO: not tested!! maybe not working!
-                base_t = model_loader->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU);
-                model_loader->load_data_for(base_t);
+                base_t = ml->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU);
+                ml->load_data_for(base_t);
             } else {
                 base_t = dest_t;
             }
@@ -3978,9 +7613,10 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
                 ggml_set_name(r, "r_cpy");
             }
 
-            struct ggml_cgraph gf = ggml_build_forward(r);
+            struct ggml_cgraph * gf = ggml_new_graph(lora_ctx);
+            ggml_build_forward_expand(gf, r);
 
-            ggml_graph_compute_helper(work_buffer, &gf, n_threads);
+            ggml_graph_compute_helper(work_buffer, gf, n_threads);
 
             // we won't need these tensors again, reset the context to save memory
             ggml_free(lora_ctx);
@@ -4009,26 +7645,37 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 //
 // interface implementation
 //
+struct llama_model_params llama_model_default_params() {
+    struct llama_model_params result = {
+        /*.n_gpu_layers                =*/ 0,
+        /*.main_gpu                    =*/ 0,
+        /*.tensor_split                =*/ nullptr,
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
+        /*.vocab_only                  =*/ false,
+        /*.use_mmap                    =*/ true,
+        /*.use_mlock                   =*/ false,
+    };
+
+#ifdef GGML_USE_METAL
+    result.n_gpu_layers = 1;
+#endif
+
+    return result;
+}
 
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
         /*.seed                        =*/ LLAMA_DEFAULT_SEED,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
-        /*.gpu_layers                  =*/ 0,
-        /*.main_gpu                    =*/ 0,
-        /*.tensor_split                =*/ nullptr,
-        /*.rope_freq_base              =*/ 10000.0f,
-        /*.rope_freq_scale             =*/ 1.0f,
-        /*.progress_callback           =*/ nullptr,
-        /*.progress_callback_user_data =*/ nullptr,
-        /*.low_vram                    =*/ false,
-        /*.mul_mat_q                   =*/ false,
+        /*.n_threads                   =*/ GGML_DEFAULT_N_THREADS, // TODO: better default
+        /*.n_threads_batch             =*/ GGML_DEFAULT_N_THREADS,
+        /*.rope_freq_base              =*/ 0.0f,
+        /*.rope_freq_scale             =*/ 0.0f,
+        /*.mul_mat_q                   =*/ true,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
-        /*.vocab_only                  =*/ false,
-        /*.use_mmap                    =*/ true,
-        /*.use_mlock                   =*/ false,
         /*.embedding                   =*/ false,
     };
 
@@ -4041,6 +7688,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
         /*.allow_requantize            =*/ false,
         /*.quantize_output_tensor      =*/ true,
+        /*.only_copy                   =*/ false,
     };
 
     return result;
@@ -4089,16 +7737,30 @@ int64_t llama_time_us(void) {
 
 struct llama_model * llama_load_model_from_file(
                              const char * path_model,
-            struct llama_context_params   params) {
+              struct llama_model_params   params) {
     ggml_time_init();
 
     llama_model * model = new llama_model;
 
-    ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    unsigned cur_percentage = 0;
+    if (params.progress_callback == NULL) {
+        params.progress_callback_user_data = &cur_percentage;
+        params.progress_callback = [](float progress, void * ctx) {
+            unsigned * cur_percentage_p = (unsigned *) ctx;
+            unsigned percentage = (unsigned) (100 * progress);
+            while (percentage > *cur_percentage_p) {
+                *cur_percentage_p = percentage;
+                LLAMA_LOG_INFO(".");
+                if (percentage >= 100) {
+                    LLAMA_LOG_INFO("\n");
+                }
+            }
+        };
+    }
 
-    if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,
-                params.low_vram, memory_type, params.use_mmap, params.use_mlock, params.vocab_only,
+    if (!llama_model_load(path_model, *model, params.n_gpu_layers,
+                params.main_gpu, params.tensor_split,
+                params.use_mmap, params.use_mlock, params.vocab_only,
                 params.progress_callback, params.progress_callback_user_data)) {
         LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
         delete model;
@@ -4122,25 +7784,24 @@ struct llama_context * llama_new_context_with_model(
 
     llama_context * ctx = new llama_context(*model);
 
+    const auto & hparams = model->hparams;
+    auto       & cparams = ctx->cparams;
+
+    cparams.n_batch         = params.n_batch;
+    cparams.n_ctx           = params.n_ctx == 0           ? hparams.n_ctx_train           : params.n_ctx;
+    cparams.rope_freq_base  = params.rope_freq_base == 0  ? hparams.rope_freq_base_train  : params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale == 0 ? hparams.rope_freq_scale_train : params.rope_freq_scale;
+    cparams.n_threads       = params.n_threads;
+    cparams.n_threads_batch = params.n_threads_batch;
+    cparams.mul_mat_q       = params.mul_mat_q;
+
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
 
-    unsigned cur_percentage = 0;
-    if (params.progress_callback == NULL) {
-        params.progress_callback_user_data = &cur_percentage;
-        params.progress_callback = [](float progress, void * ctx) {
-            unsigned * cur_percentage_p = (unsigned *) ctx;
-            unsigned percentage = (unsigned) (100 * progress);
-            while (percentage > *cur_percentage_p) {
-                *cur_percentage_p = percentage;
-                LLAMA_LOG_INFO(".");
-                if (percentage >= 100) {
-                    LLAMA_LOG_INFO("\n");
-                }
-            }
-        };
-    }
+    LLAMA_LOG_INFO("%s: n_ctx      = %u\n",     __func__, cparams.n_ctx);
+    LLAMA_LOG_INFO("%s: freq_base  = %.1f\n",   __func__, cparams.rope_freq_base);
+    LLAMA_LOG_INFO("%s: freq_scale = %g\n",     __func__, cparams.rope_freq_scale);
 
     ctx->rng = std::mt19937(params.seed);
     ctx->logits_all = params.logits_all;
@@ -4148,8 +7809,8 @@ struct llama_context * llama_new_context_with_model(
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     // reserve memory for context buffers
-    if (!params.vocab_only) {
-        if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
+    if (!hparams.vocab_only) {
+        if (!llama_kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, cparams.n_ctx, model->n_gpu_layers)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
@@ -4160,11 +7821,9 @@ struct llama_context * llama_new_context_with_model(
             LLAMA_LOG_INFO("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
         }
 
-        const auto & hparams = ctx->model.hparams;
-
         // resized during inference
         if (params.logits_all) {
-            ctx->logits.reserve(hparams.n_ctx*hparams.n_vocab);
+            ctx->logits.reserve(cparams.n_ctx*hparams.n_vocab);
         } else {
             ctx->logits.reserve(hparams.n_vocab);
         }
@@ -4182,26 +7841,29 @@ struct llama_context * llama_new_context_with_model(
             ctx->alloc = ggml_allocr_new_measure(tensor_alignment);
 
             // build worst-case graph
-            int n_tokens = std::min((int)hparams.n_ctx, params.n_batch);
-            int n_past = hparams.n_ctx - n_tokens;
+            int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_batch);
+            int n_past = cparams.n_ctx - n_tokens;
             llama_token token = llama_token_bos(ctx); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
-            ggml_cgraph * gf = llama_build_graph(*ctx, &token, NULL, n_tokens, n_past);
+            ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0));
+
 #ifdef GGML_USE_METAL
-            if (params.n_gpu_layers > 0) {
+            if (model->n_gpu_layers > 0) {
+                ggml_metal_log_set_callback(llama_log_callback_default, NULL);
+
                 ctx->ctx_metal = ggml_metal_init(1);
                 if (!ctx->ctx_metal) {
                     LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
                     llama_free(ctx);
                     return NULL;
                 }
-                ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
-                ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
+                //ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
+                //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
             }
 #endif
             // measure memory requirements for the graph
             size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment;
 
-            LLAMA_LOG_INFO("%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
+            LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0);
 
             // recreate allocator with exact memory requirements
             ggml_allocr_free(ctx->alloc);
@@ -4210,64 +7872,81 @@ struct llama_context * llama_new_context_with_model(
             ctx->alloc = ggml_allocr_new(ctx->buf_alloc.data, ctx->buf_alloc.size, tensor_alignment);
 #ifdef GGML_USE_METAL
             if (ctx->ctx_metal) {
-                ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
+                //ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
             }
 #endif
 #ifdef GGML_USE_CUBLAS
-            if (params.low_vram) {
-                LLAMA_LOG_INFO("%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
-                ggml_cuda_set_scratch_size(0); // disable scratch
-            } else {
-                ggml_cuda_set_scratch_size(alloc_size);
-                LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MB\n", __func__, alloc_size / 1024.0 / 1024.0);
+            ggml_cuda_set_scratch_size(alloc_size);
+            LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MB\n", __func__, alloc_size / 1024.0 / 1024.0);
+
+            // calculate total VRAM usage
+            auto add_tensor = [](const ggml_tensor * t, size_t & size) {
+                if (t->backend == GGML_BACKEND_GPU || t->backend == GGML_BACKEND_GPU_SPLIT) {
+                    size += ggml_nbytes(t);
+                }
+            };
+            size_t model_vram_size = 0;
+            for (const auto & kv : model->tensors_by_name) {
+                add_tensor(kv.second, model_vram_size);
             }
+
+            size_t kv_vram_size = 0;
+            add_tensor(ctx->kv_self.k, kv_vram_size);
+            add_tensor(ctx->kv_self.v, kv_vram_size);
+
+            size_t ctx_vram_size = alloc_size + kv_vram_size;
+            size_t total_vram_size = model_vram_size + ctx_vram_size;
+
+            LLAMA_LOG_INFO("%s: total VRAM used: %.2f MB (model: %.2f MB, context: %.2f MB)\n", __func__,
+                    total_vram_size / 1024.0 / 1024.0,
+                    model_vram_size / 1024.0 / 1024.0,
+                    ctx_vram_size / 1024.0 / 1024.0);
 #endif
         }
-    }
 
 #ifdef GGML_USE_METAL
-    if (params.n_gpu_layers > 0) {
-        // this allocates all Metal resources and memory buffers
+        if (model->n_gpu_layers > 0) {
+            // this allocates all Metal resources and memory buffers
 
-        void * data_ptr  = NULL;
-        size_t data_size = 0;
+            void * data_ptr  = NULL;
+            size_t data_size = 0;
 
-        if (params.use_mmap) {
-            data_ptr  = ctx->model.mapping->addr;
-            data_size = ctx->model.mapping->size;
-        } else {
-            data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
-            data_size = ggml_get_mem_size  (ctx->model.ctx);
-        }
+            if (ctx->model.mapping) {
+                data_ptr  = ctx->model.mapping->addr;
+                data_size = ctx->model.mapping->size;
+            } else {
+                data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
+                data_size = ggml_get_mem_size  (ctx->model.ctx);
+            }
 
-        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+            const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
 
-        LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+            LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
 
 #define LLAMA_METAL_CHECK_BUF(result)                            \
-    if (!(result)) {                                             \
-        LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
-        llama_free(ctx);                                         \
-        return NULL;                                             \
-    }
+            if (!(result)) {                                             \
+                LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
+                llama_free(ctx);                                         \
+                return NULL;                                             \
+            }
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
-
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
-
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data",  data_ptr, data_size, max_size));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",    ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
+            LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
 #undef LLAMA_METAL_CHECK_BUF
-    }
+        }
 #endif
+    }
 
 #ifdef GGML_USE_MPI
     ctx->ctx_mpi = ggml_mpi_init();
 
     if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
         // Enter a blocking eval loop with dummy input, letting rank=0 drive the process
-        const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos(ctx));
-        while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
+        // TODO: needs fix after #3228
+        GGML_ASSERT(false && "not implemented");
+        //const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos(ctx));
+        //while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
         llama_backend_free();
         exit(1);
     }
@@ -4276,49 +7955,63 @@ struct llama_context * llama_new_context_with_model(
     return ctx;
 }
 
-struct llama_context * llama_init_from_file(
-                             const char * path_model,
-            struct llama_context_params   params) {
-
-    struct llama_model * model = llama_load_model_from_file(path_model, params);
-    if (!model) {
-        return nullptr;
-    }
-    struct llama_context * ctx = llama_new_context_with_model(model, params);
-    ctx->model_owner = true;
-    return ctx;
-}
-
 void llama_free(struct llama_context * ctx) {
     delete ctx;
 }
 
-int llama_n_vocab(const struct llama_context * ctx) {
-    return ctx->model.vocab.id_to_token.size();
+const llama_model * llama_get_model(const struct llama_context * ctx) {
+    return &ctx->model;
 }
 
 int llama_n_ctx(const struct llama_context * ctx) {
-    return ctx->model.hparams.n_ctx;
+    return ctx->cparams.n_ctx;
 }
 
-int llama_n_embd(const struct llama_context * ctx) {
-    return ctx->model.hparams.n_embd;
+enum llama_vocab_type llama_vocab_type(const struct llama_model * model) {
+    return model->vocab.type;
 }
 
-int llama_model_n_vocab(const struct llama_model * model) {
+int llama_n_vocab(const struct llama_model * model) {
     return model->vocab.id_to_token.size();
 }
 
-int llama_model_n_ctx(const struct llama_model * model) {
-    return model->hparams.n_ctx;
+int llama_n_ctx_train(const struct llama_model * model) {
+    return model->hparams.n_ctx_train;
 }
 
-int llama_model_n_embd(const struct llama_model * model) {
+int llama_n_embd(const struct llama_model * model) {
     return model->hparams.n_embd;
 }
 
-int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) {
-    return snprintf(buf, buf_size, "LLaMA %s %s", llama_model_type_name(model->type), llama_model_ftype_name(model->ftype).c_str());
+float llama_rope_freq_scale_train(const struct llama_model * model) {
+    return model->hparams.rope_freq_scale_train;
+}
+
+int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) {
+    return snprintf(buf, buf_size, "%s %s %s",
+            llama_model_arch_name(model->arch).c_str(),
+            llama_model_type_name(model->type),
+            llama_model_ftype_name(model->ftype).c_str());
+}
+
+uint64_t llama_model_size(const struct llama_model * model) {
+    uint64_t size = 0;
+    for (const auto & it : model->tensors_by_name) {
+        size += ggml_nbytes(it.second);
+    }
+    return size;
+}
+
+uint64_t llama_model_n_params(const struct llama_model * model) {
+    uint64_t nparams = 0;
+    for (const auto & it : model->tensors_by_name) {
+        nparams += ggml_nelements(it.second);
+    }
+    return nparams;
+}
+
+struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {
+    return ggml_get_tensor(model->ctx, name);
 }
 
 int llama_model_quantize(
@@ -4334,18 +8027,18 @@ int llama_model_quantize(
     }
 }
 
-int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) {
+int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, float scale, const char * path_base_model, int n_threads) {
     try {
-        return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads);
+        return llama_apply_lora_from_file_internal(ctx->model, path_lora, scale, path_base_model, n_threads);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
     }
 }
 
-int llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, const char * path_base_model, int n_threads) {
+int llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int n_threads) {
     try {
-        return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads);
+        return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
         return 1;
@@ -4353,16 +8046,27 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha
 }
 
 int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
-    return ctx->kv_self.n;
+    return ctx->kv_self.head;
 }
 
-#define LLAMA_MAX_RNG_STATE (64*1024)
+void llama_kv_cache_tokens_rm(struct llama_context * ctx, int32_t c0, int32_t c1) {
+    llama_kv_cache_tokens_rm(ctx->kv_self, c0, c1);
+}
 
-void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
-    if (seed == LLAMA_DEFAULT_SEED) {
-        seed = time(NULL);
-    }
-    ctx->rng.seed(seed);
+void llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1);
+}
+
+void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    llama_kv_cache_seq_cp(ctx->kv_self, seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) {
+    llama_kv_cache_seq_keep(ctx->kv_self, seq_id);
+}
+
+void llama_kv_cache_seq_shift(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+    llama_kv_cache_seq_shift(ctx->kv_self, seq_id, p0, p1, delta);
 }
 
 // Returns the *maximum* size of the state
@@ -4449,7 +8153,7 @@ struct llama_data_file_context : llama_data_context {
  * llama_copy_state_data(ctx, &data_ctx);
  *
 */
-void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
+static void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
     // copy rng
     {
         std::stringstream rng_ss;
@@ -4500,36 +8204,40 @@ void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_conte
     {
         const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
-        const int    n_layer = hparams.n_layer;
-        const int    n_embd  = hparams.n_embd_gqa();
-        const int    n_ctx   = hparams.n_ctx;
+        const auto & cparams = ctx->cparams;
 
-        const size_t kv_size = kv_self.buf.size;
-        const int    kv_ntok = llama_get_kv_cache_token_count(ctx);
+        const auto   n_layer = hparams.n_layer;
+        const auto   n_embd  = hparams.n_embd_gqa();
+        const auto   n_ctx   = cparams.n_ctx;
 
-        data_ctx->write(&kv_size, sizeof(kv_size));
-        data_ctx->write(&kv_ntok, sizeof(kv_ntok));
+        const size_t   kv_buf_size = kv_self.buf.size;
+        const uint32_t kv_head     = kv_self.head;
+        const uint32_t kv_size     = kv_self.size;
 
-        if (kv_size) {
+        data_ctx->write(&kv_buf_size, sizeof(kv_buf_size));
+        data_ctx->write(&kv_head,     sizeof(kv_head));
+        data_ctx->write(&kv_size,     sizeof(kv_size));
+
+        if (kv_buf_size) {
             const size_t elt_size = ggml_element_size(kv_self.k);
 
             ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
 
-            ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
+            ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
             std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
             kout3d->data = kout3d_data.data();
 
-            ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer);
+            ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer);
             std::vector<uint8_t> vout3d_data(ggml_nbytes(vout3d), 0);
             vout3d->data = vout3d_data.data();
 
             ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
-                n_embd, kv_ntok, n_layer,
+                n_embd, kv_head, n_layer,
                 elt_size*n_embd, elt_size*n_embd*n_ctx, 0);
 
             ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v,
-                kv_ntok, n_embd, n_layer,
+                kv_head, n_embd, n_layer,
                 elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
 
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
@@ -4543,6 +8251,20 @@ void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_conte
             data_ctx->write(kout3d_data.data(), kout3d_data.size());
             data_ctx->write(vout3d_data.data(), vout3d_data.size());
         }
+
+        for (uint32_t i = 0; i < kv_size; ++i) {
+            const auto & cell = kv_self.cells[i];
+
+            const llama_pos pos         = cell.pos;
+            const size_t    seq_id_size = cell.seq_id.size();
+
+            data_ctx->write(&pos,         sizeof(pos));
+            data_ctx->write(&seq_id_size, sizeof(seq_id_size));
+
+            for (auto seq_id : cell.seq_id) {
+                data_ctx->write(&seq_id, sizeof(seq_id));
+            }
+        }
     }
 }
 
@@ -4569,7 +8291,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
         rng_ss.str(std::string(&rng_buf[0], rng_size));
         rng_ss >> ctx->rng;
 
-        GGML_ASSERT(rng_ss.fail() == false);
+        GGML_ASSERT(!rng_ss.fail());
     }
 
     // set logits
@@ -4608,38 +8330,42 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
     {
         const auto & kv_self = ctx->kv_self;
         const auto & hparams = ctx->model.hparams;
+        const auto & cparams = ctx->cparams;
+
         const int    n_layer = hparams.n_layer;
         const int    n_embd  = hparams.n_embd_gqa();
-        const int    n_ctx   = hparams.n_ctx;
+        const int    n_ctx   = cparams.n_ctx;
 
-        size_t kv_size;
-        int kv_ntok;
+        size_t   kv_buf_size;
+        uint32_t kv_head;
+        uint32_t kv_size;
 
-        memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size);
-        memcpy(&kv_ntok, inp, sizeof(kv_ntok)); inp += sizeof(kv_ntok);
+        memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size);
+        memcpy(&kv_head,     inp, sizeof(kv_head));     inp += sizeof(kv_head);
+        memcpy(&kv_size,     inp, sizeof(kv_size));     inp += sizeof(kv_size);
 
-        if (kv_size) {
-            GGML_ASSERT(kv_self.buf.size == kv_size);
+        if (kv_buf_size) {
+            GGML_ASSERT(kv_self.buf.size == kv_buf_size);
 
             const size_t elt_size = ggml_element_size(kv_self.k);
 
             ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
 
-            ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
+            ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
             kin3d->data = (void *) inp;
             inp += ggml_nbytes(kin3d);
 
-            ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer);
+            ggml_tensor * vin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_head, n_embd, n_layer);
             vin3d->data = (void *) inp;
             inp += ggml_nbytes(vin3d);
 
             ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k,
-                n_embd, kv_ntok, n_layer,
+                n_embd, kv_head, n_layer,
                 elt_size*n_embd, elt_size*n_embd*n_ctx, 0);
 
             ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v,
-                kv_ntok, n_embd, n_layer,
+                kv_head, n_embd, n_layer,
                 elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
 
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d));
@@ -4649,7 +8375,27 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
             ggml_free(cpy_ctx);
         }
 
-        ctx->kv_self.n = kv_ntok;
+        ctx->kv_self.head = kv_head;
+        ctx->kv_self.size = kv_size;
+
+        ctx->kv_self.cells.resize(kv_size);
+
+        for (uint32_t i = 0; i < kv_size; ++i) {
+            llama_pos pos;
+            size_t    seq_id_size;
+
+            memcpy(&pos,         inp, sizeof(pos));         inp += sizeof(pos);
+            memcpy(&seq_id_size, inp, sizeof(seq_id_size)); inp += sizeof(seq_id_size);
+
+            ctx->kv_self.cells[i].pos = pos;
+
+            llama_seq_id seq_id;
+
+            for (size_t j = 0; j < seq_id_size; ++j) {
+                memcpy(&seq_id, inp, sizeof(seq_id)); inp += sizeof(seq_id);
+                ctx->kv_self.cells[i].seq_id.insert(seq_id);
+            }
+        }
     }
 
     const size_t nread    = inp - src;
@@ -4744,64 +8490,102 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi
 
 int llama_eval(
         struct llama_context * ctx,
-           const llama_token * tokens,
-                         int   n_tokens,
-                         int   n_past,
-                         int   n_threads) {
-    if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) {
-        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
-        return 1;
+                 llama_token * tokens,
+                     int32_t   n_tokens,
+                         int   n_past) {
+    llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1);
+
+    const int ret = llama_decode_internal(*ctx, llama_batch_get_one(tokens, n_tokens, n_past, 0));
+    if (ret < 0) {
+        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
     }
 
-    // get a more accurate load time, upon first eval
-    // TODO: fix this
-    if (!ctx->has_evaluated_once) {
-        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
-        ctx->has_evaluated_once = true;
-    }
-
-    return 0;
+    return ret;
 }
 
 int llama_eval_embd(
             struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads) {
-    if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) {
-        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
-        return 1;
+                           float * embd,
+                         int32_t   n_tokens,
+                             int   n_past) {
+    llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1);
+
+    llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, n_past, 1, 0, };
+
+    const int ret = llama_decode_internal(*ctx, batch);
+    if (ret < 0) {
+        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
     }
 
-    // get a more accurate load time, upon first eval
-    // TODO: fix this
-    if (!ctx->has_evaluated_once) {
-        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
-        ctx->has_evaluated_once = true;
-    }
-
-    return 0;
+    return ret;
 }
 
-int llama_eval_export(struct llama_context * ctx, const char * fname) {
-    const int n_batch = 1;
-    const int n_ctx   = 512 - n_batch;
+void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) {
+    ctx->cparams.n_threads       = n_threads;
+    ctx->cparams.n_threads_batch = n_threads_batch;
+}
 
-    const std::vector<llama_token> tmp(n_batch, llama_token_bos(ctx));
+struct llama_batch llama_batch_get_one(
+             llama_token * tokens,
+                 int32_t   n_tokens,
+               llama_pos   pos_0,
+            llama_seq_id   seq_id) {
+    return {
+        /*n_tokens    =*/ n_tokens,
+        /*tokens      =*/ tokens,
+        /*embd        =*/ nullptr,
+        /*pos         =*/ nullptr,
+        /*seq_id      =*/ nullptr,
+        /*logits      =*/ nullptr,
+        /*all_pos_0   =*/ pos_0,
+        /*all_pos_1   =*/ 1,
+        /*all_seq_id  =*/ seq_id,
+    };
+}
 
-    if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) {
-        LLAMA_LOG_ERROR("%s: failed to eval\n", __func__);
-        return 1;
+struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd) {
+    llama_batch batch = { -1, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
+
+    if (embd) {
+        batch.embd = (float *) malloc(sizeof(float) * n_tokens * embd);
+    } else {
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens);
     }
 
-    return 0;
+    batch.pos    = (llama_pos *)    malloc(sizeof(llama_pos)    * n_tokens);
+    batch.seq_id = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_tokens);
+    batch.logits = (int8_t *)       malloc(sizeof(int8_t)       * n_tokens);
+
+    return batch;
+}
+
+void llama_batch_free(struct llama_batch batch) {
+    if (batch.token)  free(batch.token);
+    if (batch.embd)   free(batch.embd);
+    if (batch.pos)    free(batch.pos);
+    if (batch.seq_id) free(batch.seq_id);
+    if (batch.logits) free(batch.logits);
+}
+
+int llama_decode(
+        struct llama_context * ctx,
+          struct llama_batch   batch) {
+    const int ret = llama_decode_internal(*ctx, batch);
+    if (ret < 0) {
+        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
+    }
+
+    return ret;
 }
 
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }
 
+float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
+    return ctx->logits.data() + i*ctx->model.hparams.n_vocab;
+}
+
 float * llama_get_embeddings(struct llama_context * ctx) {
     return ctx->embedding.data();
 }
@@ -4829,47 +8613,34 @@ llama_token llama_token_eos(const struct llama_context * ctx) {
 llama_token llama_token_nl(const struct llama_context * ctx) {
     return ctx->model.vocab.linefeed_id;
 }
+llama_token llama_token_prefix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_prefix_id;
+}
+
+llama_token llama_token_middle(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_middle_id;
+}
+
+llama_token llama_token_suffix(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_suffix_id;
+}
+
+llama_token llama_token_eot(const struct llama_context * ctx) {
+    return ctx->model.vocab.special_eot_id;
+}
+
 
 int llama_tokenize(
-        struct llama_context * ctx,
-                  const char * text,
-                 llama_token * tokens,
-                         int   n_max_tokens,
-                        bool   add_bos) {
-    return llama_tokenize_with_model(&ctx->model, text, tokens, n_max_tokens, add_bos);
-}
-
-int llama_tokenize_bpe(
-        struct llama_context * ctx,
-                  const char * text,
-                 llama_token * tokens,
-                         int   n_max_tokens,
-                        bool   add_bos) {
-    auto res = llama_tokenize_internal(ctx->model.vocab, text, add_bos, false);
-
-    if (n_max_tokens < (int) res.size()) {
-        LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
-        return -((int) res.size());
-    }
-
-    for (size_t i = 0; i < res.size(); i++) {
-        tokens[i] = res[i];
-    }
-
-    return res.size();
-}
-
-int llama_tokenize_with_model(
     const struct llama_model * model,
                   const char * text,
+                         int   text_len,
                  llama_token * tokens,
                          int   n_max_tokens,
                         bool   add_bos) {
-    auto escape = llama_vocab_get_type(model->vocab) == LLAMA_VOCAB_TYPE_SPM;
-    auto res = llama_tokenize_internal(model->vocab, text, add_bos, escape);
+    auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos);
 
     if (n_max_tokens < (int) res.size()) {
-        LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
+        // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
         return -((int) res.size());
     }
 
@@ -4880,51 +8651,70 @@ int llama_tokenize_with_model(
     return res.size();
 }
 
-int llama_token_to_str(const struct llama_context * ctx, llama_token token, char * buf, int length) {
-    return llama_token_to_str_with_model(&ctx->model, token, buf, length);
-}
-
-int llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token, char * buf, int length) {
-    if (0 <= token && token < llama_model_n_vocab(&ctx->model)) {
-        std::string result = ctx->model.vocab.id_to_token[token].text;
-        if (length < (int) result.length()) {
-            return -result.length();
-        }
-        memcpy(buf, result.c_str(), result.length());
-        return result.length();
+static std::string llama_decode_text(const std::string & text) {
+    std::string decoded_text;
+    auto unicode_sequences = codepoints_from_utf8(text);
+    for (auto& unicode_sequence : unicode_sequences) {
+        decoded_text += unicode_to_bytes_bpe(codepoint_to_utf8(unicode_sequence));
     }
-    return 0;
+
+    return decoded_text;
 }
 
-// does not write null-terminator to str
-int llama_token_to_str_with_model(const struct llama_model * model, llama_token token, char * buf, int length) {
-    if (0 <= token && token < llama_model_n_vocab(model)) {
-        if (llama_is_normal_token(model->vocab, token)) {
-            std::string result = model->vocab.id_to_token[token].text;
-            if (llama_vocab_get_type(model->vocab) == LLAMA_VOCAB_TYPE_SPM) {
-                result = llama_unescape_whitespace(result);
+// does not write null-terminator to buf
+int llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int length) {
+    if (0 <= token && token < llama_n_vocab(model)) {
+        switch (llama_vocab_get_type(model->vocab)) {
+        case LLAMA_VOCAB_TYPE_SPM: {
+            if (llama_is_normal_token(model->vocab, token)) {
+                std::string result = model->vocab.id_to_token[token].text;
+                llama_unescape_whitespace(result);
+                if (length < (int) result.length()) {
+                    return -result.length();
+                }
+                memcpy(buf, result.c_str(), result.length());
+                return result.length();
+            } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
+                if (length < 3) {
+                    return -3;
+                }
+                memcpy(buf, "\xe2\x96\x85", 3);
+                return 3;
+            } else if (llama_is_control_token(model->vocab, token)) {
+                ;
+            } else if (llama_is_byte_token(model->vocab, token)) {
+                if (length < 1) {
+                    return -1;
+                }
+                buf[0] = llama_token_to_byte(model->vocab, token);
+                return 1;
+            } else {
+                // TODO: for now we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                // GGML_ASSERT(false);
             }
-            if (length < (int) result.length()) {
-                return -result.length();
+            break;
+        }
+        case LLAMA_VOCAB_TYPE_BPE: {
+            if (llama_is_normal_token(model->vocab, token)) {
+                std::string result = model->vocab.id_to_token[token].text;
+                result = llama_decode_text(result);
+                if (length < (int) result.length()) {
+                    return -result.length();
+                }
+                memcpy(buf, result.c_str(), result.length());
+                return result.length();
+            } else if (llama_is_control_token(model->vocab, token)) {
+                ;
+            } else {
+                // TODO: for now we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                // GGML_ASSERT(false);
             }
-            memcpy(buf, result.c_str(), result.length());
-            return result.length();
-        } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
-            if (length < 3) {
-                return -3;
-            }
-            buf[0] = '\xe2';
-            buf[1] = '\x96';
-            buf[2] = '\x85';
-            return 3;
-        } else if (llama_is_control_token(model->vocab, token)) {
-            ;
-        } else if (llama_is_byte_token(model->vocab, token)) {
-            if (length < 1) {
-                return -1;
-            }
-            buf[0] = llama_token_to_byte(model->vocab, token);
-            return 1;
+            break;
+        }
+        default:
+            GGML_ASSERT(false);
         }
     }
     return 0;
@@ -4951,14 +8741,14 @@ void llama_print_timings(struct llama_context * ctx) {
     const llama_timings timings = llama_get_timings(ctx);
 
     LLAMA_LOG_INFO("\n");
-    LLAMA_LOG_INFO("%s:        load time = %8.2f ms\n", __func__, timings.t_load_ms);
-    LLAMA_LOG_INFO("%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, timings.t_load_ms);
+    LLAMA_LOG_INFO("%s:      sample time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample);
-    LLAMA_LOG_INFO("%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
-    LLAMA_LOG_INFO("%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+    LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
-    LLAMA_LOG_INFO("%s:       total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
+    LLAMA_LOG_INFO("%s:       total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
 }
 
 void llama_reset_timings(struct llama_context * ctx) {
@@ -4985,26 +8775,53 @@ const char * llama_print_system_info(void) {
     s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
     s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
     s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
+    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
     s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
 
     return s.c_str();
 }
 
+void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) {
+    fprintf(stream, "\n");
+    fprintf(stream, "###########\n");
+    fprintf(stream, "# Timings #\n");
+    fprintf(stream, "###########\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "mst_eval: %.2f  # ms / token during generation\n",
+            1.0e-3 * ctx->t_eval_us / ctx->n_eval);
+    fprintf(stream, "mst_p_eval: %.2f  # ms / token during prompt processing\n",
+            1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval);
+    fprintf(stream, "mst_sample: %.2f  # ms / token during sampling\n",
+            1.0e-3 * ctx->t_sample_us / ctx->n_sample);
+    fprintf(stream, "n_eval: %d  # number of tokens generated (excluding the first one)\n", ctx->n_eval);
+    fprintf(stream, "n_p_eval: %d  # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
+    fprintf(stream, "n_sample: %d  # number of sampled tokens\n", ctx->n_sample);
+    fprintf(stream, "t_eval_us: %" PRId64 "  # total microseconds spent generating tokens\n", ctx->t_eval_us);
+    fprintf(stream, "t_load_us: %" PRId64 "  # total microseconds spent loading the model\n", ctx->t_load_us);
+    fprintf(stream, "t_p_eval_us: %" PRId64 "  # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
+    fprintf(stream, "t_sample_us: %" PRId64 "  # total microseconds spent sampling\n", ctx->t_sample_us);
+    fprintf(stream, "ts_eval: %.2f  # tokens / second during generation\n",
+            1.0e6 * ctx->n_eval / ctx->t_eval_us);
+    fprintf(stream, "ts_p_eval: %.2f  # tokens / second during prompt processing\n",
+            1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us);
+    fprintf(stream, "ts_sample: %.2f  # tokens / second during sampling\n",
+            1.0e6 * ctx->n_sample / ctx->t_sample_us);
+}
+
 // For internal test use
-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+) {
     return ctx->model.tensors_by_name;
 }
 
-void llama_log_set(llama_log_callback log_callback, void * user_data) {
+void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
     g_state.log_callback_user_data = user_data;
 }
 
-#if defined(_MSC_VER) && !defined(vsnprintf)
-#define vsnprintf _vsnprintf
-#endif
-
-static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) {
+static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {
     va_list args_copy;
     va_copy(args_copy, args);
     char buffer[128];
@@ -5021,14 +8838,14 @@ static void llama_log_internal_v(llama_log_level level, const char * format, va_
     va_end(args_copy);
 }
 
-static void llama_log_internal(llama_log_level level, const char * format, ...) {
+static void llama_log_internal(ggml_log_level level, const char * format, ...) {
     va_list args;
     va_start(args, format);
     llama_log_internal_v(level, format, args);
     va_end(args);
 }
 
-static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) {
+static void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) user_data;
     fputs(text, stderr);
diff --git a/llama.h b/llama.h
index 7ce478d54..a78015ada 100644
--- a/llama.h
+++ b/llama.h
@@ -10,6 +10,7 @@
 #endif // GGML_USE_CUBLAS
 #include <stddef.h>
 #include <stdint.h>
+#include <stdio.h>
 #include <stdbool.h>
 
 #ifdef LLAMA_SHARED
@@ -36,10 +37,12 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
+#define LLAMA_MAX_RNG_STATE (64*1024)
+
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 1
+#define LLAMA_SESSION_VERSION 2
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -59,13 +62,9 @@ extern "C" {
     struct llama_model;
     struct llama_context;
 
-    typedef int llama_token;
-
-    enum llama_log_level {
-        LLAMA_LOG_LEVEL_ERROR = 2,
-        LLAMA_LOG_LEVEL_WARN  = 3,
-        LLAMA_LOG_LEVEL_INFO  = 4
-    };
+    typedef int32_t llama_pos;
+    typedef int32_t llama_token;
+    typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
         LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
@@ -85,24 +84,24 @@ extern "C" {
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K          = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K          = 18,// except 1d tensors
+        LLAMA_FTYPE_MOSTLY_F16           = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,  // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5,  // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6,  // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -121,41 +120,68 @@ extern "C" {
 
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
-    struct llama_context_params {
-        uint32_t seed;         // RNG seed, -1 for random
-        int32_t  n_ctx;        // text context
-        int32_t  n_batch;      // prompt processing batch size
-        int32_t  n_gpu_layers; // number of layers to store in VRAM
-        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
+    // Input data for llama_decode
+    // A llama_batch object can contain input about one or many sequences
+    // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+    //
+    // - token  : the token ids of the input (used when embd is NULL)
+    // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+    // - pos    : the positions of the respective token in the sequence
+    // - seq_id : the sequence to which the respective token belongs
+    // - logits : if zero, the logits for the respective token will not be output
+    //
+    typedef struct llama_batch {
+        int32_t n_tokens;
 
+        llama_token  * token;
+        float        * embd;
+        llama_pos    * pos;
+        llama_seq_id * seq_id;
+        int8_t       * logits;
+
+        // NOTE: helpers for smooth API transition - can be deprecated in the future
+        //       for future-proof code, use the above fields instead and ignore everything below
+        //
+        // pos[i] = all_pos_0 + i*all_pos_1
+        //
+        llama_pos    all_pos_0;  // used if pos == NULL
+        llama_pos    all_pos_1;  // used if pos == NULL
+        llama_seq_id all_seq_id; // used if seq_id == NULL
+    } llama_batch;
+
+    struct llama_model_params {
+        int32_t n_gpu_layers; // number of layers to store in VRAM
+        int32_t main_gpu;     // the GPU that is used for scratch and small tensors
         const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
-        float    rope_freq_base;  // RoPE base frequency
-        float    rope_freq_scale; // RoPE frequency scaling factor
-
         // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool low_vram;   // if true, reduce VRAM usage at the cost of performance
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
-        bool f16_kv;     // use fp16 for KV cache
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
-        bool embedding;  // embedding mode only
     };
 
-    // Signature for logging events
-    // Note that text includes the new line character at the end for most events.
-    // If your logging mechanism cannot handle that, check if the last character is '\n' and strip it
-    // if it exists.
-    // It might not exist for progress report where '.' is output repeatedly.
-    typedef void (*llama_log_callback)(enum llama_log_level level, const char * text, void * user_data);
+    struct llama_context_params {
+        uint32_t seed;            // RNG seed, -1 for random
+        uint32_t n_ctx;           // text context, 0 = from model
+        uint32_t n_batch;         // prompt processing maximum batch size
+        uint32_t n_threads;       // number of threads to use for generation
+        uint32_t n_threads_batch; // number of threads to use for batch processing
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float rope_freq_base;  // RoPE base frequency, 0 = from model
+        float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
+        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool embedding;  // embedding mode only
+    };
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
@@ -163,6 +189,7 @@ extern "C" {
         enum llama_ftype ftype;      // quantize to this llama_ftype
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
     } llama_model_quantize_params;
 
     // grammar types
@@ -213,6 +240,8 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // Helpers for getting default parameters
+    LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
     LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
 
@@ -226,7 +255,7 @@ extern "C" {
 
     LLAMA_API struct llama_model * llama_load_model_from_file(
                              const char * path_model,
-            struct llama_context_params   params);
+            struct llama_model_params     params);
 
     LLAMA_API void llama_free_model(struct llama_model * model);
 
@@ -243,16 +272,30 @@ extern "C" {
     LLAMA_API bool llama_mmap_supported (void);
     LLAMA_API bool llama_mlock_supported(void);
 
-    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
+    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
 
-    LLAMA_API int llama_model_n_vocab(const struct llama_model * model);
-    LLAMA_API int llama_model_n_ctx  (const struct llama_model * model);
-    LLAMA_API int llama_model_n_embd (const struct llama_model * model);
+    LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
+
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+
+    LLAMA_API int llama_n_vocab    (const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int llama_n_embd     (const struct llama_model * model);
+
+    // Get the model's RoPE frequency scaling factor
+    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
     // Get a string describing the model type
-    LLAMA_API int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size);
+    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+
+    // Returns the total size of all the tensors in the model in bytes
+    LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+
+    // Returns the total number of parameters in the model
+    LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
+
+    // Get a llama model tensor
+    LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
 
     // Returns 0 on success
     LLAMA_API int llama_model_quantize(
@@ -269,21 +312,73 @@ extern "C" {
     LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
             struct llama_context * ctx,
                       const char * path_lora,
+                           float   scale,
                       const char * path_base_model,
                              int   n_threads),
-            "please use llama_model_apply_lora_from_file instead");
+            "use llama_model_apply_lora_from_file instead");
 
     LLAMA_API int llama_model_apply_lora_from_file(
             const struct llama_model * model,
-                          const char * path_lora,
-                          const char * path_base_model,
-                                 int   n_threads);
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    //
+    // KV cache
+    //
 
     // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
 
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+    // Remove all tokens data of cells in [c0, c1)
+    // c0 < 0 : [0,  c1]
+    // c1 < 0 : [c0, inf)
+    LLAMA_API void llama_kv_cache_tokens_rm(
+            struct llama_context * ctx,
+                         int32_t   c0,
+                         int32_t   c1);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_shift(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta);
+
+    //
+    // State / sessions
+    //
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens
@@ -292,48 +387,102 @@ extern "C" {
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
     // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+    LLAMA_API size_t llama_copy_state_data(
+            struct llama_context * ctx,
+                         uint8_t * dst);
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(
+            struct llama_context * ctx,
+                         uint8_t * src);
 
     // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    LLAMA_API bool llama_load_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+                     llama_token * tokens_out,
+                          size_t   n_token_capacity,
+                          size_t * n_token_count_out);
 
-    // Run the llama inference to obtain the logits and probabilities for the next token.
+    LLAMA_API bool llama_save_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+               const llama_token * tokens,
+                          size_t   n_token_count);
+
+    //
+    // Decoding
+    //
+
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls
     // Returns 0 on success
-    LLAMA_API int llama_eval(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
             struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
+                     llama_token * tokens,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
 
     // Same as llama_eval, but use float matrix input directly.
-    LLAMA_API int llama_eval_embd(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
             struct llama_context * ctx,
-                     const float * embd,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
+                           float * embd,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
 
-    // Export a static computation graph for context of 511 and batch size of 1
-    // NOTE: since this functionality is mostly for debugging and demonstration purposes, we hardcode these
-    //       parameters here to keep things simple
-    // IMPORTANT: do not use for anything else other than debugging and testing!
-    LLAMA_API int llama_eval_export(struct llama_context * ctx, const char * fname);
+    // Return batch for single sequence of tokens starting at pos_0
+    //
+    // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
+    //
+    LLAMA_API struct llama_batch llama_batch_get_one(
+                  llama_token * tokens,
+                      int32_t   n_tokens,
+                    llama_pos   pos_0,
+                 llama_seq_id   seq_id);
+
+    // Allocates a batch of tokens on the heap
+    // The batch has to be freed with llama_batch_free()
+    // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
+    // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
+    // The rest of the llama_batch members are allocated with size n_tokens
+    // All members are left uninitialized
+    LLAMA_API struct llama_batch llama_batch_init(
+            int32_t n_tokens,
+            int32_t embd);
+
+    // Frees a batch of tokens allocated with llama_batch_init()
+    LLAMA_API void llama_batch_free(struct llama_batch batch);
+
+    // Positive return values does not mean a fatal error, but rather a warning.
+    //   0 - success
+    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    // < 0 - error
+    LLAMA_API int llama_decode(
+            struct llama_context * ctx,
+              struct llama_batch   batch);
+
+    // Set the number of threads used for decoding
+    // n_threads is the number of threads used for generation (single token)
+    // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
+    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
+    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
+
     // Get the embeddings for the input
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
@@ -346,12 +495,17 @@ extern "C" {
 
     LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token);
 
-    LLAMA_API llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx);  // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx);  // end-of-sentence
     LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx);  // next-line
+    // codellama infill tokens
+    LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_context * ctx); // End of infill middle
 
     //
     // Tokenization
@@ -362,41 +516,18 @@ extern "C" {
     // Returns the number of tokens on success, no more than n_max_tokens
     // Returns a negative number on failure - the number of tokens that would have been returned
     LLAMA_API int llama_tokenize(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_tokenize_bpe(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_tokenize_with_model(
         const struct llama_model * model,
                       const char * text,
+                             int   text_len,
                      llama_token * tokens,
                              int   n_max_tokens,
                             bool   add_bos);
 
-    // Token Id -> String. Uses the vocabulary in the provided context
-    // Does not write null terminator to the buffer
-    LLAMA_API int llama_token_to_str(
-            const struct llama_context * ctx,
-                           llama_token   token,
-                                  char * buf,
-                                  int    length);
-
-    LLAMA_API int llama_token_to_str_bpe(
-            const struct llama_context * ctx,
-                           llama_token   token,
-                                  char * buf,
-                                  int    length);
-
-    LLAMA_API int llama_token_to_str_with_model(
+    // Token Id -> Piece.
+    // Uses the vocabulary in the provided context.
+    // Does not write null terminator to the buffer.
+    // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
+    LLAMA_API int llama_token_to_piece(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
@@ -413,15 +544,31 @@ extern "C" {
 
     LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
 
+    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+
     //
     // Sampling functions
     //
 
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
+
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
+    LLAMA_API void llama_sample_repetition_penalty(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                          float    penalty);
 
     /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   last_tokens_size,
+                           float   alpha_frequency,
+                           float   alpha_presence);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
@@ -434,23 +581,54 @@ extern "C" {
                              float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API void llama_sample_softmax(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+    LLAMA_API void llama_sample_top_k(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                             int   k,
+                          size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_top_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+    LLAMA_API void llama_sample_tail_free(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   z,
+                          size_t   min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+    LLAMA_API void llama_sample_typical(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    LLAMA_API void llama_sample_temp(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   temp);
+
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");
 
     /// @details Apply constraints from grammar
-    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+    LLAMA_API void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -458,26 +636,89 @@ extern "C" {
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                             int   m,
+                           float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                           float * mu);
 
     /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token_greedy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Accepts the sampled token into the grammar
-    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+    LLAMA_API void llama_grammar_accept_token(
+            struct llama_context * ctx,
+            struct llama_grammar * grammar,
+                     llama_token   token);
+
+    //
+    // Beam search
+    //
+
+    struct llama_beam_view {
+        const llama_token * tokens;
+
+        size_t n_tokens;
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
+    };
+
+    // Passed to beam_search_callback function.
+    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
+    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
+    // These pointers are valid only during the synchronous callback, so should not be saved.
+    struct llama_beams_state {
+        struct llama_beam_view * beam_views;
+
+        size_t n_beams;               // Number of elements in beam_views[].
+        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
+        bool   last_call;             // True iff this is the last callback invocation.
+    };
+
+    // Type of pointer to the beam_search_callback function.
+    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
+    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
+    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state);
+
+    /// @details Deterministically returns entire sentence constructed by a beam search.
+    /// @param ctx Pointer to the llama_context.
+    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
+    /// @param callback_data A pointer that is simply passed back to callback.
+    /// @param n_beams Number of beams to use.
+    /// @param n_past Number of tokens already evaluated.
+    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                    int   n_past,
+                                    int   n_predict);
 
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
     LLAMA_API void llama_reset_timings(struct llama_context * ctx);
 
@@ -486,7 +727,9 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data);
+    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
+
+    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
 
 #ifdef __cplusplus
 }
@@ -500,7 +743,9 @@ extern "C" {
 
 struct ggml_tensor;
 
-const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);
 
 #endif // LLAMA_API_INTERNAL
 
diff --git a/models/ggml-vocab-aquila.gguf b/models/ggml-vocab-aquila.gguf
new file mode 100644
index 000000000..7a9abb122
Binary files /dev/null and b/models/ggml-vocab-aquila.gguf differ
diff --git a/models/ggml-vocab-falcon.gguf b/models/ggml-vocab-falcon.gguf
new file mode 100644
index 000000000..d4ea2e822
Binary files /dev/null and b/models/ggml-vocab-falcon.gguf differ
diff --git a/mypy.ini b/mypy.ini
new file mode 100644
index 000000000..55c168f2d
--- /dev/null
+++ b/mypy.ini
@@ -0,0 +1,5 @@
+[mypy]
+strict = true
+allow_untyped_calls = true
+allow_untyped_defs = true
+allow_incomplete_defs = true
diff --git a/pocs/vdot/q8dot.cpp b/pocs/vdot/q8dot.cpp
index 4e0e02357..111770d55 100644
--- a/pocs/vdot/q8dot.cpp
+++ b/pocs/vdot/q8dot.cpp
@@ -43,7 +43,7 @@ static_assert(QK4_1 == QK8_0, "QK4_1 and QK8_0 must be the same");
 static_assert(QK4_0 == QK8_0, "QK4_0 and QK8_0 must be the same");
 
 template <typename T>
-void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
+static void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
     for (auto& b : blocks) {
         b.d = 1;
         for (int i=0; i<QK4_1/2; ++i) {
@@ -54,7 +54,7 @@ void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
     }
 }
 
-void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
+static void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
     for (auto& b : blocks) {
         b.d = 1;
         int sum = 0;
@@ -66,7 +66,7 @@ void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
     }
 }
 
-float simpleDot(const block_q4_0& x, const block_q8_0& y) {
+static float simpleDot(const block_q4_0& x, const block_q8_0& y) {
     int s1 = 0; //, s2 = 0;
     for (int i=0; i<QK4_1/2; i+=2) {
         int v1 = x.qs[i+0] & 0xf;
@@ -81,7 +81,7 @@ float simpleDot(const block_q4_0& x, const block_q8_0& y) {
     //return y.d * x.d * (s1 - 8 * s2);
 }
 
-float simpleDot(const block_q4_1& x, const block_q8_0& y) {
+static float simpleDot(const block_q4_1& x, const block_q8_0& y) {
     int s1 = 0; //, s2 = 0;
     for (int i=0; i<QK4_1/2; i+=2) {
         int v1 = x.qs[i+0] & 0xf;
diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp
index 48758cda8..e96372c4b 100644
--- a/pocs/vdot/vdot.cpp
+++ b/pocs/vdot/vdot.cpp
@@ -16,7 +16,7 @@
 
 constexpr int kVecSize = 1 << 18;
 
-float drawFromGaussianPdf(std::mt19937& rndm) {
+static float drawFromGaussianPdf(std::mt19937& rndm) {
     constexpr double kScale = 1./(1. + std::mt19937::max());
     constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale;
     static float lastX;
@@ -28,7 +28,8 @@ float drawFromGaussianPdf(std::mt19937& rndm) {
     haveX = true;
     return r*cos(phi);
 }
-void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
+
+static void fillRandomGaussianFloats(std::vector<float>& values, std::mt19937& rndm, float mean = 0) {
     for (auto& v : values) v = mean + drawFromGaussianPdf(rndm);
 }
 
diff --git a/prompts/LLM-questions.txt b/prompts/LLM-questions.txt
new file mode 100644
index 000000000..fdf3d52f4
--- /dev/null
+++ b/prompts/LLM-questions.txt
@@ -0,0 +1,49 @@
+In the context of LLMs, what is "Attention"?
+In the context of LLMs, what is a completion?
+In the context of LLMs, what is a prompt?
+In the context of LLMs, what is GELU?
+In the context of LLMs, what is RELU?
+In the context of LLMs, what is softmax?
+In the context of LLMs, what is decoding?
+In the context of LLMs, what is encoding?
+In the context of LLMs, what is tokenizing?
+In the context of LLMs, what is an embedding?
+In the context of LLMs, what is quantization?
+In the context of LLMs, what is a tensor?
+In the context of LLMs, what is a sparse tensor?
+In the context of LLMs, what is a vector?
+In the context of LLMs, how is attention implemented?
+In the context of LLMs, why is attention all you need?
+In the context of LLMs, what is "RoPe" and what is it used for?
+In the context of LLMs, what is "LoRA" and what is it used for?
+In the context of LLMs, what are weights?
+In the context of LLMs, what are biases?
+In the context of LLMs, what are checkpoints?
+In the context of LLMs, what is "perplexity"?
+In the context of LLMs, what are models?
+In the context of machine-learning, what is "catastrophic forgetting"?
+In the context of machine-learning, what is "elastic weight consolidation (EWC)"?
+In the context of neural nets, what is a hidden layer?
+In the context of neural nets, what is a convolution?
+In the context of neural nets, what is dropout?
+In the context of neural nets, what is cross-entropy?
+In the context of neural nets, what is over-fitting?
+In the context of neural nets, what is under-fitting?
+What is the difference between an interpreted computer language and a compiled computer language?
+In the context of software development, what is a debugger?
+When processing using a GPU, what is off-loading?
+When processing using a GPU, what is a batch?
+When processing using a GPU, what is a block?
+When processing using a GPU, what is the difference between a batch and a block?
+When processing using a GPU, what is a scratch tensor?
+When processing using a GPU, what is a layer?
+When processing using a GPU, what is a cache?
+When processing using a GPU, what is unified memory?
+When processing using a GPU, what is VRAM?
+When processing using a GPU, what is a kernel?
+When processing using a GPU, what is "metal"?
+In the context of LLMs, what are "Zero-Shot", "One-Shot" and "Few-Shot" learning models?
+In the context of LLMs, what is the "Transformer-model" architecture?
+In the context of LLMs, what is "Multi-Head Attention"?
+In the context of LLMs, what is "Self-Attention"?
+In the context of transformer-model architectures, how do attention mechanisms use masks?
\ No newline at end of file
diff --git a/prompts/chat-with-baichuan.txt b/prompts/chat-with-baichuan.txt
new file mode 100644
index 000000000..11626b692
--- /dev/null
+++ b/prompts/chat-with-baichuan.txt
@@ -0,0 +1,4 @@
+以下内容为人类用户与与一位智能助手的对话。
+
+用户:你好！
+助手:
diff --git a/prompts/parallel-questions.txt b/prompts/parallel-questions.txt
new file mode 100644
index 000000000..c9fc7b8b4
--- /dev/null
+++ b/prompts/parallel-questions.txt
@@ -0,0 +1,43 @@
+What do you know about Hobbits?
+What is quantum field theory?
+Why did the chicken cross the road?
+Who is the president of the United States?
+How do I run CMake on MacOS?
+Do you agree that C++ is a really finicky language compared with Python3?
+Is it a good idea to invest in technology?
+Do you like Wagner's Ring?
+Do you think this file input option is really neat?
+What should we all do about climate change?
+Is time-travel possible within the laws of current physics?
+Is it like anything to be a bat?
+Once the chicken has crossed the road, does it try to go back?
+Who is the greatest of all musical composers?
+What is art?
+Is there life elsewhere in the universe?
+What is intelligence?
+What is the difference between knowledge and intelligence?
+Will religion ever die?
+Do we understand ourselves?
+What is the best way to cook eggs?
+If you cannot see things, on what basis do you evaluate them?
+Explain the role of the np junction in photovoltaic cells?
+Is professional sport a good or bad influence on human behaviour?
+Is capital punishment immoral?
+Should we care about other people?
+Who are you?
+Which sense would you surrender if you could?
+Was Henry Ford a hero or a villain?
+Do we need leaders?
+What is nucleosynthesis?
+Who is the greatest scientist of all time?
+Who first observed what came to be known as the photovoltaic effect?
+What is nuclear fusion and why does it release energy?
+Can you know that you exist?
+What is an exoplanet?
+Do you like cream?
+What is the difference?
+Can I know that I exist while I'm dreaming that I'm Descartes?
+Who said "I didn't know I thought that until I heard myself saying it"?
+Does anything really matter?
+Can you explain the unreasonable effectiveness of mathematics?
+
diff --git a/requirements.txt b/requirements.txt
index 6c32cbd04..81c909d0b 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,2 +1,3 @@
-numpy==1.24
+numpy==1.24.4
 sentencepiece==0.1.98
+gguf>=0.1.0
diff --git a/run_with_preset.py b/run_with_preset.py
new file mode 100755
index 000000000..9b4d7ecbe
--- /dev/null
+++ b/run_with_preset.py
@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import subprocess
+import sys
+
+import yaml
+
+CLI_ARGS_MAIN_PERPLEXITY = [
+    "batch-size", "cfg-negative-prompt", "cfg-scale", "chunks", "color", "ctx-size", "escape",
+    "export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag",
+    "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix", "instruct",
+    "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
+    "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
+    "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
+    "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
+    "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "random-prompt", "repeat-last-n",
+    "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed",
+    "simple-io", "tensor-split", "threads", "temp", "tfs", "top-k", "top-p", "typical",
+    "verbose-prompt"
+]
+
+CLI_ARGS_LLAMA_BENCH = [
+    "batch-size", "memory-f32", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
+    "n-prompt", "output", "repetitions", "tensor-split", "threads", "verbose"
+]
+
+CLI_ARGS_SERVER = [
+    "alias", "batch-size", "ctx-size", "embedding", "host", "memory-f32", "lora", "lora-base",
+    "low-vram", "main-gpu", "mlock", "model", "n-gpu-layers", "n-probs", "no-mmap", "no-mul-mat-q",
+    "numa", "path", "port", "rope-freq-base", "timeout", "rope-freq-scale", "tensor-split",
+    "threads", "verbose"
+]
+
+description = """Run llama.cpp binaries with presets from YAML file(s).
+To specify which binary should be run, specify the "binary" property (main, perplexity, llama-bench, and server are supported).
+To get a preset file template, run a llama.cpp binary with the "--logdir" CLI argument.
+
+Formatting considerations:
+- The YAML property names are the same as the CLI argument names of the corresponding binary.
+- Properties must use the long name of their corresponding llama.cpp CLI arguments.
+- Like the llama.cpp binaries the property names do not differentiate between hyphens and underscores.
+- Flags must be defined as "<PROPERTY_NAME>: true" to be effective.
+- To define the logit_bias property, the expected format is "<TOKEN_ID>: <BIAS>" in the "logit_bias" namespace.
+- To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings.
+- To define a tensor split, pass a list of floats.
+"""
+usage = "run_with_preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
+epilog = ("  --<ARG_NAME> specify additional CLI ars to be passed to the binary (override all preset files). "
+          "Unknown args will be ignored.")
+
+parser = argparse.ArgumentParser(
+    description=description, usage=usage, epilog=epilog, formatter_class=argparse.RawTextHelpFormatter)
+parser.add_argument("-bin", "--binary", help="The binary to run.")
+parser.add_argument("yaml_files", nargs="*",
+                    help="Arbitrary number of YAML files from which to read preset values. "
+                    "If two files specify the same values the later one will be used.")
+
+known_args, unknown_args = parser.parse_known_args()
+
+if not known_args.yaml_files and not unknown_args:
+    parser.print_help()
+    sys.exit(0)
+
+props = dict()
+
+for yaml_file in known_args.yaml_files:
+    with open(yaml_file, "r") as f:
+        props.update(yaml.load(f, yaml.SafeLoader))
+
+props = {prop.replace("_", "-"): val for prop, val in props.items()}
+
+binary = props.pop("binary", "main")
+if known_args.binary:
+    binary = known_args.binary
+
+if os.path.exists(f"./{binary}"):
+    binary = f"./{binary}"
+
+if binary.lower().endswith("main") or binary.lower().endswith("perplexity"):
+    cli_args = CLI_ARGS_MAIN_PERPLEXITY
+elif binary.lower().endswith("llama-bench"):
+    cli_args = CLI_ARGS_LLAMA_BENCH
+elif binary.lower().endswith("server"):
+    cli_args = CLI_ARGS_SERVER
+else:
+    print(f"Unknown binary: {binary}")
+    sys.exit(1)
+
+command_list = [binary]
+
+for cli_arg in cli_args:
+    value = props.pop(cli_arg, None)
+
+    if not value or value == -1:
+        continue
+
+    if cli_arg == "logit-bias":
+        for token, bias in value.items():
+            command_list.append("--logit-bias")
+            command_list.append(f"{token}{bias:+}")
+        continue
+
+    if cli_arg == "reverse-prompt" and not isinstance(value, str):
+        for rp in value:
+            command_list.append("--reverse-prompt")
+            command_list.append(str(rp))
+        continue
+
+    command_list.append(f"--{cli_arg}")
+
+    if cli_arg == "tensor-split":
+        command_list.append(",".join([str(v) for v in value]))
+        continue
+
+    value = str(value)
+
+    if value != "True":
+        command_list.append(str(value))
+
+num_unused = len(props)
+if num_unused > 10:
+    print(f"The preset file contained a total of {num_unused} unused properties.")
+elif num_unused > 0:
+    print("The preset file contained the following unused properties:")
+    for prop, value in props.items():
+        print(f"  {prop}: {value}")
+
+command_list += unknown_args
+
+sp = subprocess.Popen(command_list)
+
+while sp.returncode is None:
+    try:
+        sp.wait()
+    except KeyboardInterrupt:
+        pass
+
+sys.exit(sp.returncode)
diff --git a/scripts/LlamaConfig.cmake.in b/scripts/LlamaConfig.cmake.in
new file mode 100644
index 000000000..6a6d8e39e
--- /dev/null
+++ b/scripts/LlamaConfig.cmake.in
@@ -0,0 +1,71 @@
+set(LLAMA_VERSION @LLAMA_INSTALL_VERSION@)
+set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
+set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
+set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
+set(LLAMA_BLAS @LLAMA_BLAS@)
+set(LLAMA_CUBLAS @LLAMA_CUBLAS@)
+set(LLAMA_METAL @LLAMA_METAL@)
+set(LLAMA_MPI @LLAMA_MPI@)
+set(LLAMA_CLBLAST @LLAMA_CLBLAST@)
+set(LLAMA_HIPBLAS @LLAMA_HIPBLAS@)
+set(LLAMA_ACCELERATE @LLAMA_ACCELERATE@)
+
+@PACKAGE_INIT@
+
+set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
+set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+if (APPLE AND LLAMA_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (LLAMA_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (LLAMA_CUBLAS)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (LLAMA_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (LLAMA_MPI)
+    find_package(MPI REQUIRED)
+endif()
+
+if (LLAMA_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
+if (LLAMA_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+find_library(llama_LIBRARY llama
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
+set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
+add_library(llama UNKNOWN IMPORTED)
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(Llama)
diff --git a/scripts/build-info.cmake b/scripts/build-info.cmake
index 5023b77ab..c86ab4379 100644
--- a/scripts/build-info.cmake
+++ b/scripts/build-info.cmake
@@ -2,20 +2,18 @@ set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in")
 set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
 set(BUILD_NUMBER 0)
 set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")
 
 # Look for git
 find_package(Git)
 if(NOT Git_FOUND)
-    execute_process(
-        COMMAND which git
-        OUTPUT_VARIABLE GIT_EXECUTABLE
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    if(NOT GIT_EXECUTABLE STREQUAL "")
+    find_program(GIT_EXECUTABLE NAMES git git.exe)
+    if(GIT_EXECUTABLE)
         set(Git_FOUND TRUE)
-        message(STATUS "Found Git using 'which': ${GIT_EXECUTABLE}")
+        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
     else()
-        message(WARNING "Git not found using 'find_package' or 'which'. Build info will not be accurate. Consider installing Git or ensuring it is in the PATH.")
+        message(WARNING "Git not found. Build info will not be accurate.")
     endif()
 endif()
 
@@ -26,26 +24,49 @@ if(Git_FOUND)
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         OUTPUT_VARIABLE HEAD
         OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE GIT_HEAD_RESULT
     )
     execute_process(
         COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
         OUTPUT_VARIABLE COUNT
         OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE GIT_COUNT_RESULT
     )
-    if(GIT_HEAD_RESULT EQUAL 0 AND GIT_COUNT_RESULT EQUAL 0)
-        set(BUILD_COMMIT ${HEAD})
-        set(BUILD_NUMBER ${COUNT})
-    endif()
+    set(BUILD_COMMIT ${HEAD})
+    set(BUILD_NUMBER ${COUNT})
+endif()
+
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    execute_process(
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_COMPILER ${OUT})
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
 endif()
 
 # Only write the header if it's changed to prevent unnecessary recompilation
 if(EXISTS ${HEADER_FILE})
-    file(STRINGS ${HEADER_FILE} CONTENTS REGEX "BUILD_COMMIT \"([^\"]*)\"")
-    list(GET CONTENTS 0 EXISTING)
-    if(NOT EXISTING STREQUAL "#define BUILD_COMMIT \"${BUILD_COMMIT}\"")
+    file(READ ${HEADER_FILE} CONTENTS)
+    string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
         configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
     endif()
 else()
diff --git a/scripts/build-info.h.in b/scripts/build-info.h.in
index 75d1e16fd..e996faef0 100644
--- a/scripts/build-info.h.in
+++ b/scripts/build-info.h.in
@@ -3,5 +3,7 @@
 
 #define BUILD_NUMBER @BUILD_NUMBER@
 #define BUILD_COMMIT "@BUILD_COMMIT@"
+#define BUILD_COMPILER "@BUILD_COMPILER@"
+#define BUILD_TARGET "@BUILD_TARGET@"
 
 #endif // BUILD_INFO_H
diff --git a/scripts/build-info.sh b/scripts/build-info.sh
index ed0d6c56a..3c8b1fb85 100755
--- a/scripts/build-info.sh
+++ b/scripts/build-info.sh
@@ -1,23 +1,35 @@
 #!/bin/sh
 
-BUILD_NUMBER="0"
-BUILD_COMMIT="unknown"
+CC=$1
 
-REV_LIST=$(git rev-list --count HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_NUMBER=$REV_LIST
+build_number="0"
+build_commit="unknown"
+build_compiler="unknown"
+build_target="unknown"
+
+if out=$(git rev-list --count HEAD); then
+  # git is broken on WSL so we need to strip extra newlines
+  build_number=$(printf '%s' "$out" | tr -d '\n')
 fi
 
-REV_PARSE=$(git rev-parse --short HEAD)
-if [ $? -eq 0 ]; then
-  BUILD_COMMIT=$REV_PARSE
+if out=$(git rev-parse --short HEAD); then
+  build_commit=$(printf '%s' "$out" | tr -d '\n')
+fi
+
+if out=$($CC --version | head -1); then
+  build_compiler=$out
+fi
+
+if out=$($CC -dumpmachine); then
+  build_target=$out
 fi
 
 echo "#ifndef BUILD_INFO_H"
 echo "#define BUILD_INFO_H"
-echo ""
-echo "#define BUILD_NUMBER $BUILD_NUMBER" | tr -d '\n'
-echo ""
-echo "#define BUILD_COMMIT \"$BUILD_COMMIT\"" | tr -d '\n'
-echo ""
+echo
+echo "#define BUILD_NUMBER $build_number"
+echo "#define BUILD_COMMIT \"$build_commit\""
+echo "#define BUILD_COMPILER \"$build_compiler\""
+echo "#define BUILD_TARGET \"$build_target\""
+echo
 echo "#endif // BUILD_INFO_H"
diff --git a/scripts/convert-gg.sh b/scripts/convert-gg.sh
new file mode 100755
index 000000000..01fda16fd
--- /dev/null
+++ b/scripts/convert-gg.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+set -e
+
+# LLaMA v1
+python3 convert.py ../llama1/7B  --outfile models/llama-7b/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16
+
+# LLaMA v2
+python3 convert.py ../llama2/llama-2-7b  --outfile models/llama-7b-v2/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16
+
+# Code Llama
+python3 convert.py ../codellama/CodeLlama-7b/  --outfile models/codellama-7b/ggml-model-f16.gguf  --outtype f16
+python3 convert.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16
+python3 convert.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16
+
+# Falcon
+python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-7b  1
+mv -v ../falcon/falcon-7b/ggml-model-f16.gguf models/falcon-7b/ggml-model-f16.gguf
+
+python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-40b 1
+mv -v ../falcon/falcon-40b/ggml-model-f16.gguf models/falcon-40b/ggml-model-f16.gguf
diff --git a/scripts/perf-run-all.sh b/scripts/perf-run-all.sh
deleted file mode 100755
index 7dbfc7c20..000000000
--- a/scripts/perf-run-all.sh
+++ /dev/null
@@ -1,93 +0,0 @@
-#!/bin/bash
-#
-# Measure the performance (time per token) of the various quantization techniques
-#
-
-QUANTIZE=0
-if [ "$1" != "" ]; then
-    echo "Quantizing"
-    QUANTIZE=1
-fi
-
-if [ "$QUANTIZE" != "0" ]; then
-    #
-    # quantize
-    #
-
-    # 7B
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
-    time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
-
-    # 13B
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
-    time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
-fi
-
-#
-# perf
-# run each command twice
-#
-
-set -x
-
-# 7B - 4 threads
-     ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings
-
-# 7B - 8 threads
-     ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings
-
-# 13B - 4 threads
-     ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings
-
-# 13B - 8 threads
-     ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-f16.bin  -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-f16.txt  | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings
-     ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe"
-time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings
diff --git a/scripts/ppl-run-all.sh b/scripts/ppl-run-all.sh
deleted file mode 100755
index c59e3075d..000000000
--- a/scripts/ppl-run-all.sh
+++ /dev/null
@@ -1,39 +0,0 @@
-#!/bin/bash
-
-#
-# quantize
-#
-
-# 7B
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt
-time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt
-
-# 13B
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt
-time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt
-
-#
-# perplexity
-#
-
-# 7B
-time ./bin/perplexity -m ../models/7B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-f16.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_0.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_1.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_0.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_1.txt
-time ./bin/perplexity -m ../models/7B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q8_0.txt
-
-# 13B
-time ./bin/perplexity -m ../models/13B/ggml-model-f16.bin  -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-f16.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_0.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_1.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_0.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_1.txt
-time ./bin/perplexity -m ../models/13B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q8_0.txt
diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh
new file mode 100755
index 000000000..b4c2a159e
--- /dev/null
+++ b/scripts/qnt-all.sh
@@ -0,0 +1,30 @@
+#!/bin/bash
+
+qnt=(q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args=""
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+model="$1"
+out="../tmp/results-${model}"
+
+set -o pipefail
+set -e
+
+mkdir -p ${out}
+
+for q in ${qnt[@]}; do
+    time ./bin/quantize ../models/${model}/ggml-model-f16.gguf ../models/${model}/ggml-model-${q}.gguf ${q} 2>&1 ${args} | tee ${out}/qnt-${q}.txt
+done
diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh
new file mode 100755
index 000000000..6384e364d
--- /dev/null
+++ b/scripts/run-all-perf.sh
@@ -0,0 +1,34 @@
+#!/bin/bash
+
+qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args="-ngl 999 -n 64 -p 512"
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+model="$1"
+out="../tmp/results-${model}"
+
+set -o pipefail
+set -e
+
+mkdir -p ${out}
+
+mstr=""
+
+for q in ${qnt[@]}; do
+    mstr="${mstr} -m ../models/${model}/ggml-model-${q}.gguf"
+done
+
+./bin/llama-bench ${mstr} ${args} 2> /dev/null
diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh
new file mode 100755
index 000000000..e04d61d7f
--- /dev/null
+++ b/scripts/run-all-ppl.sh
@@ -0,0 +1,30 @@
+#!/bin/bash
+
+qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
+args="-ngl 999 -t 8"
+
+if [ -z "$1" ]; then
+    echo "usage: $0 <model> [qnt] [args]"
+    echo "default: $0 <model> \"${qnt[@]}\" \"${args}\""
+    exit 1
+fi
+
+if [ ! -z "$2" ]; then
+    qnt=($2)
+fi
+
+if [ ! -z "$3" ]; then
+    args="$3"
+fi
+
+set -o pipefail
+set -e
+
+model="$1"
+out="../tmp/results-${model}"
+
+mkdir -p ${out}
+
+for q in ${qnt[@]}; do
+    time ./bin/perplexity -m ../models/${model}/ggml-model-f16.gguf -f ./wiki.test.raw ${args} 2>&1 | tee ${out}/ppl-${q}.txt
+done
diff --git a/scripts/verify-checksum-models.py b/scripts/verify-checksum-models.py
index 307b7c08d..dff4b4734 100755
--- a/scripts/verify-checksum-models.py
+++ b/scripts/verify-checksum-models.py
@@ -1,4 +1,4 @@
-#!/bin/env python3
+#!/usr/bin/env python3
 
 import os
 import hashlib
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 4ccefe932..61407e573 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -7,9 +7,6 @@ endfunction()
 
 function(llama_test_executable name source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
-    # add_executable(${TEST_TARGET} ${source})
-    # install(TARGETS ${TEST_TARGET} RUNTIME)
-    # target_link_libraries(${TEST_TARGET} PRIVATE llama)
     add_test(NAME ${name} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()
 
@@ -25,12 +22,23 @@ endfunction()
 llama_build_and_test_executable(test-quantize-fns.cpp)
 llama_build_and_test_executable(test-quantize-perf.cpp)
 llama_build_and_test_executable(test-sampling.cpp)
-llama_build_executable(test-tokenizer-0.cpp)
-llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
-llama_build_executable(test-tokenizer-1.cpp)
-llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
-#llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+llama_build_executable(test-tokenizer-0-llama.cpp)
+llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_build_executable(test-tokenizer-0-falcon.cpp)
+llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_build_executable(test-tokenizer-1-llama.cpp)
+llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_build_executable(test-tokenizer-1-bpe.cpp)
+llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 llama_build_and_test_executable(test-grammar-parser.cpp)
 llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp) # SLOW
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
+
+llama_build_and_test_executable(test-rope.cpp)
+
+# dummy executable - not installed
+get_filename_component(TEST_TARGET test-c.c NAME_WE)
+add_executable(${TEST_TARGET} test-c.c)
+target_link_libraries(${TEST_TARGET} PRIVATE llama)
diff --git a/tests/test-c.c b/tests/test-c.c
new file mode 100644
index 000000000..a05071080
--- /dev/null
+++ b/tests/test-c.c
@@ -0,0 +1,3 @@
+#include "llama.h"
+
+int main(void) {}
diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp
index 75a698d73..0a559b27a 100644
--- a/tests/test-grad0.cpp
+++ b/tests/test-grad0.cpp
@@ -107,7 +107,7 @@ static struct ggml_tensor * get_random_tensor_f32(
             break;
         default:
             assert(false);
-    };
+    }
 
     return result;
 }
@@ -155,7 +155,7 @@ static struct ggml_tensor * get_random_tensor_f16(
             break;
         default:
             assert(false);
-    };
+    }
 
     return result;
 }
@@ -203,31 +203,11 @@ static struct ggml_tensor * get_random_tensor_i32(
             break;
         default:
             assert(false);
-    };
+    }
 
     return result;
 }
 
-static void print_elements(const char* label, const struct ggml_tensor * t) {
-    if (!t) {
-        printf("%s: %s = null\n", __func__, label);
-        return;
-    }
-    const int nelements = ggml_nelements(t);
-    printf("%s: %s = [", __func__, label);
-    for (int k = 0; k < nelements; ++k) {
-        if (k > 0) { printf(", "); }
-        printf("%.5f", ggml_get_f32_1d(t, k));
-    }
-    printf("] shape: [");
-    for (int k = 0; k < t->n_dims; ++k) {
-        if (k > 0) { printf(", "); }
-        printf("%d", (int)t->ne[k]);
-    }
-    printf("]\n");
-
-}
-
 static bool check_gradient(
         const char * op_name,
         struct ggml_context * ctx0,
@@ -251,18 +231,20 @@ static bool check_gradient(
         printf("GGML_N_THREADS = %d\n", n_threads);
     }
 
-    struct ggml_cgraph gf = ggml_build_forward (f);
-    struct ggml_cgraph gb = ggml_build_backward(ctx0, &gf, false);
+    struct ggml_cgraph * gf = ggml_build_forward_ctx(ctx0, f);
+    struct ggml_cgraph * gb = ggml_new_graph(ctx0);
+    *gb = *gf;
+    ggml_build_backward_expand(ctx0, gf, gb, false);
 
-    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+    ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
 
-    ggml_graph_reset  (&gf);
+    ggml_graph_reset  (gf);
     ggml_set_f32      (f->grad, 1.0f);
 
-    ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
+    ggml_graph_compute_with_ctx(ctx0, gb, n_threads);
 
-    // ggml_graph_dump_dot(&gf, NULL, "test-grad0-forward.dot");
-    // ggml_graph_dump_dot(&gb, &gf,  "test-grad0-backward.dot");
+    // ggml_graph_dump_dot(gf, NULL, "test-grad0-forward.dot");
+    // ggml_graph_dump_dot(gb, gf,  "test-grad0-backward.dot");
 
     for (int i = 0; i < nargs; ++i) {
         const int nelements = ggml_nelements(x[i]);
@@ -273,29 +255,29 @@ static bool check_gradient(
             const float xp = x0 + eps;
             ggml_set_f32_1d(x[i], k, xp);
 
-            ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
 
-            const float f0 = ggml_get_f32_1d(f, 0);
+            const double f0 = ggml_get_f32_1d(f, 0);
 
             ggml_set_f32_1d(x[i], k, xm);
 
-            ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
 
-            const float f1 = ggml_get_f32_1d(f, 0);
-            const float g0 = (f0 - f1)/(2.0f*eps);
+            const double f1 = ggml_get_f32_1d(f, 0);
+            const double g0 = (f0 - f1)/(2.0*(double) eps);
 
             ggml_set_f32_1d(x[i], k, x0);
 
             // compute gradient using backward graph
-            ggml_graph_reset  (&gf);
+            ggml_graph_reset  (gf);
             ggml_set_f32      (f->grad, 1.0f);
 
-            ggml_graph_compute_with_ctx(ctx0, &gb, n_threads);
+            ggml_graph_compute_with_ctx(ctx0, gb, n_threads);
 
-            const float g1 = ggml_get_f32_1d(x[i]->grad, k);
+            const double g1 = ggml_get_f32_1d(x[i]->grad, k);
 
-            const float error_abs = fabsf(g0 - g1);
-            const float error_rel = g0 != 0 ? fabsf(g0 - g1)/fabsf(g0) : 0;
+            const double error_abs = fabs(g0 - g1);
+            const double error_rel = g0 != 0 ? fabs(g0 - g1)/fabs(g0) : 0;
 
             if (error_abs > max_error_abs || error_rel > max_error_rel) {
                 printf("%s: ndims=%d, i=%d, k=%d, x0=%f, xm=%f, xp=%f, f0=%f, f1=%f, g0=%f, g1=%f, eps=%f, error_abs=%f, error_rel=%f\n",
@@ -373,7 +355,7 @@ static bool check_mat_mul(
 
 int main(int argc, const char ** argv) {
     struct ggml_init_params params = {
-        /* .mem_size   = */ 128*1024*1024,
+        /* .mem_size   = */ 256*1024*1024,
         /* .mem_buffer = */ NULL,
         /* .no_alloc   = */ false,
     };
@@ -405,6 +387,7 @@ int main(int argc, const char ** argv) {
         }
     }
 
+    unsigned seed_iter = 1;
 
     // original loop: 1000
     int niter = 4;
@@ -416,6 +399,10 @@ int main(int argc, const char ** argv) {
         niter = atoi(argv[1]);
     }
     for (int iter = 0; iter < niter; ++iter) {
+        srand(seed_iter);
+        seed_iter = rand();
+        unsigned seed = rand();
+
         printf("test-grad0: iter:%d/%d\n", iter, niter);
         struct ggml_context * ctx0 = ggml_init(params);
 
@@ -425,6 +412,7 @@ int main(int argc, const char ** argv) {
 
         // add f32
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -441,6 +429,7 @@ int main(int argc, const char ** argv) {
 
         // add f16
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -457,6 +446,7 @@ int main(int argc, const char ** argv) {
 
         // sub
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -473,6 +463,7 @@ int main(int argc, const char ** argv) {
 
         // mul
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -489,6 +480,7 @@ int main(int argc, const char ** argv) {
 
         // div
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -505,6 +497,7 @@ int main(int argc, const char ** argv) {
 
         // sqr
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -521,6 +514,7 @@ int main(int argc, const char ** argv) {
 
         // sqrt
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -531,12 +525,13 @@ int main(int argc, const char ** argv) {
 
                 struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqrt(ctx0, x[0]));
 
-                check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f);
+                check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f);
             }
         }
 
         // log
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -553,6 +548,7 @@ int main(int argc, const char ** argv) {
 
         // sum
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -570,6 +566,7 @@ int main(int argc, const char ** argv) {
 
         // sum_rows
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -587,6 +584,7 @@ int main(int argc, const char ** argv) {
         // mean, not yet fully implemented
         if(0)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -604,6 +602,7 @@ int main(int argc, const char ** argv) {
         // argmax
         if (0)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -620,6 +619,7 @@ int main(int argc, const char ** argv) {
 
         // repeat
         {
+            srand(seed);
             int64_t ne2[4];
             get_random_dims(ne2, 4);
 
@@ -642,6 +642,7 @@ int main(int argc, const char ** argv) {
 
         // repeat back
         {
+            srand(seed);
             int64_t ne2[4];
             get_random_dims(ne2, 4);
 
@@ -680,6 +681,7 @@ int main(int argc, const char ** argv) {
 
         // sgn
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -696,6 +698,7 @@ int main(int argc, const char ** argv) {
 
         // neg
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -712,6 +715,7 @@ int main(int argc, const char ** argv) {
 
         // step
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -729,6 +733,7 @@ int main(int argc, const char ** argv) {
         // tanh, not yet fully implemented
         if(0)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -745,33 +750,45 @@ int main(int argc, const char ** argv) {
 
         // mul_mat
         {
+            srand(seed);
             const int nargs = 2;
 
-            for (int ndims = 2; ndims <= 2; ++ndims) {
+            for (int ndims = 2; ndims <= 4; ++ndims) {
+                int max_nrep = (ndims >= 3) ? 2 : 1;
                 x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                {
-                    int64_t ne2[4];
-                    get_random_dims(ne2, 4);
-                    ne2[0] = ne[0];
-                    x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                for (int nrep2 = 1; nrep2 < max_nrep; ++nrep2) {
+                    for (int nrep3 = 1; nrep3 < max_nrep; ++nrep3) {
+                        {
+                            int64_t ne2[4];
+                            get_random_dims(ne2, 4);
+                            ne2[0] = ne[0];
+                            ne2[2] = nrep2 * ne[2];
+                            ne2[3] = nrep3 * ne[3];
+                            x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+                        }
+
+                        ggml_set_param(ctx0, x[0]);
+                        ggml_set_param(ctx0, x[1]);
+
+                        struct ggml_tensor * m = ggml_mul_mat(ctx0, x[1], x[0]);
+                        struct ggml_tensor * f = ggml_sum(ctx0, m);
+
+                        GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
+
+                        check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                        if (ndims == 2) {
+                            // check_mat_mul does not support ndims > 2
+                            check_mat_mul(m, x[1], x[0]);
+                        }
+                    }
                 }
-
-                ggml_set_param(ctx0, x[0]);
-                ggml_set_param(ctx0, x[1]);
-
-                struct ggml_tensor * m = ggml_mul_mat(ctx0, x[1], x[0]);
-                struct ggml_tensor * f = ggml_sum(ctx0, m);
-
-                GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
-
-                check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
-                check_mat_mul(m, x[1], x[0]);
             }
         }
 
         // elu, not yet fully implemented
         if(0)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -788,6 +805,7 @@ int main(int argc, const char ** argv) {
 
         // relu
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -805,6 +823,7 @@ int main(int argc, const char ** argv) {
         // gelu, not yet fully implemented
         if(0)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 4; ++ndims) {
@@ -821,6 +840,7 @@ int main(int argc, const char ** argv) {
 
         // silu
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -842,6 +862,7 @@ int main(int argc, const char ** argv) {
 
         // rms_norm
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -858,6 +879,7 @@ int main(int argc, const char ** argv) {
 
         // scale
         {
+            srand(seed);
             const int nargs = 2;
 
             int64_t ne2[4];
@@ -878,6 +900,7 @@ int main(int argc, const char ** argv) {
 
         // cpy f32
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -895,6 +918,7 @@ int main(int argc, const char ** argv) {
 
         // cpy f16
         {
+            srand(seed);
             const int nargs = 2;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -912,6 +936,7 @@ int main(int argc, const char ** argv) {
 
         // reshape (1d->nd)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -935,6 +960,7 @@ int main(int argc, const char ** argv) {
 
         // reshape (nd->1d)
         {
+            srand(seed);
             const int nargs = 1;
 
             for (int ndims = 1; ndims <= 2; ++ndims) {
@@ -958,6 +984,7 @@ int main(int argc, const char ** argv) {
 
         // acc 1d
         {
+            srand(seed);
             int64_t ne2[4] = { 1, 1, 1, 1 };
 
             const int nargs = 2;
@@ -985,6 +1012,7 @@ int main(int argc, const char ** argv) {
 
         // acc 2d
         {
+            srand(seed);
             int64_t ne2[4]         = { 1, 1, 1, 1 };
             int64_t max_offsets[4] = { 0, 0, 0, 0 };
             int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1017,6 +1045,7 @@ int main(int argc, const char ** argv) {
 
         // acc 3d
         {
+            srand(seed);
             int64_t ne2[4]         = { 1, 1, 1, 1 };
             int64_t max_offsets[4] = { 0, 0, 0, 0 };
             int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1051,6 +1080,7 @@ int main(int argc, const char ** argv) {
 
         // acc 4d
         {
+            srand(seed);
             int64_t ne2[4]         = { 1, 1, 1, 1 };
             int64_t max_offsets[4] = { 0, 0, 0, 0 };
             int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1087,6 +1117,7 @@ int main(int argc, const char ** argv) {
 
         // set_1d
         {
+            srand(seed);
             int64_t ne2[4];
 
             const int nargs = 2;
@@ -1114,6 +1145,7 @@ int main(int argc, const char ** argv) {
 
         // set_2d
         {
+            srand(seed);
             int64_t ne2[4];
             int64_t max_offsets[4] = { 0, 0, 0, 0 };
             int64_t offsets[4]     = { 0, 0, 0, 0 };
@@ -1146,6 +1178,7 @@ int main(int argc, const char ** argv) {
 
         // view_1d
         {
+            srand(seed);
             const int nargs = 1;
             for (int ndims = 1; ndims <= 4; ++ndims) {
 
@@ -1169,6 +1202,7 @@ int main(int argc, const char ** argv) {
 
         // view_2d
         {
+            srand(seed);
             int64_t ne2[4];
             int64_t nb2[4];
 
@@ -1199,6 +1233,7 @@ int main(int argc, const char ** argv) {
 
         // view_3d
         {
+            srand(seed);
             int64_t ne2[4] = {1,1,1,1};
             int64_t nb2[4] = {0,0,0,0};
 
@@ -1230,6 +1265,7 @@ int main(int argc, const char ** argv) {
 
         // permute
         {
+            srand(seed);
             int64_t ne2[4];
 
             const int nargs = 1;
@@ -1263,6 +1299,7 @@ int main(int argc, const char ** argv) {
 
         // transpose
         {
+            srand(seed);
             int64_t ne2[4];
 
             const int nargs = 1;
@@ -1290,6 +1327,7 @@ int main(int argc, const char ** argv) {
 
         // get_rows
         {
+            srand(seed);
             int64_t ne2[4] = {ne[0], ne[1], 1, 1};
             int64_t ne3[4] = {1+irand(ne[1]), 1, 1, 1};
             const int nargs = 1;
@@ -1306,6 +1344,7 @@ int main(int argc, const char ** argv) {
 
         // diag_mask_inf
         {
+            srand(seed);
             const int nargs = 1;
             const int ndims = 2;
 
@@ -1321,6 +1360,7 @@ int main(int argc, const char ** argv) {
 
         // diag_mask_zero
         {
+            srand(seed);
             const int nargs = 1;
             const int ndims = 2;
 
@@ -1336,6 +1376,7 @@ int main(int argc, const char ** argv) {
 
         // softmax
         {
+            srand(seed);
             const int nargs = 1;
 
             int64_t ne2[4];
@@ -1345,33 +1386,59 @@ int main(int argc, const char ** argv) {
                 x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
                 ggml_set_param(ctx0, x[0]);
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_soft_max(ctx0, x[0]));
+                float eps = 1e-6f;
+                // dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
+                // instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
+                struct ggml_tensor * f = ggml_sum(ctx0,
+                                            ggml_log(ctx0,
+                                                ggml_add1(ctx0,
+                                                    ggml_scale(ctx0,
+                                                        ggml_soft_max(ctx0, x[0]),
+                                                        ggml_new_f32(ctx0, 1.0f - eps)),
+                                                    ggml_new_f32(ctx0, eps))));
 
-                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY);
+                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY);
+                // NOTE: softmax forward is computed using f16 table lookup instead of using actual expf, but backward assumes actual expf.
+                // this may result in different gradients too finite differences.
+                // when this test reports errors, first try to replace the table lookup with actual expf and test again to see if just that was the cause.
+                // if only the table lookup causes gradients to differ this is acceptable.
             }
         }
 
         // cross_entropy_loss
         {
+            srand(seed);
             const int nargs = 1;
 
             int64_t ne2[4];
             get_random_dims(ne2, 4);
 
-            for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
+            for (int ndims = 1; ndims <= 4; ++ndims) {
+                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -0.1f, 0.1f);
                 x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
+                // the second argument to cross_entropy_loss must sum up to 1 for each row
+                int nr = ggml_nrows(x[1]);
+                int nc = ggml_nelements(x[1]) / nr;
+                for (int ir = 0; ir < nr; ++ir) {
+                    float sum = 0;
+                    for (int ic = 0; ic < nc; ++ic) {
+                        sum += ((float *) x[1]->data)[ic + ir*nc];
+                    }
+                    for (int ic = 0; ic < nc; ++ic) {
+                        ((float *) x[1]->data)[ic + ir*nc] /= sum;
+                    }
+                }
                 ggml_set_param(ctx0, x[0]);
 
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+                struct ggml_tensor * f = ggml_cross_entropy_loss(ctx0, x[0], x[1]);
 
-                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
-                // finite differences regularly fails!
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-4f, 1e-3f, INFINITY);
             }
         }
 
         // rope f32
         {
+            srand(seed);
             const int nargs = 1;
 
             int64_t ne2[4];
@@ -1384,6 +1451,11 @@ int main(int argc, const char ** argv) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                         x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
 
+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                         ggml_set_param(ctx0, x[0]);
 
                         const bool skip_past = (mode & 1);
@@ -1395,7 +1467,7 @@ int main(int argc, const char ** argv) {
                             continue;
                         }
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));
 
                         GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                         check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY);
@@ -1406,6 +1478,7 @@ int main(int argc, const char ** argv) {
 
         // rope f16
         {
+            srand(seed);
             const int nargs = 1;
 
             int64_t ne2[4];
@@ -1418,6 +1491,11 @@ int main(int argc, const char ** argv) {
                     for (int n_past = 1; n_past < ne2[2]; ++n_past) {
                         x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);
 
+                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
+                        for (int i = 0; i < ne2[2]; ++i) {
+                            ((int32_t *) p->data)[i] = n_past + i;
+                        }
+
                         ggml_set_param(ctx0, x[0]);
 
                         const bool skip_past = (mode & 1);
@@ -1429,7 +1507,7 @@ int main(int argc, const char ** argv) {
                             continue;
                         }
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], n_past, n_rot, mode, 0));
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode, 0));
 
                         GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
                         check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY);
@@ -1440,6 +1518,7 @@ int main(int argc, const char ** argv) {
 
         // flash_attn f32
         {
+            srand(seed);
             const int nargs = 3;
 
             int64_t ne2[4];
@@ -1452,28 +1531,31 @@ int main(int argc, const char ** argv) {
 
             for (int masked = 0; masked <= 1; ++masked) {
                 for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int64_t neq[4] = { D, N, B, ne[3] };
-                    int64_t nek[4] = { D, M, B, ne[3] };
-                    int64_t nev[4] = { M, D, B, ne[3] };
-                    if (ndims == 2) {
-                        neq[2] = 1; neq[3] = 1;
-                        nek[2] = 1; nek[3] = 1;
-                        nev[2] = 1; nev[3] = 1;
-                    } else if (ndims == 3) {
-                        neq[3] = 1;
-                        nek[3] = 1;
-                        nev[3] = 1;
+                    int max_nrep = (ndims >= 3) ? 2 : 1;
+                    for (int nrep = 1; nrep < max_nrep; ++nrep) {
+                        int64_t neq[4] = { D, N, B*nrep, ne[3] };
+                        int64_t nek[4] = { D, M, B, ne[3] };
+                        int64_t nev[4] = { M, D, B, ne[3] };
+                        if (ndims == 2) {
+                            neq[2] = 1; neq[3] = 1;
+                            nek[2] = 1; nek[3] = 1;
+                            nev[2] = 1; nev[3] = 1;
+                        } else if (ndims == 3) {
+                            neq[3] = 1;
+                            nek[3] = 1;
+                            nev[3] = 1;
+                        }
+                        x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                        x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                        x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                        ggml_set_param(ctx0, x[0]);
+                        ggml_set_param(ctx0, x[1]);
+                        ggml_set_param(ctx0, x[2]);
+
+                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                        check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
                     }
-                    x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                    ggml_set_param(ctx0, x[0]);
-                    ggml_set_param(ctx0, x[1]);
-                    ggml_set_param(ctx0, x[2]);
-
-                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
-
-                    check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
                 }
             }
         }
@@ -1481,6 +1563,7 @@ int main(int argc, const char ** argv) {
         // flash_attn f16, not yet fully implemented
         if(0)
         {
+            srand(seed);
             const int nargs = 3;
 
             int64_t ne2[4];
@@ -1514,7 +1597,7 @@ int main(int argc, const char ** argv) {
 
                     struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
                 }
             }
         }
diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp
index 8ab240202..bb8af5962 100644
--- a/tests/test-opt.cpp
+++ b/tests/test-opt.cpp
@@ -36,37 +36,13 @@
 #define GGML_PRINT(...) printf(__VA_ARGS__)
 
 
-float frand(void) {
+static float frand(void) {
     return (float)rand()/(float)RAND_MAX;
 }
 
-int irand(int n) {
-    return rand()%n;
-}
-
-void get_random_dims(int64_t * dims, int ndims) {
-    dims[0] = dims[1] = dims[2] = dims[3] = 1;
-
-    for (int i = 0; i < ndims; i++) {
-        dims[i] = 1 + irand(4);
-    }
-}
-
-void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
-    dims[0] = dims[1] = dims[2] = dims[3] = 1;
-
-    for (int i = 0; i < ndims; i++) {
-        dims[i] = min + irand(max-min);
-    }
-}
-
-
-struct ggml_tensor * get_random_tensor(
-        struct ggml_context * ctx0,
-        int ndims,
-        int64_t ne[],
-        float fmin,
-        float fmax) {
+static struct ggml_tensor * get_random_tensor(
+    struct ggml_context * ctx0, int ndims, int64_t ne[], float fmin, float fmax
+) {
     struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
 
     switch (ndims) {
@@ -104,19 +80,11 @@ struct ggml_tensor * get_random_tensor(
             break;
         default:
             assert(false);
-    };
+    }
 
     return result;
 }
 
-float get_element(const struct ggml_tensor * t, int idx) {
-    return ((float *)t->data)[idx];
-}
-
-void set_element(struct ggml_tensor * t, int idx, float value) {
-    ((float *)t->data)[idx] = value;
-}
-
 int main(void) {
     struct ggml_init_params params = {
         /* .mem_size   = */ 1024*1024*1024,
@@ -127,7 +95,7 @@ int main(void) {
     struct ggml_context * ctx = ggml_init(params);
 
     int64_t ne1[4] = {4, 128, 1, 1};
-    int64_t ne2[4] = {4, 256, 1, 1};;
+    int64_t ne2[4] = {4, 256, 1, 1};
     int64_t ne3[4] = {128, 256, 1, 1};
 
     struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp
index 8d3c162d2..884af4054 100644
--- a/tests/test-quantize-fns.cpp
+++ b/tests/test-quantize-fns.cpp
@@ -13,24 +13,24 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
-const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
-const float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
-const float MAX_DOT_PRODUCT_ERROR = 0.02f;
+constexpr float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR = 0.002f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
+constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
 
-const char* RESULT_STR[] = {"ok", "FAILED"};
+static const char* RESULT_STR[] = {"ok", "FAILED"};
 
 
 // Generate synthetic data
-void generate_data(float offset, size_t n, float * dst) {
+static void generate_data(float offset, size_t n, float * dst) {
     for (size_t i = 0; i < n; i++) {
         dst[i] = 0.1 + 2*cosf(i + offset);
     }
 }
 
 // Calculate RMSE between two float arrays
-float array_rmse(const float * a1, const float * a2, size_t n) {
+static float array_rmse(const float * a1, const float * a2, size_t n) {
     double sum = 0;
     for (size_t i = 0; i < n; i++) {
         double diff = a1[i] - a2[i];
@@ -40,7 +40,7 @@ float array_rmse(const float * a1, const float * a2, size_t n) {
 }
 
 // Total quantization error on test data
-float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
 
@@ -50,7 +50,7 @@ float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, cons
 }
 
 // Total quantization error on test data
-float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
     std::vector<float> tmp_out_ref(test_size);
@@ -64,7 +64,7 @@ float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size,
     return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
 }
 
-float dot_product(const float * a1, const float * a2, size_t test_size) {
+static float dot_product(const float * a1, const float * a2, size_t test_size) {
     double sum = 0;
     for (size_t i = 0; i < test_size; i++) {
         sum += a1[i] * a2[i];
@@ -73,7 +73,9 @@ float dot_product(const float * a1, const float * a2, size_t test_size) {
 }
 
 // Total dot product error
-float dot_product_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
+static float dot_product_error(
+    ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2
+) {
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);
 
diff --git a/tests/test-quantize-perf.cpp b/tests/test-quantize-perf.cpp
index 0bb9537f6..88fac0e23 100644
--- a/tests/test-quantize-perf.cpp
+++ b/tests/test-quantize-perf.cpp
@@ -61,37 +61,36 @@ inline int64_t cpu_cycles() {
 
 
 // Generate synthetic data
-void generate_data(float offset, size_t n, float * dst) {
+static void generate_data(float offset, size_t n, float * dst) {
     for (size_t i = 0; i < n; i++) {
         dst[i] = 0.1 + 2*cosf(i + offset);
     }
 }
 
-float gigabytes_per_second(size_t bytes, int64_t usecs) {
+static float gigabytes_per_second(size_t bytes, int64_t usecs) {
     return bytes / (float) usecs * 1000000 / (1024*1024*1024);
 }
 
-void * align_with_offset(void * ptr, int offset) {
+static void * align_with_offset(void * ptr, int offset) {
     size_t dummy_size = MAX_ALIGNMENT * 4;
     return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset;
 }
 
-void benchmark_function(size_t size, size_t q_size, int64_t iterations, std::function<size_t(void)> function) {
+static void benchmark_function(size_t size, size_t q_size, int64_t iterations, const std::function<float(void)> & func) {
     int64_t min_time_us = INT64_MAX;
     int64_t total_time_us = 0;
     int64_t min_time_cycles = INT64_MAX;
     int64_t total_time_cycles = 0;
 
     for (int i = 0; i < WARMUP; i++) {
-        function();
+        func();
     }
 
-
     for (int i = 0; i < iterations; i++) {
         const int64_t start_time = ggml_time_us();
         const int64_t start_cycles = cpu_cycles();
 
-        function();
+        func();
 
         const int64_t end_cycles = cpu_cycles();
         const int64_t end_time = ggml_time_us();
@@ -108,7 +107,7 @@ void benchmark_function(size_t size, size_t q_size, int64_t iterations, std::fun
     printf("      quantized throughput : %9.2f GB/s\n",  gigabytes_per_second(q_size * iterations, total_time_us));
 }
 
-void usage(char * argv[]) {
+static void usage(char * argv[]) {
     printf("Benchmark quantization specific functions on synthetic data\n");
     printf("\n");
     printf("usage: %s [options]\n", argv[0]);
@@ -245,15 +244,15 @@ int main(int argc, char * argv[]) {
 
     std::vector<uint8_t> test_data1_v(largest*4 + MAX_ALIGNMENT*2);
     std::vector<uint8_t> test_data2_v(largest*4 + MAX_ALIGNMENT*2);
-    std::vector<uint8_t> test_q1_v(largest*4 + MAX_ALIGNMENT*2);
-    std::vector<uint8_t> test_q2_v(largest*4 + MAX_ALIGNMENT*2);
-    std::vector<uint8_t> test_out_v(largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q1_v   (largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_q2_v   (largest*4 + MAX_ALIGNMENT*2);
+    std::vector<uint8_t> test_out_v  (largest*4 + MAX_ALIGNMENT*2);
 
     float * test_data1 = (float *) align_with_offset(test_data1_v.data(), params.alignment_offset);
     float * test_data2 = (float *) align_with_offset(test_data2_v.data(), params.alignment_offset);
-    float * test_q1 = (float *) align_with_offset(test_q1_v.data(), params.alignment_offset);
-    float * test_q2 = (float *) align_with_offset(test_q2_v.data(), params.alignment_offset);
-    float * test_out = (float *) align_with_offset(test_out_v.data(), params.alignment_offset);
+    float * test_q1    = (float *) align_with_offset(test_q1_v.data(),    params.alignment_offset);
+    float * test_q2    = (float *) align_with_offset(test_q2_v.data(),    params.alignment_offset);
+    float * test_out   = (float *) align_with_offset(test_out_v.data(),   params.alignment_offset);
 
     generate_data(0, largest, test_data1);
     generate_data(1, largest, test_data2);
@@ -283,7 +282,7 @@ int main(int argc, char * argv[]) {
                 printf("  quantize_row_q_reference\n");
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
-                    auto quantize_fn = [&](void ) {
+                    auto quantize_fn = [&](void) -> float {
                         qfns.from_float_reference(test_data1, test_q1, size);
                         return test_q1[0];
                     };
@@ -297,7 +296,7 @@ int main(int argc, char * argv[]) {
                 printf("  quantize_row_q\n");
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
-                    auto quantize_fn = [&](void ) {
+                    auto quantize_fn = [&](void) -> float {
                         qfns.from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
@@ -312,7 +311,7 @@ int main(int argc, char * argv[]) {
                 qfns.from_float(test_data1, test_q1, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
-                    auto quantize_fn = [&](void ) {
+                    auto quantize_fn = [&](void) -> float {
                         qfns.to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
@@ -326,7 +325,7 @@ int main(int argc, char * argv[]) {
                 printf("  quantize_row_q_dot\n");
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
-                    auto quantize_fn = [&](void ) {
+                    auto quantize_fn = [&](void) -> float {
                         auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
                         vdot.from_float(test_data1, test_q1, size);
                         return test_q1[0];
@@ -343,7 +342,7 @@ int main(int argc, char * argv[]) {
                 qfns.from_float(test_data2, test_q2, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
-                    auto quantize_fn = [&](void ) {
+                    auto quantize_fn = [&](void) -> float {
                         float result;
                         qfns.vec_dot(size, &result, test_q1, test_q2);
                         return result;
diff --git a/tests/test-rope.cpp b/tests/test-rope.cpp
new file mode 100644
index 000000000..26c1f42dc
--- /dev/null
+++ b/tests/test-rope.cpp
@@ -0,0 +1,221 @@
+#include "ggml.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cassert>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wdouble-promotion"
+#endif
+
+#define MAX_NARGS 3
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+#define GGML_SILU_FP16
+
+//
+// logging
+//
+
+#if (GGML_DEBUG >= 1)
+#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG(...)
+#endif
+
+#if (GGML_DEBUG >= 5)
+#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_5(...)
+#endif
+
+#if (GGML_DEBUG >= 10)
+#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_10(...)
+#endif
+
+#define GGML_PRINT(...) printf(__VA_ARGS__)
+
+static float frand(void) {
+    return (float)rand()/(float)RAND_MAX;
+}
+
+static int irand(int n) {
+    if (n == 0) return 0;
+    return rand()%n;
+}
+
+static void get_random_dims(int64_t * dims, int ndims) {
+    dims[0] = dims[1] = dims[2] = dims[3] = 1;
+
+    for (int i = 0; i < ndims; i++) {
+        dims[i] = 1 + irand(4);
+    }
+}
+
+static struct ggml_tensor * get_random_tensor_f32(
+        struct ggml_context * ctx0,
+        int ndims,
+        const int64_t ne[],
+        float fmin,
+        float fmax) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
+
+    switch (ndims) {
+        case 1:
+            for (int i0 = 0; i0 < ne[0]; i0++) {
+                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < ne[1]; i1++) {
+                for (int i0 = 0; i0 < ne[0]; i0++) {
+                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < ne[2]; i2++) {
+                for (int i1 = 0; i1 < ne[1]; i1++) {
+                    for (int i0 = 0; i0 < ne[0]; i0++) {
+                        ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < ne[3]; i3++) {
+                for (int i2 = 0; i2 < ne[2]; i2++) {
+                    for (int i1 = 0; i1 < ne[1]; i1++) {
+                        for (int i0 = 0; i0 < ne[0]; i0++) {
+                            ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            assert(false);
+    };
+
+    return result;
+}
+
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
+int main(int /*argc*/, const char ** /*argv*/) {
+    struct ggml_init_params params = {
+        /* .mem_size   = */ 128*1024*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ false,
+    };
+
+    std::vector<uint8_t> work_buffer;
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    struct ggml_tensor * x;
+
+    // rope f32
+    for (int m = 0; m < 3; ++m) {
+        const int ndims = 4;
+
+        const int64_t n_rot = 128;
+        const int64_t ne[4] = { 2*n_rot, 32, 73, 1 };
+
+        const int n_past_0 = 100;
+        const int n_past_2 = 33;
+
+        struct ggml_tensor * p0 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+        struct ggml_tensor * p1 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+        struct ggml_tensor * p2 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne[2]);
+
+        for (int i = 0; i < ne[2]; ++i) {
+            ((int32_t *) p0->data)[i] = n_past_0 + i;
+            ((int32_t *) p1->data)[i] = n_past_2 - n_past_0;
+            ((int32_t *) p2->data)[i] = n_past_2 + i;
+        }
+
+        // test mode 0, 2, 4 (standard, GPT-NeoX, GLM)
+        const int mode = m == 0 ? 0 : m == 1 ? 2 : 4;
+
+        x = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
+
+        // 100, 101, 102, ..., 172
+        struct ggml_tensor * r0 = ggml_rope(ctx0, x,  p0, n_rot, mode, 1024);
+        // -67, -67, -67, ..., -67
+        struct ggml_tensor * r1 = ggml_rope(ctx0, r0, p1, n_rot, mode, 1024); // "context swap", i.e. forget n_past_0 - n_past_2 tokens
+
+        //  33,  34,  35, ..., 105
+        struct ggml_tensor * r2 = ggml_rope(ctx0, x,  p2, n_rot, mode, 1024);
+
+        ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+        ggml_build_forward_expand(gf, r0);
+        ggml_build_forward_expand(gf, r1);
+        ggml_build_forward_expand(gf, r2);
+
+        ggml_graph_compute_helper(work_buffer, gf, 4);
+
+        // check that r1 and r2 are the same
+        {
+            double sum0 = 0.0f;
+            double sum1 = 0.0f;
+            double diff = 0.0f;
+
+            const float * r1_data = (float *) r1->data;
+            const float * r2_data = (float *) r2->data;
+
+            const int n_elements = ggml_nelements(r1);
+
+            for (int i = 0; i < n_elements; ++i) {
+                sum0 += fabs(r1_data[i]);
+                sum1 += fabs(r2_data[i]);
+                diff += fabs(r1_data[i] - r2_data[i]);
+                //if (fabs(r1_data[i] - r2_data[i]) > 0.0001f) {
+                //    printf("%d: %f %f\n", i, r1_data[i], r2_data[i]);
+                //    printf("diff: %f\n", fabs(r1_data[i] - r2_data[i]));
+                //}
+            }
+
+            //for (int i = 4096; i < 4096 + 128; ++i) {
+            //    printf("%f %f\n", r1_data[i], r2_data[i]);
+            //}
+
+            printf("mode: %d\n", mode);
+            printf("sum0: %f\n", sum0);
+            printf("sum1: %f\n", sum1);
+            printf("diff: %f\n", diff);
+            printf("rel err: %f\n", diff / sum0);
+            printf("rel err: %f\n", diff / sum1);
+
+            GGML_ASSERT(diff / sum0 < 0.0001f);
+            GGML_ASSERT(diff / sum1 < 0.0001f);
+        }
+    }
+
+    ggml_free(ctx0);
+
+    return 0;
+}
+
diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp
index 4437c3948..019c0d462 100644
--- a/tests/test-sampling.cpp
+++ b/tests/test-sampling.cpp
@@ -12,7 +12,8 @@
 #include <vector>
 #include <algorithm>
 
-void dump(const llama_token_data_array * candidates) {
+
+static void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
         printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);
     }
@@ -21,9 +22,7 @@ void dump(const llama_token_data_array * candidates) {
 #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0)
 
 
-void test_top_k(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                int k) {
+static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -45,10 +44,7 @@ void test_top_k(const std::vector<float> & probs,
 }
 
 
-void test_top_p(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float p) {
-
+static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -70,9 +66,7 @@ void test_top_p(const std::vector<float> & probs,
 }
 
 
-void test_tfs(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float z) {
+static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -93,9 +87,7 @@ void test_tfs(const std::vector<float> & probs,
 }
 
 
-void test_typical(const std::vector<float> & probs,
-                const std::vector<float> & expected_probs,
-                float p) {
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
     size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
@@ -116,11 +108,10 @@ void test_typical(const std::vector<float> & probs,
 }
 
 
-void test_repetition_penalty(
-                const std::vector<float> & probs,
-                const std::vector<llama_token> & last_tokens,
-                const std::vector<float> & expected_probs,
-                float penalty) {
+static void test_repetition_penalty(
+    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
+    const std::vector<float> & expected_probs, float penalty
+) {
     assert(probs.size() == expected_probs.size());
 
     size_t n_vocab = probs.size();
@@ -145,11 +136,10 @@ void test_repetition_penalty(
 }
 
 
-void test_frequency_presence_penalty(
-                const std::vector<float> & probs,
-                const std::vector<llama_token> & last_tokens,
-                const std::vector<float> & expected_probs,
-                float alpha_frequency, float alpha_presence) {
+static void test_frequency_presence_penalty(
+    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
+    const std::vector<float> & expected_probs, float alpha_frequency, float alpha_presence
+) {
     assert(probs.size() == expected_probs.size());
 
     size_t n_vocab = probs.size();
diff --git a/tests/test-tokenizer-0-falcon.cpp b/tests/test-tokenizer-0-falcon.cpp
new file mode 100644
index 000000000..0f3c50bce
--- /dev/null
+++ b/tests/test-tokenizer-0-falcon.cpp
@@ -0,0 +1,187 @@
+#include "llama.h"
+#include "common.h"
+#include "console.h"
+
+#include <cstdio>
+#include <string>
+#include <map>
+#include <vector>
+#include <fstream>
+
+// generate using test-tokenizer-0-falcon.py
+static const std::map<std::string, std::vector<llama_token>> & k_tests() {
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { ""                      , {  }, },
+        { " "                     , {     204, }, },
+        { "  "                    , {     258, }, },
+        { "   "                   , {     466, }, },
+        { "\t"                    , {     192, }, },
+        { "\n"                    , {     193, }, },
+        { "\t\n"                  , {   19125, }, },
+        { "Hello world"           , {    9856,   1079, }, },
+        { " Hello world"          , {   23090,   1079, }, },
+        { "Hello World"           , {    9856,   2889, }, },
+        { " Hello World"          , {   23090,   2889, }, },
+        { " Hello World!"         , {   23090,   2889,     12, }, },
+        { "Hello, world!"         , {    9856,     23,   1079,     12, }, },
+        { " Hello, world!"        , {   23090,     23,   1079,     12, }, },
+        { " this is 🦙.cpp"        , {     414,    304,   3346,    111,    231,     25,  29247, }, },
+        { "w048 7tuijk dsdfhu"    , {      98,  55866,    204,     34,  16682,   7149,  36190,   6869,  11481, }, },
+        { "нещо на Български"     , {     150,    133,   6207,    151,    215,    150,    134,   5052,    133,   6279,   5052,    223,    151,    216,  49679,    123,  53110,  47043,   7795, }, },
+        { "កាន់តែពិសេសអាចខលចេញ"   , {   38154,    206,  38154,    126,  38154,    225,    167,    237,    217,  38154,    221,    167,    237,    208,  38154,    228,  38154,    127,  38154,    237,    167,    237,    207,  38154,    237,  38154,    107,  38154,    126,  38154,    211,  38154,    207,  38154,    233,  38154,    211,    167,    237,    207,  38154,    215, }, },
+        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", {    2571,    232,    206,    204,     19,  11003,     20,   8196,    126,    283,    219,  48778,    116,  13392,    204,     19,  51831,    732,  63209,   1741,   7955,    522,     20,  22438,    211,    204,     19,   7927,  53360,    325,    504,    701,    946,  10930,     20, }, },
+        { "Hello"                 , {    9856, }, },
+        { " Hello"                , {   23090, }, },
+        { "  Hello"               , {     204,  23090, }, },
+        { "   Hello"              , {     258,  23090, }, },
+        { "    Hello"             , {     466,  23090, }, },
+        { "    Hello\n    Hello"  , {     466,  23090,    742,  23090, }, },
+    };
+
+    return _k_tests;
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    std::string fname_text;
+    if (argc > 2) {
+        fname_text = argv[2];
+    }
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto mparams = llama_model_default_params();
+
+        mparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), mparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        auto cparams = llama_context_default_params();
+
+        ctx = llama_new_context_with_model(model, cparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    if (llama_vocab_type(model) != LLAMA_VOCAB_TYPE_BPE) {
+        fprintf(stderr, "%s : error: vocab type is not BPE\n", __func__);
+        llama_free_model(model);
+        llama_free(ctx);
+        return 2;
+    }
+
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
+    bool success = true;
+
+    for (const auto & test_kv : k_tests()) {
+        const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, false);
+
+        printf("\n");
+        printf("src: '%s'\n", test_kv.first.c_str());
+        printf("res: '%s'\n", llama_detokenize_bpe(ctx, res).c_str());
+        printf("tok: ");
+        for (const auto & tok : res) {
+            printf("%d ", tok);
+        }
+        printf("\n");
+
+        bool correct = res.size() == test_kv.second.size();
+
+        for (int i = 0; i < (int) res.size() && correct; ++i) {
+            if (test_kv.second[i] != res[i]) {
+                correct = false;
+            }
+        }
+
+        if (!correct) {
+            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                llama_detokenize_bpe(ctx, res).c_str(),
+                llama_detokenize_bpe(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : expected tokens: ", __func__);
+            for (const auto & t : test_kv.second) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : got tokens:      ", __func__);
+            for (const auto & t : res) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+
+            success = false;
+        }
+    }
+
+    if (!fname_text.empty()) {
+        fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
+
+        std::string text;
+        {
+            std::ifstream ifs(fname_text);
+            if (!ifs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str());
+                return 1;
+            }
+            text = std::string(std::istreambuf_iterator<char>(ifs), std::istreambuf_iterator<char>());
+        }
+
+        fprintf(stderr, "%s : text size: %zu\n", __func__, text.size());
+
+        const std::vector<llama_token> res = llama_tokenize(ctx, text, true);
+
+        fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size());
+
+        {
+            const std::string fname_out = fname_text + ".tokcpp";
+
+            std::ofstream ofs(fname_out);
+            if (!ofs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str());
+                return 1;
+            }
+
+            for (const auto & tok : res) {
+                ofs << tok << " ";
+            }
+
+            ofs << "\n";
+        }
+
+        fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str());
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return success ? 0 : 3;
+}
diff --git a/tests/test-tokenizer-0-falcon.py b/tests/test-tokenizer-0-falcon.py
new file mode 100644
index 000000000..9c8c1c7d1
--- /dev/null
+++ b/tests/test-tokenizer-0-falcon.py
@@ -0,0 +1,83 @@
+# tests with BPE tokenizer
+
+import os
+import sys
+import argparse
+
+from transformers import AutoTokenizer
+
+parser = argparse.ArgumentParser()
+parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
+parser.add_argument("--fname-tok",   help="path to a text file to tokenize")
+args = parser.parse_args()
+
+dir_tokenizer = args.dir_tokenizer
+
+tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer)
+
+tests = [
+        "",
+        " ",
+        "  ",
+        "   ",
+        "\t",
+        "\n",
+        "\t\n",
+        "Hello world",
+        " Hello world",
+        "Hello World",
+        " Hello World",
+        " Hello World!",
+        "Hello, world!",
+        " Hello, world!",
+        " this is 🦙.cpp",
+        "w048 7tuijk dsdfhu",
+        "нещо на Български",
+        "កាន់តែពិសេសអាចខលចេញ",
+        "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+        "Hello",
+        " Hello",
+        "  Hello",
+        "   Hello",
+        "    Hello",
+        "    Hello\n    Hello",
+    ]
+
+for text in tests:
+    print('text: ', text)
+    print(tokenizer.encode(text))
+    print(tokenizer.decode(tokenizer.encode(text)))
+
+print("\n\ntests for C++:\n")
+for text in tests:
+    res = tokenizer.encode(text)
+
+    k = text.replace('\n', '\\n')
+    k = k.replace('\t', '\\t')
+    k = '"' + k + '"'
+    print("{ %-24s, { " % k, end='')
+    for x in res:
+        print("%7d," % x, end='')
+    print(" }, },")
+
+print(tokenizer.encode('hello'))
+print(tokenizer.encode('world'))
+print(tokenizer.encode(' world'))
+print(tokenizer.encode('hello world'))
+
+fname_tok = args.fname_tok
+if fname_tok:
+    print('tokenizing file: ', fname_tok)
+    fname_out = fname_tok + '.tok'
+    with open(fname_tok, 'r') as f:
+        lines = f.readlines()
+        s = ''.join(lines)
+        res = tokenizer.encode(s)
+        # write to file
+        with open(fname_out, 'w') as f:
+            for x in res:
+                f.write(str(x) + ' ')
+            f.write('\n')
+        print('len(res): ', len(res))
+        print('len(lines): ', len(lines))
+    print('results written to: ', fname_out)
diff --git a/tests/test-tokenizer-0-llama.cpp b/tests/test-tokenizer-0-llama.cpp
new file mode 100644
index 000000000..91c841f7b
--- /dev/null
+++ b/tests/test-tokenizer-0-llama.cpp
@@ -0,0 +1,192 @@
+#include "llama.h"
+#include "common.h"
+#include "console.h"
+
+#include <cstdio>
+#include <string>
+#include <map>
+#include <vector>
+#include <fstream>
+
+// generate using test-tokenizer-0-llama.py
+static const std::map<std::string, std::vector<llama_token>> & k_tests() {
+    static std::map<std::string, std::vector<llama_token>> _k_tests = {
+        { ""                      , {  }, },
+        { " "                     , {     259, }, },
+        { "  "                    , {    1678, }, },
+        { "   "                   , {     268, }, },
+        { "\t"                    , {   29871,     12, }, },
+        { "\n"                    , {   29871,     13, }, },
+        { "\t\n"                  , {   29871,     12,     13, }, },
+        { "Hello world"           , {   15043,   3186, }, },
+        { " Hello world"          , {   29871,  15043,   3186, }, },
+        { "Hello World"           , {   15043,   2787, }, },
+        { " Hello World"          , {   29871,  15043,   2787, }, },
+        { " Hello World!"         , {   29871,  15043,   2787,  29991, }, },
+        { "Hello, world!"         , {   15043,  29892,   3186,  29991, }, },
+        { " Hello, world!"        , {   29871,  15043,  29892,   3186,  29991, }, },
+        { " this is 🦙.cpp"        , {   29871,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
+        { "w048 7tuijk dsdfhu"    , {     281,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
+        { "нещо на Български"     , {    1538,   4851,    665,   1386,  29713,   1305, }, },
+        { "កាន់តែពិសេសអាចខលចេញ"   , {   29871,  31849,  31324,  31934,    228,    162,    142,    228,    161,    146,    228,    162,    133,    228,    161,    153,    228,    161,    186,  31708,    228,    162,    132,  31708,    228,    161,    165,  31324,    228,    161,    136,    228,    161,    132,    228,    161,    158,    228,    161,    136,    228,    162,    132,    228,    161,    140, }, },
+        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", {   29871,    243,    162,    157,    131,    313,   8945,  29897,  29871,    243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,    313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,    313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
+        { "Hello"                 , {   15043, }, },
+        { " Hello"                , {   29871,  15043, }, },
+        { "  Hello"               , {     259,  15043, }, },
+        { "   Hello"              , {    1678,  15043, }, },
+        { "    Hello"             , {     268,  15043, }, },
+        { "    Hello\n    Hello"  , {     268,  15043,     13,   1678,  15043, }, },
+        { " ("                    , {   29871,  313, }, },
+    };
+
+    return _k_tests;
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    std::string fname_text;
+    if (argc > 2) {
+        fname_text = argv[2];
+    }
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto mparams = llama_model_default_params();
+
+        mparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), mparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        auto cparams = llama_context_default_params();
+
+        ctx = llama_new_context_with_model(model, cparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    if (llama_vocab_type(model) != LLAMA_VOCAB_TYPE_SPM) {
+        fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__);
+        llama_free_model(model);
+        llama_free(ctx);
+        return 2;
+    }
+
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
+    bool success = true;
+
+    for (const auto & test_kv : k_tests()) {
+        const std::vector<llama_token> res_bos   = llama_tokenize(ctx, test_kv.first, true);
+        const std::vector<llama_token> res_nobos = llama_tokenize(ctx, test_kv.first, false);
+
+        printf("\n");
+        printf("src: '%s'\n", test_kv.first.c_str());
+        printf("res: '%s'\n", llama_detokenize_spm(ctx, res_bos).c_str());
+        printf("tok: ");
+        for (const auto & tok : res_bos) {
+            printf("%d ", tok);
+        }
+        printf("\n");
+
+        bool correct = res_nobos.size() == test_kv.second.size() && res_bos.size() == res_nobos.size() + 1 && res_bos[0] == 1;
+
+        for (int i = 0; i < (int) res_nobos.size() && correct; ++i) {
+            if (test_kv.second[i] != res_bos[i + 1]) {
+                correct = false;
+            }
+            if (test_kv.second[i] != res_nobos[i]) {
+                correct = false;
+            }
+        }
+
+        if (!correct) {
+            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                llama_detokenize_spm(ctx, res_nobos).c_str(),
+                llama_detokenize_spm(ctx, test_kv.second).c_str());
+            fprintf(stderr, "%s : expected tokens: ", __func__);
+            for (const auto & t : test_kv.second) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : got tokens:      ", __func__);
+            for (const auto & t : res_nobos) {
+                fprintf(stderr, "%6d, ", t);
+            }
+            fprintf(stderr, "\n");
+
+            success = false;
+        }
+    }
+
+    if (!fname_text.empty()) {
+        fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
+
+        std::string text;
+        {
+            std::ifstream ifs(fname_text);
+            if (!ifs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str());
+                return 1;
+            }
+            text = std::string(std::istreambuf_iterator<char>(ifs), std::istreambuf_iterator<char>());
+        }
+
+        fprintf(stderr, "%s : text size: %zu\n", __func__, text.size());
+
+        const std::vector<llama_token> res = llama_tokenize(ctx, text, true);
+
+        fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size());
+
+        {
+            const std::string fname_out = fname_text + ".tokcpp";
+
+            std::ofstream ofs(fname_out);
+            if (!ofs) {
+                fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str());
+                return 1;
+            }
+
+            for (const auto & tok : res) {
+                ofs << tok << " ";
+            }
+
+            ofs << "\n";
+        }
+
+        fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str());
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return success ? 0 : 3;
+}
diff --git a/tests/test-tokenizer-0-llama.py b/tests/test-tokenizer-0-llama.py
new file mode 100644
index 000000000..bc164ee29
--- /dev/null
+++ b/tests/test-tokenizer-0-llama.py
@@ -0,0 +1,95 @@
+# tests with SPM tokenizer
+
+import os
+import sys
+import argparse
+
+from sentencepiece import SentencePieceProcessor
+
+parser = argparse.ArgumentParser()
+parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
+parser.add_argument("--fname-tok",   help="path to a text file to tokenize")
+args = parser.parse_args()
+
+dir_tokenizer = args.dir_tokenizer
+
+tokenizer = SentencePieceProcessor(dir_tokenizer + '/tokenizer.model')
+
+tests = [
+        "",
+        " ",
+        "  ",
+        "   ",
+        "\t",
+        "\n",
+        "\t\n",
+        "Hello world",
+        " Hello world",
+        "Hello World",
+        " Hello World",
+        " Hello World!",
+        "Hello, world!",
+        " Hello, world!",
+        " this is 🦙.cpp",
+        "w048 7tuijk dsdfhu",
+        "нещо на Български",
+        "កាន់តែពិសេសអាចខលចេញ",
+        "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+        "Hello",
+        " Hello",
+        "  Hello",
+        "   Hello",
+        "    Hello",
+        "    Hello\n    Hello",
+    ]
+
+
+for text in tests:
+    print('text: ', text)
+    print('\nwith bos:')
+    print(tokenizer.encode(text, add_bos=True))
+    print(tokenizer.decode(tokenizer.encode(text, add_bos=True)))
+    print('\nwithout bos:')
+    print(tokenizer.encode(text, add_bos=False))
+    print(tokenizer.decode(tokenizer.encode(text, add_bos=False)))
+
+print("'" + tokenizer.id_to_piece(15043) + "'") # '_Hello'
+print("'" + tokenizer.id_to_piece(29871) + "'") # '_'
+print("'" + tokenizer.decode([15043]) + "'")        # 'Hello'
+print("'" + tokenizer.decode([15043, 15043]) + "'") # 'Hello Hello'
+print("'" + tokenizer.decode([29871, 15043]) + "'")               # ' Hello'
+print("'" + tokenizer.decode([29871, 15043, 29871, 15043]) + "'") # ' Hello  Hello'
+
+print("\n\ntests for C++:\n")
+for text in tests:
+    res = tokenizer.encode(text, add_bos=False)
+
+    k = text.replace('\n', '\\n')
+    k = k.replace('\t', '\\t')
+    k = '"' + k + '"'
+    print("{ %-24s, { " % k, end='')
+    for x in res:
+        print("%7d," % x, end='')
+    print(" }, },")
+
+print(tokenizer.encode('hello'))
+print(tokenizer.encode('world'))
+print(tokenizer.encode(' world'))
+print(tokenizer.encode('hello world'))
+
+fname_tok = args.fname_tok
+if fname_tok:
+    print('tokenizing file: ', fname_tok)
+    fname_out = fname_tok + '.tok'
+    with open(fname_tok, 'r') as f:
+        lines = f.readlines()
+        s = ''.join(lines)
+        res = tokenizer.encode(s, add_bos=True)
+        # write to file
+        with open(fname_out, 'w') as f:
+            for x in res:
+                f.write(str(x) + ' ')
+            f.write('\n')
+        print('len(res): ', len(res))
+        print('len(lines): ', len(lines))
+    print('results written to: ', fname_out)
diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp
deleted file mode 100644
index f3ee851a3..000000000
--- a/tests/test-tokenizer-0.cpp
+++ /dev/null
@@ -1,140 +0,0 @@
-#include "llama.h"
-#include "common.h"
-
-#include <cstdio>
-#include <string>
-#include <map>
-#include <vector>
-
-static std::string unescape_whitespace(llama_context* ctx, const std::vector<llama_token>& tokens) {
-    std::string result;
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        result += llama_token_to_str(ctx, tokens[i]);
-    }
-    return result;
-}
-
-static const std::map<std::string, std::vector<llama_token>> & k_tests() {
-    static std::map<std::string, std::vector<llama_token>> _k_tests = {
-        { " ",                      {1,    259, }, },
-        { "  ",                     { 1,    1678, }, },
-        { "   ",                    { 1,     268, }, },
-        { "\t",                     { 1,    29871,   12, }, },
-        { "\n",                     { 1,    29871,   13, }, },
-        { "\t\n",                   { 1,    29871,   12,     13, }, },
-        { "Hello world",            { 1,  15043,   3186, }, },
-        { " Hello world",           { 1,  29871,  15043,   3186, }, },
-        { "Hello World",            { 1,  15043,   2787, }, },
-        { " Hello World",           { 1,  29871,  15043,   2787, }, },
-        { " Hello World!",          { 1,  29871,  15043,   2787,  29991, }, },
-        { " this is 🦙.cpp",        { 1,  29871,    445,    338,  29871,    243,    162,    169,    156,  29889,   8223, }, },
-        { "w048 7tuijk dsdfhu",     { 1,    281,  29900,  29946,  29947,  29871,  29955,   9161,  13535,  18031,   2176,   6905, }, },
-        { "нещо на Български",      { 1,   1538,   4851,    665,   1386,  29713,   1305, }, },
-        { "កាន់តែពិសេសអាចខលចេញ",   { 1,  29871,  31849,  31324,  31934,    228,    162,    142,    228,    161,
-                                     146,    228,    162,    133,    228,    161,    153,    228,    161,    186,
-                                     31708,    228,    162,    132,  31708,    228,    161,    165,  31324,    228,
-                                     161,    136,    228,    161,    132,    228,    161,    158,    228,    161,
-                                     136,    228,    162,    132,    228,    161,    140, }, },
-        { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
-            { 1,  29871,    243,    162,    157,    131,    313,   8945,  29897,  29871,
-                243,    162,    155,    185,  30722,    243,    162,    143,    174,  30598,
-                313,  20787,    953,   3848,    275,  16125,    630,  29897,  29871,  31681,
-                313,   6194,    953,  29877,   2397,    393,    756,    967,   1914,   5993,  29897, }, },
-        { "Hello",                  { 1,    15043 }, },
-        { " Hello",                 { 1,    29871,  15043 }, },
-        { "  Hello",                { 1,    259,    15043 }, },
-        { "   Hello",               { 1,    1678,   15043 }, },
-        { "    Hello",              { 1,    268,    15043 }, },
-        { "    Hello\n    Hello",   { 1,    268,    15043,  13,     1678,   15043 }, },
-    };
-
-    return _k_tests;
-}
-
-int main(int argc, char **argv) {
-    if (argc < 2) {
-        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
-        return 1;
-    }
-
-    const std::string fname = argv[1];
-
-    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
-
-    llama_model * model;
-    llama_context * ctx;
-
-    llama_backend_init(false);
-
-    // load the vocab
-    {
-        auto lparams = llama_context_default_params();
-
-        lparams.vocab_only = true;
-
-        model = llama_load_model_from_file(fname.c_str(), lparams);
-
-        if (model == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            return 1;
-        }
-
-        ctx = llama_new_context_with_model(model, lparams);
-
-        if (ctx == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            llama_free_model(model);
-            return 1;
-        }
-    }
-
-    const int n_vocab = llama_n_vocab(ctx);
-
-    if (n_vocab != 32000) {
-        fprintf(stderr, "%s : expected 32000 tokens, got %d\n", __func__, n_vocab);
-        llama_free_model(model);
-        llama_free(ctx);
-        return 2;
-    }
-
-    bool success = true;
-
-    for (const auto & test_kv : k_tests()) {
-        std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, true);
-        fprintf(stderr, "%s : '%s' tokenized to '%s'\n",
-            __func__, test_kv.first.c_str(), unescape_whitespace(ctx, res).c_str());
-
-        bool correct = res.size() == test_kv.second.size();
-
-        for (int i = 0; i < (int) res.size() && correct; ++i) {
-            if (res[i] != test_kv.second[i]) {
-                correct = false;
-            }
-        }
-
-        if (!correct) {
-            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
-                unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str());
-            fprintf(stderr, "%s : expected tokens: ", __func__);
-            for (const auto & t : test_kv.second) {
-                fprintf(stderr, "%6d, ", t);
-            }
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s : got tokens:      ", __func__);
-            for (const auto & t : res) {
-                fprintf(stderr, "%6d, ", t);
-            }
-            fprintf(stderr, "\n");
-
-            success = false;
-        }
-    }
-
-    llama_free_model(model);
-    llama_free(ctx);
-
-    llama_backend_free();
-
-    return success ? 0 : 3;
-}
diff --git a/tests/test-tokenizer-1-bpe.cpp b/tests/test-tokenizer-1-bpe.cpp
new file mode 100644
index 000000000..85a59a14d
--- /dev/null
+++ b/tests/test-tokenizer-1-bpe.cpp
@@ -0,0 +1,113 @@
+#include "llama.h"
+#include "common.h"
+#include "unicode.h"
+#include "console.h"
+
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <codecvt>
+#include <map>
+#include <vector>
+#include <locale>
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto mparams = llama_model_default_params();
+
+        mparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), mparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        auto cparams = llama_context_default_params();
+
+        ctx = llama_new_context_with_model(model, cparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    GGML_ASSERT(llama_vocab_type(model) == LLAMA_VOCAB_TYPE_BPE);
+
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
+    const int n_vocab = llama_n_vocab(model);
+
+    for (int i = 0; i < n_vocab; ++i) {
+        std::string str = llama_detokenize_bpe(ctx, std::vector<int>(1, i));
+        try {
+            auto cps = codepoints_from_utf8(str);
+            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+            std::string check = llama_detokenize_bpe(ctx, tokens);
+            if (check != str) {
+                fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
+                    __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
+                return 2;
+            }
+        }
+        catch (const std::invalid_argument &) {
+            fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
+        }
+    }
+
+    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
+        // NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
+        if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
+            std::string str = " " + codepoint_to_utf8(cp);
+            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+            std::string check = llama_detokenize_bpe(ctx, tokens);
+            if (str != check) {
+                fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+                return 3;
+            }
+        }
+    }
+    // TODO: why doesn't this work for the full range of Unicodes?
+    // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
+    for (uint32_t cp = 0x10000; cp < 0x00080000; ++cp) {
+        std::string str = codepoint_to_utf8(cp);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+        std::string check = llama_detokenize_bpe(ctx, tokens);
+        if (str != check) {
+            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+            return 4;
+        }
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return 0;
+}
diff --git a/tests/test-tokenizer-1-llama.cpp b/tests/test-tokenizer-1-llama.cpp
new file mode 100644
index 000000000..4b58fe495
--- /dev/null
+++ b/tests/test-tokenizer-1-llama.cpp
@@ -0,0 +1,104 @@
+#include "llama.h"
+#include "common.h"
+#include "unicode.h"
+#include "console.h"
+
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <codecvt>
+#include <map>
+#include <vector>
+#include <locale>
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
+        return 1;
+    }
+
+    const std::string fname = argv[1];
+
+    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
+
+    llama_model * model;
+    llama_context * ctx;
+
+    llama_backend_init(false);
+
+    // load the vocab
+    {
+        auto mparams = llama_model_default_params();
+
+        mparams.vocab_only = true;
+
+        model = llama_load_model_from_file(fname.c_str(), mparams);
+
+        if (model == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            return 1;
+        }
+
+        auto cparams = llama_context_default_params();
+
+        ctx = llama_new_context_with_model(model, cparams);
+
+        if (ctx == NULL) {
+            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
+            llama_free_model(model);
+            return 1;
+        }
+    }
+
+    GGML_ASSERT(llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
+
+#ifdef _WIN32
+    // We need this for unicode console support
+    console::init(false, false);
+    atexit([]() { console::cleanup(); });
+#endif
+
+    const int n_vocab = llama_n_vocab(model);
+
+    for (int i = 0; i < n_vocab; ++i) {
+        std::string str = llama_detokenize_spm(ctx, std::vector<int>(1, i));
+        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+        std::string check = llama_detokenize_spm(ctx, tokens);
+        if (check != str) {
+            fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
+                __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
+            return 2;
+        }
+    }
+
+    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
+        if (cp < 0xd800 || cp > 0xdfff) {
+            std::string str = codepoint_to_utf8(cp);
+            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+            std::string check = llama_detokenize_spm(ctx, tokens);
+            if (cp != 9601 && str != check) {
+                fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+                return 3;
+            }
+        }
+    }
+    for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
+        std::string str = codepoint_to_utf8(cp);
+        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+        std::string check = llama_detokenize_spm(ctx, tokens);
+        if (str != check) {
+            fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
+            return 4;
+        }
+    }
+
+    llama_free_model(model);
+    llama_free(ctx);
+
+    llama_backend_free();
+
+    return 0;
+}
diff --git a/tests/test-tokenizer-1.cpp b/tests/test-tokenizer-1.cpp
deleted file mode 100644
index 993d17f18..000000000
--- a/tests/test-tokenizer-1.cpp
+++ /dev/null
@@ -1,124 +0,0 @@
-#include "llama.h"
-#include "common.h"
-
-#include <cassert>
-#include <cstdio>
-#include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
-#include <locale>
-
-static std::string escape_whitespace(const std::string& text) {
-    std::string result = "\xe2\x96\x81";
-    for (size_t offs = 0; offs < text.length(); ++offs) {
-        if (text[offs] == ' ') {
-            result += "\xe2\x96\x81";
-        } else {
-            result += text[offs];
-        }
-    }
-    return result;
-}
-
-static std::string unescape_whitespace(llama_context * ctx, const std::vector<llama_token> & tokens) {
-    std::string result;
-    for (size_t i = 0; i < tokens.size(); ++i) {
-        result += llama_token_to_str(ctx, tokens[i]);
-    }
-    return result;
-}
-
-int main(int argc, char **argv) {
-    if (argc < 2) {
-        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
-        return 1;
-    }
-
-    const std::string fname = argv[1];
-
-    fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
-
-    llama_model * model;
-    llama_context * ctx;
-
-    llama_backend_init(false);
-
-    // load the vocab
-    {
-        auto lparams = llama_context_default_params();
-
-        lparams.vocab_only = true;
-
-        model = llama_load_model_from_file(fname.c_str(), lparams);
-
-        if (model == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            return 1;
-        }
-
-        ctx = llama_new_context_with_model(model, lparams);
-
-        if (ctx == NULL) {
-            fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
-            llama_free_model(model);
-            return 1;
-        }
-    }
-
-    const int n_vocab = llama_n_vocab(ctx);
-
-    for (int i = 0; i < n_vocab; ++i) {
-        std::string forward = llama_token_to_str_bpe(ctx, i);
-        std::vector<llama_token> tokens = llama_tokenize_bpe(ctx, forward, false);
-        if (tokens.size() == 1) {
-            if (i != tokens[0]) {
-                std::string backward = llama_token_to_str(ctx, tokens[0]);
-                fprintf(stderr, "%s : error: token %d is string %s but bpe returns token %d %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str());
-                return 2;
-            }
-        } else {
-            llama_token_type type = llama_token_get_type(ctx, i);
-            if (type == LLAMA_TOKEN_TYPE_UNKNOWN || type == LLAMA_TOKEN_TYPE_CONTROL || type == LLAMA_TOKEN_TYPE_BYTE) {
-                fprintf(stderr, "%s : info: token %d is string %s and bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-            } else {
-                fprintf(stderr, "%s : error: token %d is string %s but bpe returns tokens %s\n",
-                    __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str());
-                return 2;
-            }
-        }
-    }
-
-#ifdef _WIN32
-    std::wstring_convert<typename std::codecvt_utf8<char16_t>, char16_t> u16converter;
-    for (char16_t ch = 0x0000; ch < 0xffff; ++ch) {
-        std::u16string u16str(1, ch);
-        std::string str = u16converter.to_bytes(u16str);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, escape_whitespace(str).c_str(), false);
-        if (tokens.size() == 1) {
-            fprintf(stderr, "%s : info: %s tokenized to %d \n",
-                __func__, str.c_str(), tokens[0]);
-        }
-    }
-
-    std::wstring_convert<typename std::codecvt_utf8<char32_t>, char32_t> u32converter;
-    for (char32_t ch = 0x0000; ch < 0x0010ffff; ++ch) {
-        std::u32string u32str(1, ch);
-        std::string str = u32converter.to_bytes(u32str);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, escape_whitespace(str).c_str(), false);
-        if (tokens.size() == 1) {
-            fprintf(stderr, "%s : info: %s tokenized to %d \n", __func__, str.c_str(), tokens[0]);
-        }
-    }
-#endif
-
-    llama_free_model(model);
-    llama_free(ctx);
-
-    llama_backend_free();
-
-    return 0;
-}
diff --git a/unicode.h b/unicode.h
new file mode 100644
index 000000000..aeca879ea
--- /dev/null
+++ b/unicode.h
@@ -0,0 +1,462 @@
+#pragma once
+
+#include <cassert>
+#include <stdexcept>
+#include <vector>
+#include <unordered_map>
+
+static const std::vector<std::pair<uint32_t, uint32_t>> digit_ranges = {
+{0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F},
+{0xCE6, 0xCEF}, {0xD66, 0xD6F}, {0xDE6, 0xDEF}, {0xE50, 0xE59}, {0xED0, 0xED9}, {0xF20, 0xF29}, {0x1040, 0x1049}, {0x1090, 0x1099}, {0x1369, 0x1371}, {0x17E0, 0x17E9}, {0x1810, 0x1819}, {0x1946, 0x194F},
+{0x19D0, 0x19DA}, {0x1A80, 0x1A89}, {0x1A90, 0x1A99}, {0x1B50, 0x1B59}, {0x1BB0, 0x1BB9}, {0x1C40, 0x1C49}, {0x1C50, 0x1C59}, {0x2070, 0x2070}, {0x2074, 0x2079}, {0x2080, 0x2089}, {0x2460, 0x2468},
+{0x2474, 0x247C}, {0x2488, 0x2490}, {0x24EA, 0x24EA}, {0x24F5, 0x24FD}, {0x24FF, 0x24FF}, {0x2776, 0x277E}, {0x2780, 0x2788}, {0x278A, 0x2792}, {0xA620, 0xA629}, {0xA8D0, 0xA8D9}, {0xA900, 0xA909},
+{0xA9D0, 0xA9D9}, {0xA9F0, 0xA9F9}, {0xAA50, 0xAA59}, {0xABF0, 0xABF9}, {0xFF10, 0xFF19}, {0x104A0, 0x104A9}, {0x10A40, 0x10A43}, {0x10D30, 0x10D39}, {0x10E60, 0x10E68}, {0x11052, 0x1105A},
+{0x11066, 0x1106F}, {0x110F0, 0x110F9}, {0x11136, 0x1113F}, {0x111D0, 0x111D9}, {0x112F0, 0x112F9}, {0x11450, 0x11459}, {0x114D0, 0x114D9}, {0x11650, 0x11659}, {0x116C0, 0x116C9}, {0x11730, 0x11739},
+{0x118E0, 0x118E9}, {0x11950, 0x11959}, {0x11C50, 0x11C59}, {0x11D50, 0x11D59}, {0x11DA0, 0x11DA9}, {0x16A60, 0x16A69}, {0x16B50, 0x16B59}, {0x1D7CE, 0x1D7FF}, {0x1E140, 0x1E149}, {0x1E2F0, 0x1E2F9},
+{0x1E950, 0x1E959}, {0x1F100, 0x1F10A}, {0x1FBF0, 0x1FBF9},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> letter_ranges = {
+{0x41, 0x5A}, {0x61, 0x7A}, {0xAA, 0xAA}, {0xB5, 0xB5}, {0xBA, 0xBA}, {0xC0, 0xD6}, {0xD8, 0xF6}, {0xF8, 0x2C1}, {0x2C6, 0x2D1}, {0x2E0, 0x2E4}, {0x2EC, 0x2EC}, {0x2EE, 0x2EE}, {0x370, 0x374},
+{0x376, 0x377}, {0x37A, 0x37D}, {0x37F, 0x37F}, {0x386, 0x386}, {0x388, 0x38A}, {0x38C, 0x38C}, {0x38E, 0x3A1}, {0x3A3, 0x3F5}, {0x3F7, 0x481}, {0x48A, 0x52F}, {0x531, 0x556}, {0x559, 0x559},
+{0x560, 0x588}, {0x5D0, 0x5EA}, {0x5EF, 0x5F2}, {0x620, 0x64A}, {0x66E, 0x66F}, {0x671, 0x6D3}, {0x6D5, 0x6D5}, {0x6E5, 0x6E6}, {0x6EE, 0x6EF}, {0x6FA, 0x6FC}, {0x6FF, 0x6FF}, {0x710, 0x710},
+{0x712, 0x72F}, {0x74D, 0x7A5}, {0x7B1, 0x7B1}, {0x7CA, 0x7EA}, {0x7F4, 0x7F5}, {0x7FA, 0x7FA}, {0x800, 0x815}, {0x81A, 0x81A}, {0x824, 0x824}, {0x828, 0x828}, {0x840, 0x858}, {0x860, 0x86A},
+{0x8A0, 0x8B4}, {0x8B6, 0x8C7}, {0x904, 0x939}, {0x93D, 0x93D}, {0x950, 0x950}, {0x958, 0x961}, {0x971, 0x980}, {0x985, 0x98C}, {0x98F, 0x990}, {0x993, 0x9A8}, {0x9AA, 0x9B0}, {0x9B2, 0x9B2},
+{0x9B6, 0x9B9}, {0x9BD, 0x9BD}, {0x9CE, 0x9CE}, {0x9DC, 0x9DD}, {0x9DF, 0x9E1}, {0x9F0, 0x9F1}, {0x9FC, 0x9FC}, {0xA05, 0xA0A}, {0xA0F, 0xA10}, {0xA13, 0xA28}, {0xA2A, 0xA30}, {0xA32, 0xA33},
+{0xA35, 0xA36}, {0xA38, 0xA39}, {0xA59, 0xA5C}, {0xA5E, 0xA5E}, {0xA72, 0xA74}, {0xA85, 0xA8D}, {0xA8F, 0xA91}, {0xA93, 0xAA8}, {0xAAA, 0xAB0}, {0xAB2, 0xAB3}, {0xAB5, 0xAB9}, {0xABD, 0xABD},
+{0xAD0, 0xAD0}, {0xAE0, 0xAE1}, {0xAF9, 0xAF9}, {0xB05, 0xB0C}, {0xB0F, 0xB10}, {0xB13, 0xB28}, {0xB2A, 0xB30}, {0xB32, 0xB33}, {0xB35, 0xB39}, {0xB3D, 0xB3D}, {0xB5C, 0xB5D}, {0xB5F, 0xB61},
+{0xB71, 0xB71}, {0xB83, 0xB83}, {0xB85, 0xB8A}, {0xB8E, 0xB90}, {0xB92, 0xB95}, {0xB99, 0xB9A}, {0xB9C, 0xB9C}, {0xB9E, 0xB9F}, {0xBA3, 0xBA4}, {0xBA8, 0xBAA}, {0xBAE, 0xBB9}, {0xBD0, 0xBD0},
+{0xC05, 0xC0C}, {0xC0E, 0xC10}, {0xC12, 0xC28}, {0xC2A, 0xC39}, {0xC3D, 0xC3D}, {0xC58, 0xC5A}, {0xC60, 0xC61}, {0xC80, 0xC80}, {0xC85, 0xC8C}, {0xC8E, 0xC90}, {0xC92, 0xCA8}, {0xCAA, 0xCB3},
+{0xCB5, 0xCB9}, {0xCBD, 0xCBD}, {0xCDE, 0xCDE}, {0xCE0, 0xCE1}, {0xCF1, 0xCF2}, {0xD04, 0xD0C}, {0xD0E, 0xD10}, {0xD12, 0xD3A}, {0xD3D, 0xD3D}, {0xD4E, 0xD4E}, {0xD54, 0xD56}, {0xD5F, 0xD61},
+{0xD7A, 0xD7F}, {0xD85, 0xD96}, {0xD9A, 0xDB1}, {0xDB3, 0xDBB}, {0xDBD, 0xDBD}, {0xDC0, 0xDC6}, {0xE01, 0xE30}, {0xE32, 0xE33}, {0xE40, 0xE46}, {0xE81, 0xE82}, {0xE84, 0xE84}, {0xE86, 0xE8A},
+{0xE8C, 0xEA3}, {0xEA5, 0xEA5}, {0xEA7, 0xEB0}, {0xEB2, 0xEB3}, {0xEBD, 0xEBD}, {0xEC0, 0xEC4}, {0xEC6, 0xEC6}, {0xEDC, 0xEDF}, {0xF00, 0xF00}, {0xF40, 0xF47}, {0xF49, 0xF6C}, {0xF88, 0xF8C},
+{0x1000, 0x102A}, {0x103F, 0x103F}, {0x1050, 0x1055}, {0x105A, 0x105D}, {0x1061, 0x1061}, {0x1065, 0x1066}, {0x106E, 0x1070}, {0x1075, 0x1081}, {0x108E, 0x108E}, {0x10A0, 0x10C5}, {0x10C7, 0x10C7},
+{0x10CD, 0x10CD}, {0x10D0, 0x10FA}, {0x10FC, 0x1248}, {0x124A, 0x124D}, {0x1250, 0x1256}, {0x1258, 0x1258}, {0x125A, 0x125D}, {0x1260, 0x1288}, {0x128A, 0x128D}, {0x1290, 0x12B0}, {0x12B2, 0x12B5},
+{0x12B8, 0x12BE}, {0x12C0, 0x12C0}, {0x12C2, 0x12C5}, {0x12C8, 0x12D6}, {0x12D8, 0x1310}, {0x1312, 0x1315}, {0x1318, 0x135A}, {0x1380, 0x138F}, {0x13A0, 0x13F5}, {0x13F8, 0x13FD}, {0x1401, 0x166C},
+{0x166F, 0x167F}, {0x1681, 0x169A}, {0x16A0, 0x16EA}, {0x16F1, 0x16F8}, {0x1700, 0x170C}, {0x170E, 0x1711}, {0x1720, 0x1731}, {0x1740, 0x1751}, {0x1760, 0x176C}, {0x176E, 0x1770}, {0x1780, 0x17B3},
+{0x17D7, 0x17D7}, {0x17DC, 0x17DC}, {0x1820, 0x1878}, {0x1880, 0x1884}, {0x1887, 0x18A8}, {0x18AA, 0x18AA}, {0x18B0, 0x18F5}, {0x1900, 0x191E}, {0x1950, 0x196D}, {0x1970, 0x1974}, {0x1980, 0x19AB},
+{0x19B0, 0x19C9}, {0x1A00, 0x1A16}, {0x1A20, 0x1A54}, {0x1AA7, 0x1AA7}, {0x1B05, 0x1B33}, {0x1B45, 0x1B4B}, {0x1B83, 0x1BA0}, {0x1BAE, 0x1BAF}, {0x1BBA, 0x1BE5}, {0x1C00, 0x1C23}, {0x1C4D, 0x1C4F},
+{0x1C5A, 0x1C7D}, {0x1C80, 0x1C88}, {0x1C90, 0x1CBA}, {0x1CBD, 0x1CBF}, {0x1CE9, 0x1CEC}, {0x1CEE, 0x1CF3}, {0x1CF5, 0x1CF6}, {0x1CFA, 0x1CFA}, {0x1D00, 0x1DBF}, {0x1E00, 0x1F15}, {0x1F18, 0x1F1D},
+{0x1F20, 0x1F45}, {0x1F48, 0x1F4D}, {0x1F50, 0x1F57}, {0x1F59, 0x1F59}, {0x1F5B, 0x1F5B}, {0x1F5D, 0x1F5D}, {0x1F5F, 0x1F7D}, {0x1F80, 0x1FB4}, {0x1FB6, 0x1FBC}, {0x1FBE, 0x1FBE}, {0x1FC2, 0x1FC4},
+{0x1FC6, 0x1FCC}, {0x1FD0, 0x1FD3}, {0x1FD6, 0x1FDB}, {0x1FE0, 0x1FEC}, {0x1FF2, 0x1FF4}, {0x1FF6, 0x1FFC}, {0x2071, 0x2071}, {0x207F, 0x207F}, {0x2090, 0x209C}, {0x2102, 0x2102}, {0x2107, 0x2107},
+{0x210A, 0x2113}, {0x2115, 0x2115}, {0x2119, 0x211D}, {0x2124, 0x2124}, {0x2126, 0x2126}, {0x2128, 0x2128}, {0x212A, 0x212D}, {0x212F, 0x2139}, {0x213C, 0x213F}, {0x2145, 0x2149}, {0x214E, 0x214E},
+{0x2183, 0x2184}, {0x2C00, 0x2C2E}, {0x2C30, 0x2C5E}, {0x2C60, 0x2CE4}, {0x2CEB, 0x2CEE}, {0x2CF2, 0x2CF3}, {0x2D00, 0x2D25}, {0x2D27, 0x2D27}, {0x2D2D, 0x2D2D}, {0x2D30, 0x2D67}, {0x2D6F, 0x2D6F},
+{0x2D80, 0x2D96}, {0x2DA0, 0x2DA6}, {0x2DA8, 0x2DAE}, {0x2DB0, 0x2DB6}, {0x2DB8, 0x2DBE}, {0x2DC0, 0x2DC6}, {0x2DC8, 0x2DCE}, {0x2DD0, 0x2DD6}, {0x2DD8, 0x2DDE}, {0x2E2F, 0x2E2F}, {0x3005, 0x3006},
+{0x3031, 0x3035}, {0x303B, 0x303C}, {0x3041, 0x3096}, {0x309D, 0x309F}, {0x30A1, 0x30FA}, {0x30FC, 0x30FF}, {0x3105, 0x312F}, {0x3131, 0x318E}, {0x31A0, 0x31BF}, {0x31F0, 0x31FF}, {0x3400, 0x4DBF},
+{0x4E00, 0x9FFC}, {0xA000, 0xA48C}, {0xA4D0, 0xA4FD}, {0xA500, 0xA60C}, {0xA610, 0xA61F}, {0xA62A, 0xA62B}, {0xA640, 0xA66E}, {0xA67F, 0xA69D}, {0xA6A0, 0xA6E5}, {0xA717, 0xA71F}, {0xA722, 0xA788},
+{0xA78B, 0xA7BF}, {0xA7C2, 0xA7CA}, {0xA7F5, 0xA801}, {0xA803, 0xA805}, {0xA807, 0xA80A}, {0xA80C, 0xA822}, {0xA840, 0xA873}, {0xA882, 0xA8B3}, {0xA8F2, 0xA8F7}, {0xA8FB, 0xA8FB}, {0xA8FD, 0xA8FE},
+{0xA90A, 0xA925}, {0xA930, 0xA946}, {0xA960, 0xA97C}, {0xA984, 0xA9B2}, {0xA9CF, 0xA9CF}, {0xA9E0, 0xA9E4}, {0xA9E6, 0xA9EF}, {0xA9FA, 0xA9FE}, {0xAA00, 0xAA28}, {0xAA40, 0xAA42}, {0xAA44, 0xAA4B},
+{0xAA60, 0xAA76}, {0xAA7A, 0xAA7A}, {0xAA7E, 0xAAAF}, {0xAAB1, 0xAAB1}, {0xAAB5, 0xAAB6}, {0xAAB9, 0xAABD}, {0xAAC0, 0xAAC0}, {0xAAC2, 0xAAC2}, {0xAADB, 0xAADD}, {0xAAE0, 0xAAEA}, {0xAAF2, 0xAAF4},
+{0xAB01, 0xAB06}, {0xAB09, 0xAB0E}, {0xAB11, 0xAB16}, {0xAB20, 0xAB26}, {0xAB28, 0xAB2E}, {0xAB30, 0xAB5A}, {0xAB5C, 0xAB69}, {0xAB70, 0xABE2}, {0xAC00, 0xD7A3}, {0xD7B0, 0xD7C6}, {0xD7CB, 0xD7FB},
+{0xF900, 0xFA6D}, {0xFA70, 0xFAD9}, {0xFB00, 0xFB06}, {0xFB13, 0xFB17}, {0xFB1D, 0xFB1D}, {0xFB1F, 0xFB28}, {0xFB2A, 0xFB36}, {0xFB38, 0xFB3C}, {0xFB3E, 0xFB3E}, {0xFB40, 0xFB41}, {0xFB43, 0xFB44},
+{0xFB46, 0xFBB1}, {0xFBD3, 0xFD3D}, {0xFD50, 0xFD8F}, {0xFD92, 0xFDC7}, {0xFDF0, 0xFDFB}, {0xFE70, 0xFE74}, {0xFE76, 0xFEFC}, {0xFF21, 0xFF3A}, {0xFF41, 0xFF5A}, {0xFF66, 0xFFBE}, {0xFFC2, 0xFFC7},
+{0xFFCA, 0xFFCF}, {0xFFD2, 0xFFD7}, {0xFFDA, 0xFFDC}, {0x10000, 0x1000B}, {0x1000D, 0x10026}, {0x10028, 0x1003A}, {0x1003C, 0x1003D}, {0x1003F, 0x1004D}, {0x10050, 0x1005D}, {0x10080, 0x100FA},
+{0x10280, 0x1029C}, {0x102A0, 0x102D0}, {0x10300, 0x1031F}, {0x1032D, 0x10340}, {0x10342, 0x10349}, {0x10350, 0x10375}, {0x10380, 0x1039D}, {0x103A0, 0x103C3}, {0x103C8, 0x103CF}, {0x10400, 0x1049D},
+{0x104B0, 0x104D3}, {0x104D8, 0x104FB}, {0x10500, 0x10527}, {0x10530, 0x10563}, {0x10600, 0x10736}, {0x10740, 0x10755}, {0x10760, 0x10767}, {0x10800, 0x10805}, {0x10808, 0x10808}, {0x1080A, 0x10835},
+{0x10837, 0x10838}, {0x1083C, 0x1083C}, {0x1083F, 0x10855}, {0x10860, 0x10876}, {0x10880, 0x1089E}, {0x108E0, 0x108F2}, {0x108F4, 0x108F5}, {0x10900, 0x10915}, {0x10920, 0x10939}, {0x10980, 0x109B7},
+{0x109BE, 0x109BF}, {0x10A00, 0x10A00}, {0x10A10, 0x10A13}, {0x10A15, 0x10A17}, {0x10A19, 0x10A35}, {0x10A60, 0x10A7C}, {0x10A80, 0x10A9C}, {0x10AC0, 0x10AC7}, {0x10AC9, 0x10AE4}, {0x10B00, 0x10B35},
+{0x10B40, 0x10B55}, {0x10B60, 0x10B72}, {0x10B80, 0x10B91}, {0x10C00, 0x10C48}, {0x10C80, 0x10CB2}, {0x10CC0, 0x10CF2}, {0x10D00, 0x10D23}, {0x10E80, 0x10EA9}, {0x10EB0, 0x10EB1}, {0x10F00, 0x10F1C},
+{0x10F27, 0x10F27}, {0x10F30, 0x10F45}, {0x10FB0, 0x10FC4}, {0x10FE0, 0x10FF6}, {0x11003, 0x11037}, {0x11083, 0x110AF}, {0x110D0, 0x110E8}, {0x11103, 0x11126}, {0x11144, 0x11144}, {0x11147, 0x11147},
+{0x11150, 0x11172}, {0x11176, 0x11176}, {0x11183, 0x111B2}, {0x111C1, 0x111C4}, {0x111DA, 0x111DA}, {0x111DC, 0x111DC}, {0x11200, 0x11211}, {0x11213, 0x1122B}, {0x11280, 0x11286}, {0x11288, 0x11288},
+{0x1128A, 0x1128D}, {0x1128F, 0x1129D}, {0x1129F, 0x112A8}, {0x112B0, 0x112DE}, {0x11305, 0x1130C}, {0x1130F, 0x11310}, {0x11313, 0x11328}, {0x1132A, 0x11330}, {0x11332, 0x11333}, {0x11335, 0x11339},
+{0x1133D, 0x1133D}, {0x11350, 0x11350}, {0x1135D, 0x11361}, {0x11400, 0x11434}, {0x11447, 0x1144A}, {0x1145F, 0x11461}, {0x11480, 0x114AF}, {0x114C4, 0x114C5}, {0x114C7, 0x114C7}, {0x11580, 0x115AE},
+{0x115D8, 0x115DB}, {0x11600, 0x1162F}, {0x11644, 0x11644}, {0x11680, 0x116AA}, {0x116B8, 0x116B8}, {0x11700, 0x1171A}, {0x11800, 0x1182B}, {0x118A0, 0x118DF}, {0x118FF, 0x11906}, {0x11909, 0x11909},
+{0x1190C, 0x11913}, {0x11915, 0x11916}, {0x11918, 0x1192F}, {0x1193F, 0x1193F}, {0x11941, 0x11941}, {0x119A0, 0x119A7}, {0x119AA, 0x119D0}, {0x119E1, 0x119E1}, {0x119E3, 0x119E3}, {0x11A00, 0x11A00},
+{0x11A0B, 0x11A32}, {0x11A3A, 0x11A3A}, {0x11A50, 0x11A50}, {0x11A5C, 0x11A89}, {0x11A9D, 0x11A9D}, {0x11AC0, 0x11AF8}, {0x11C00, 0x11C08}, {0x11C0A, 0x11C2E}, {0x11C40, 0x11C40}, {0x11C72, 0x11C8F},
+{0x11D00, 0x11D06}, {0x11D08, 0x11D09}, {0x11D0B, 0x11D30}, {0x11D46, 0x11D46}, {0x11D60, 0x11D65}, {0x11D67, 0x11D68}, {0x11D6A, 0x11D89}, {0x11D98, 0x11D98}, {0x11EE0, 0x11EF2}, {0x11FB0, 0x11FB0},
+{0x12000, 0x12399}, {0x12480, 0x12543}, {0x13000, 0x1342E}, {0x14400, 0x14646}, {0x16800, 0x16A38}, {0x16A40, 0x16A5E}, {0x16AD0, 0x16AED}, {0x16B00, 0x16B2F}, {0x16B40, 0x16B43}, {0x16B63, 0x16B77},
+{0x16B7D, 0x16B8F}, {0x16E40, 0x16E7F}, {0x16F00, 0x16F4A}, {0x16F50, 0x16F50}, {0x16F93, 0x16F9F}, {0x16FE0, 0x16FE1}, {0x16FE3, 0x16FE3}, {0x17000, 0x187F7}, {0x18800, 0x18CD5}, {0x18D00, 0x18D08},
+{0x1B000, 0x1B11E}, {0x1B150, 0x1B152}, {0x1B164, 0x1B167}, {0x1B170, 0x1B2FB}, {0x1BC00, 0x1BC6A}, {0x1BC70, 0x1BC7C}, {0x1BC80, 0x1BC88}, {0x1BC90, 0x1BC99}, {0x1D400, 0x1D454}, {0x1D456, 0x1D49C},
+{0x1D49E, 0x1D49F}, {0x1D4A2, 0x1D4A2}, {0x1D4A5, 0x1D4A6}, {0x1D4A9, 0x1D4AC}, {0x1D4AE, 0x1D4B9}, {0x1D4BB, 0x1D4BB}, {0x1D4BD, 0x1D4C3}, {0x1D4C5, 0x1D505}, {0x1D507, 0x1D50A}, {0x1D50D, 0x1D514},
+{0x1D516, 0x1D51C}, {0x1D51E, 0x1D539}, {0x1D53B, 0x1D53E}, {0x1D540, 0x1D544}, {0x1D546, 0x1D546}, {0x1D54A, 0x1D550}, {0x1D552, 0x1D6A5}, {0x1D6A8, 0x1D6C0}, {0x1D6C2, 0x1D6DA}, {0x1D6DC, 0x1D6FA},
+{0x1D6FC, 0x1D714}, {0x1D716, 0x1D734}, {0x1D736, 0x1D74E}, {0x1D750, 0x1D76E}, {0x1D770, 0x1D788}, {0x1D78A, 0x1D7A8}, {0x1D7AA, 0x1D7C2}, {0x1D7C4, 0x1D7CB}, {0x1E100, 0x1E12C}, {0x1E137, 0x1E13D},
+{0x1E14E, 0x1E14E}, {0x1E2C0, 0x1E2EB}, {0x1E800, 0x1E8C4}, {0x1E900, 0x1E943}, {0x1E94B, 0x1E94B}, {0x1EE00, 0x1EE03}, {0x1EE05, 0x1EE1F}, {0x1EE21, 0x1EE22}, {0x1EE24, 0x1EE24}, {0x1EE27, 0x1EE27},
+{0x1EE29, 0x1EE32}, {0x1EE34, 0x1EE37}, {0x1EE39, 0x1EE39}, {0x1EE3B, 0x1EE3B}, {0x1EE42, 0x1EE42}, {0x1EE47, 0x1EE47}, {0x1EE49, 0x1EE49}, {0x1EE4B, 0x1EE4B}, {0x1EE4D, 0x1EE4F}, {0x1EE51, 0x1EE52},
+{0x1EE54, 0x1EE54}, {0x1EE57, 0x1EE57}, {0x1EE59, 0x1EE59}, {0x1EE5B, 0x1EE5B}, {0x1EE5D, 0x1EE5D}, {0x1EE5F, 0x1EE5F}, {0x1EE61, 0x1EE62}, {0x1EE64, 0x1EE64}, {0x1EE67, 0x1EE6A}, {0x1EE6C, 0x1EE72},
+{0x1EE74, 0x1EE77}, {0x1EE79, 0x1EE7C}, {0x1EE7E, 0x1EE7E}, {0x1EE80, 0x1EE89}, {0x1EE8B, 0x1EE9B}, {0x1EEA1, 0x1EEA3}, {0x1EEA5, 0x1EEA9}, {0x1EEAB, 0x1EEBB}, {0x20000, 0x2A6DD}, {0x2A700, 0x2B734},
+{0x2B740, 0x2B81D}, {0x2B820, 0x2CEA1}, {0x2CEB0, 0x2EBE0}, {0x2F800, 0x2FA1D}, {0x30000, 0x3134A},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> whitespace_ranges = {
+{0x9, 0xD}, {0x1C, 0x20}, {0x85, 0x85}, {0xA0, 0xA0}, {0x1680, 0x1680}, {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> accent_mark_ranges = {
+{0x300, 0x36F}, {0x483, 0x489}, {0x591, 0x5BD}, {0x5BF, 0x5BF}, {0x5C1, 0x5C2}, {0x5C4, 0x5C5}, {0x5C7, 0x5C7}, {0x610, 0x61A}, {0x64B, 0x65F}, {0x670, 0x670}, {0x6D6, 0x6DC}, {0x6DF, 0x6E4},
+{0x6E7, 0x6E8}, {0x6EA, 0x6ED}, {0x711, 0x711}, {0x730, 0x74A}, {0x7A6, 0x7B0}, {0x7EB, 0x7F3}, {0x7FD, 0x7FD}, {0x816, 0x819}, {0x81B, 0x823}, {0x825, 0x827}, {0x829, 0x82D}, {0x859, 0x85B},
+{0x8D3, 0x8E1}, {0x8E3, 0x903}, {0x93A, 0x93C}, {0x93E, 0x94F}, {0x951, 0x957}, {0x962, 0x963}, {0x981, 0x983}, {0x9BC, 0x9BC}, {0x9BE, 0x9C4}, {0x9C7, 0x9C8}, {0x9CB, 0x9CD}, {0x9D7, 0x9D7},
+{0x9E2, 0x9E3}, {0x9FE, 0x9FE}, {0xA01, 0xA03}, {0xA3C, 0xA3C}, {0xA3E, 0xA42}, {0xA47, 0xA48}, {0xA4B, 0xA4D}, {0xA51, 0xA51}, {0xA70, 0xA71}, {0xA75, 0xA75}, {0xA81, 0xA83}, {0xABC, 0xABC},
+{0xABE, 0xAC5}, {0xAC7, 0xAC9}, {0xACB, 0xACD}, {0xAE2, 0xAE3}, {0xAFA, 0xAFF}, {0xB01, 0xB03}, {0xB3C, 0xB3C}, {0xB3E, 0xB44}, {0xB47, 0xB48}, {0xB4B, 0xB4D}, {0xB55, 0xB57}, {0xB62, 0xB63},
+{0xB82, 0xB82}, {0xBBE, 0xBC2}, {0xBC6, 0xBC8}, {0xBCA, 0xBCD}, {0xBD7, 0xBD7}, {0xC00, 0xC04}, {0xC3E, 0xC44}, {0xC46, 0xC48}, {0xC4A, 0xC4D}, {0xC55, 0xC56}, {0xC62, 0xC63}, {0xC81, 0xC83},
+{0xCBC, 0xCBC}, {0xCBE, 0xCC4}, {0xCC6, 0xCC8}, {0xCCA, 0xCCD}, {0xCD5, 0xCD6}, {0xCE2, 0xCE3}, {0xD00, 0xD03}, {0xD3B, 0xD3C}, {0xD3E, 0xD44}, {0xD46, 0xD48}, {0xD4A, 0xD4D}, {0xD57, 0xD57},
+{0xD62, 0xD63}, {0xD81, 0xD83}, {0xDCA, 0xDCA}, {0xDCF, 0xDD4}, {0xDD6, 0xDD6}, {0xDD8, 0xDDF}, {0xDF2, 0xDF3}, {0xE31, 0xE31}, {0xE34, 0xE3A}, {0xE47, 0xE4E}, {0xEB1, 0xEB1}, {0xEB4, 0xEBC},
+{0xEC8, 0xECD}, {0xF18, 0xF19}, {0xF35, 0xF35}, {0xF37, 0xF37}, {0xF39, 0xF39}, {0xF3E, 0xF3F}, {0xF71, 0xF84}, {0xF86, 0xF87}, {0xF8D, 0xF97}, {0xF99, 0xFBC}, {0xFC6, 0xFC6}, {0x102B, 0x103E},
+{0x1056, 0x1059}, {0x105E, 0x1060}, {0x1062, 0x1064}, {0x1067, 0x106D}, {0x1071, 0x1074}, {0x1082, 0x108D}, {0x108F, 0x108F}, {0x109A, 0x109D}, {0x135D, 0x135F}, {0x1712, 0x1714}, {0x1732, 0x1734},
+{0x1752, 0x1753}, {0x1772, 0x1773}, {0x17B4, 0x17D3}, {0x17DD, 0x17DD}, {0x180B, 0x180D}, {0x1885, 0x1886}, {0x18A9, 0x18A9}, {0x1920, 0x192B}, {0x1930, 0x193B}, {0x1A17, 0x1A1B}, {0x1A55, 0x1A5E},
+{0x1A60, 0x1A7C}, {0x1A7F, 0x1A7F}, {0x1AB0, 0x1AC0}, {0x1B00, 0x1B04}, {0x1B34, 0x1B44}, {0x1B6B, 0x1B73}, {0x1B80, 0x1B82}, {0x1BA1, 0x1BAD}, {0x1BE6, 0x1BF3}, {0x1C24, 0x1C37}, {0x1CD0, 0x1CD2},
+{0x1CD4, 0x1CE8}, {0x1CED, 0x1CED}, {0x1CF4, 0x1CF4}, {0x1CF7, 0x1CF9}, {0x1DC0, 0x1DF9}, {0x1DFB, 0x1DFF}, {0x20D0, 0x20F0}, {0x2CEF, 0x2CF1}, {0x2D7F, 0x2D7F}, {0x2DE0, 0x2DFF}, {0x302A, 0x302F},
+{0x3099, 0x309A}, {0xA66F, 0xA672}, {0xA674, 0xA67D}, {0xA69E, 0xA69F}, {0xA6F0, 0xA6F1}, {0xA802, 0xA802}, {0xA806, 0xA806}, {0xA80B, 0xA80B}, {0xA823, 0xA827}, {0xA82C, 0xA82C}, {0xA880, 0xA881},
+{0xA8B4, 0xA8C5}, {0xA8E0, 0xA8F1}, {0xA8FF, 0xA8FF}, {0xA926, 0xA92D}, {0xA947, 0xA953}, {0xA980, 0xA983}, {0xA9B3, 0xA9C0}, {0xA9E5, 0xA9E5}, {0xAA29, 0xAA36}, {0xAA43, 0xAA43}, {0xAA4C, 0xAA4D},
+{0xAA7B, 0xAA7D}, {0xAAB0, 0xAAB0}, {0xAAB2, 0xAAB4}, {0xAAB7, 0xAAB8}, {0xAABE, 0xAABF}, {0xAAC1, 0xAAC1}, {0xAAEB, 0xAAEF}, {0xAAF5, 0xAAF6}, {0xABE3, 0xABEA}, {0xABEC, 0xABED}, {0xFB1E, 0xFB1E},
+{0xFE00, 0xFE0F}, {0xFE20, 0xFE2F}, {0x101FD, 0x101FD}, {0x102E0, 0x102E0}, {0x10376, 0x1037A}, {0x10A01, 0x10A03}, {0x10A05, 0x10A06}, {0x10A0C, 0x10A0F}, {0x10A38, 0x10A3A}, {0x10A3F, 0x10A3F},
+{0x10AE5, 0x10AE6}, {0x10D24, 0x10D27}, {0x10EAB, 0x10EAC}, {0x10F46, 0x10F50}, {0x11000, 0x11002}, {0x11038, 0x11046}, {0x1107F, 0x11082}, {0x110B0, 0x110BA}, {0x11100, 0x11102}, {0x11127, 0x11134},
+{0x11145, 0x11146}, {0x11173, 0x11173}, {0x11180, 0x11182}, {0x111B3, 0x111C0}, {0x111C9, 0x111CC}, {0x111CE, 0x111CF}, {0x1122C, 0x11237}, {0x1123E, 0x1123E}, {0x112DF, 0x112EA}, {0x11300, 0x11303},
+{0x1133B, 0x1133C}, {0x1133E, 0x11344}, {0x11347, 0x11348}, {0x1134B, 0x1134D}, {0x11357, 0x11357}, {0x11362, 0x11363}, {0x11366, 0x1136C}, {0x11370, 0x11374}, {0x11435, 0x11446}, {0x1145E, 0x1145E},
+{0x114B0, 0x114C3}, {0x115AF, 0x115B5}, {0x115B8, 0x115C0}, {0x115DC, 0x115DD}, {0x11630, 0x11640}, {0x116AB, 0x116B7}, {0x1171D, 0x1172B}, {0x1182C, 0x1183A}, {0x11930, 0x11935}, {0x11937, 0x11938},
+{0x1193B, 0x1193E}, {0x11940, 0x11940}, {0x11942, 0x11943}, {0x119D1, 0x119D7}, {0x119DA, 0x119E0}, {0x119E4, 0x119E4}, {0x11A01, 0x11A0A}, {0x11A33, 0x11A39}, {0x11A3B, 0x11A3E}, {0x11A47, 0x11A47},
+{0x11A51, 0x11A5B}, {0x11A8A, 0x11A99}, {0x11C2F, 0x11C36}, {0x11C38, 0x11C3F}, {0x11C92, 0x11CA7}, {0x11CA9, 0x11CB6}, {0x11D31, 0x11D36}, {0x11D3A, 0x11D3A}, {0x11D3C, 0x11D3D}, {0x11D3F, 0x11D45},
+{0x11D47, 0x11D47}, {0x11D8A, 0x11D8E}, {0x11D90, 0x11D91}, {0x11D93, 0x11D97}, {0x11EF3, 0x11EF6}, {0x16AF0, 0x16AF4}, {0x16B30, 0x16B36}, {0x16F4F, 0x16F4F}, {0x16F51, 0x16F87}, {0x16F8F, 0x16F92},
+{0x16FE4, 0x16FE4}, {0x16FF0, 0x16FF1}, {0x1BC9D, 0x1BC9E}, {0x1D165, 0x1D169}, {0x1D16D, 0x1D172}, {0x1D17B, 0x1D182}, {0x1D185, 0x1D18B}, {0x1D1AA, 0x1D1AD}, {0x1D242, 0x1D244}, {0x1DA00, 0x1DA36},
+{0x1DA3B, 0x1DA6C}, {0x1DA75, 0x1DA75}, {0x1DA84, 0x1DA84}, {0x1DA9B, 0x1DA9F}, {0x1DAA1, 0x1DAAF}, {0x1E000, 0x1E006}, {0x1E008, 0x1E018}, {0x1E01B, 0x1E021}, {0x1E023, 0x1E024}, {0x1E026, 0x1E02A},
+{0x1E130, 0x1E136}, {0x1E2EC, 0x1E2EF}, {0x1E8D0, 0x1E8D6}, {0x1E944, 0x1E94A}, {0xE0100, 0xE01EF},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> punctuation_ranges = {
+{0x21, 0x23}, {0x25, 0x2A}, {0x2C, 0x2F}, {0x3A, 0x3B}, {0x3F, 0x40}, {0x5B, 0x5D}, {0x5F, 0x5F}, {0x7B, 0x7B}, {0x7D, 0x7D}, {0xA1, 0xA1}, {0xA7, 0xA7}, {0xAB, 0xAB}, {0xB6, 0xB7}, {0xBB, 0xBB},
+{0xBF, 0xBF}, {0x37E, 0x37E}, {0x387, 0x387}, {0x55A, 0x55F}, {0x589, 0x58A}, {0x5BE, 0x5BE}, {0x5C0, 0x5C0}, {0x5C3, 0x5C3}, {0x5C6, 0x5C6}, {0x5F3, 0x5F4}, {0x609, 0x60A}, {0x60C, 0x60D},
+{0x61B, 0x61B}, {0x61E, 0x61F}, {0x66A, 0x66D}, {0x6D4, 0x6D4}, {0x700, 0x70D}, {0x7F7, 0x7F9}, {0x830, 0x83E}, {0x85E, 0x85E}, {0x964, 0x965}, {0x970, 0x970}, {0x9FD, 0x9FD}, {0xA76, 0xA76},
+{0xAF0, 0xAF0}, {0xC77, 0xC77}, {0xC84, 0xC84}, {0xDF4, 0xDF4}, {0xE4F, 0xE4F}, {0xE5A, 0xE5B}, {0xF04, 0xF12}, {0xF14, 0xF14}, {0xF3A, 0xF3D}, {0xF85, 0xF85}, {0xFD0, 0xFD4}, {0xFD9, 0xFDA},
+{0x104A, 0x104F}, {0x10FB, 0x10FB}, {0x1360, 0x1368}, {0x1400, 0x1400}, {0x166E, 0x166E}, {0x169B, 0x169C}, {0x16EB, 0x16ED}, {0x1735, 0x1736}, {0x17D4, 0x17D6}, {0x17D8, 0x17DA}, {0x1800, 0x180A},
+{0x1944, 0x1945}, {0x1A1E, 0x1A1F}, {0x1AA0, 0x1AA6}, {0x1AA8, 0x1AAD}, {0x1B5A, 0x1B60}, {0x1BFC, 0x1BFF}, {0x1C3B, 0x1C3F}, {0x1C7E, 0x1C7F}, {0x1CC0, 0x1CC7}, {0x1CD3, 0x1CD3}, {0x2010, 0x2027},
+{0x2030, 0x2043}, {0x2045, 0x2051}, {0x2053, 0x205E}, {0x207D, 0x207E}, {0x208D, 0x208E}, {0x2308, 0x230B}, {0x2329, 0x232A}, {0x2768, 0x2775}, {0x27C5, 0x27C6}, {0x27E6, 0x27EF}, {0x2983, 0x2998},
+{0x29D8, 0x29DB}, {0x29FC, 0x29FD}, {0x2CF9, 0x2CFC}, {0x2CFE, 0x2CFF}, {0x2D70, 0x2D70}, {0x2E00, 0x2E2E}, {0x2E30, 0x2E4F}, {0x2E52, 0x2E52}, {0x3001, 0x3003}, {0x3008, 0x3011}, {0x3014, 0x301F},
+{0x3030, 0x3030}, {0x303D, 0x303D}, {0x30A0, 0x30A0}, {0x30FB, 0x30FB}, {0xA4FE, 0xA4FF}, {0xA60D, 0xA60F}, {0xA673, 0xA673}, {0xA67E, 0xA67E}, {0xA6F2, 0xA6F7}, {0xA874, 0xA877}, {0xA8CE, 0xA8CF},
+{0xA8F8, 0xA8FA}, {0xA8FC, 0xA8FC}, {0xA92E, 0xA92F}, {0xA95F, 0xA95F}, {0xA9C1, 0xA9CD}, {0xA9DE, 0xA9DF}, {0xAA5C, 0xAA5F}, {0xAADE, 0xAADF}, {0xAAF0, 0xAAF1}, {0xABEB, 0xABEB}, {0xFD3E, 0xFD3F},
+{0xFE10, 0xFE19}, {0xFE30, 0xFE52}, {0xFE54, 0xFE61}, {0xFE63, 0xFE63}, {0xFE68, 0xFE68}, {0xFE6A, 0xFE6B}, {0xFF01, 0xFF03}, {0xFF05, 0xFF0A}, {0xFF0C, 0xFF0F}, {0xFF1A, 0xFF1B}, {0xFF1F, 0xFF20},
+{0xFF3B, 0xFF3D}, {0xFF3F, 0xFF3F}, {0xFF5B, 0xFF5B}, {0xFF5D, 0xFF5D}, {0xFF5F, 0xFF65}, {0x10100, 0x10102}, {0x1039F, 0x1039F}, {0x103D0, 0x103D0}, {0x1056F, 0x1056F}, {0x10857, 0x10857},
+{0x1091F, 0x1091F}, {0x1093F, 0x1093F}, {0x10A50, 0x10A58}, {0x10A7F, 0x10A7F}, {0x10AF0, 0x10AF6}, {0x10B39, 0x10B3F}, {0x10B99, 0x10B9C}, {0x10EAD, 0x10EAD}, {0x10F55, 0x10F59}, {0x11047, 0x1104D},
+{0x110BB, 0x110BC}, {0x110BE, 0x110C1}, {0x11140, 0x11143}, {0x11174, 0x11175}, {0x111C5, 0x111C8}, {0x111CD, 0x111CD}, {0x111DB, 0x111DB}, {0x111DD, 0x111DF}, {0x11238, 0x1123D}, {0x112A9, 0x112A9},
+{0x1144B, 0x1144F}, {0x1145A, 0x1145B}, {0x1145D, 0x1145D}, {0x114C6, 0x114C6}, {0x115C1, 0x115D7}, {0x11641, 0x11643}, {0x11660, 0x1166C}, {0x1173C, 0x1173E}, {0x1183B, 0x1183B}, {0x11944, 0x11946},
+{0x119E2, 0x119E2}, {0x11A3F, 0x11A46}, {0x11A9A, 0x11A9C}, {0x11A9E, 0x11AA2}, {0x11C41, 0x11C45}, {0x11C70, 0x11C71}, {0x11EF7, 0x11EF8}, {0x11FFF, 0x11FFF}, {0x12470, 0x12474}, {0x16A6E, 0x16A6F},
+{0x16AF5, 0x16AF5}, {0x16B37, 0x16B3B}, {0x16B44, 0x16B44}, {0x16E97, 0x16E9A}, {0x16FE2, 0x16FE2}, {0x1BC9F, 0x1BC9F}, {0x1DA87, 0x1DA8B}, {0x1E95E, 0x1E95F},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> symbol_ranges = {
+{0x24, 0x24}, {0x2B, 0x2B}, {0x3C, 0x3E}, {0x5E, 0x5E}, {0x60, 0x60}, {0x7C, 0x7C}, {0x7E, 0x7E}, {0xA2, 0xA6}, {0xA8, 0xA9}, {0xAC, 0xAC}, {0xAE, 0xB1}, {0xB4, 0xB4}, {0xB8, 0xB8}, {0xD7, 0xD7},
+{0xF7, 0xF7}, {0x2C2, 0x2C5}, {0x2D2, 0x2DF}, {0x2E5, 0x2EB}, {0x2ED, 0x2ED}, {0x2EF, 0x2FF}, {0x375, 0x375}, {0x384, 0x385}, {0x3F6, 0x3F6}, {0x482, 0x482}, {0x58D, 0x58F}, {0x606, 0x608},
+{0x60B, 0x60B}, {0x60E, 0x60F}, {0x6DE, 0x6DE}, {0x6E9, 0x6E9}, {0x6FD, 0x6FE}, {0x7F6, 0x7F6}, {0x7FE, 0x7FF}, {0x9F2, 0x9F3}, {0x9FA, 0x9FB}, {0xAF1, 0xAF1}, {0xB70, 0xB70}, {0xBF3, 0xBFA},
+{0xC7F, 0xC7F}, {0xD4F, 0xD4F}, {0xD79, 0xD79}, {0xE3F, 0xE3F}, {0xF01, 0xF03}, {0xF13, 0xF13}, {0xF15, 0xF17}, {0xF1A, 0xF1F}, {0xF34, 0xF34}, {0xF36, 0xF36}, {0xF38, 0xF38}, {0xFBE, 0xFC5},
+{0xFC7, 0xFCC}, {0xFCE, 0xFCF}, {0xFD5, 0xFD8}, {0x109E, 0x109F}, {0x1390, 0x1399}, {0x166D, 0x166D}, {0x17DB, 0x17DB}, {0x1940, 0x1940}, {0x19DE, 0x19FF}, {0x1B61, 0x1B6A}, {0x1B74, 0x1B7C},
+{0x1FBD, 0x1FBD}, {0x1FBF, 0x1FC1}, {0x1FCD, 0x1FCF}, {0x1FDD, 0x1FDF}, {0x1FED, 0x1FEF}, {0x1FFD, 0x1FFE}, {0x2044, 0x2044}, {0x2052, 0x2052}, {0x207A, 0x207C}, {0x208A, 0x208C}, {0x20A0, 0x20BF},
+{0x2100, 0x2101}, {0x2103, 0x2106}, {0x2108, 0x2109}, {0x2114, 0x2114}, {0x2116, 0x2118}, {0x211E, 0x2123}, {0x2125, 0x2125}, {0x2127, 0x2127}, {0x2129, 0x2129}, {0x212E, 0x212E}, {0x213A, 0x213B},
+{0x2140, 0x2144}, {0x214A, 0x214D}, {0x214F, 0x214F}, {0x218A, 0x218B}, {0x2190, 0x2307}, {0x230C, 0x2328}, {0x232B, 0x2426}, {0x2440, 0x244A}, {0x249C, 0x24E9}, {0x2500, 0x2767}, {0x2794, 0x27C4},
+{0x27C7, 0x27E5}, {0x27F0, 0x2982}, {0x2999, 0x29D7}, {0x29DC, 0x29FB}, {0x29FE, 0x2B73}, {0x2B76, 0x2B95}, {0x2B97, 0x2BFF}, {0x2CE5, 0x2CEA}, {0x2E50, 0x2E51}, {0x2E80, 0x2E99}, {0x2E9B, 0x2EF3},
+{0x2F00, 0x2FD5}, {0x2FF0, 0x2FFB}, {0x3004, 0x3004}, {0x3012, 0x3013}, {0x3020, 0x3020}, {0x3036, 0x3037}, {0x303E, 0x303F}, {0x309B, 0x309C}, {0x3190, 0x3191}, {0x3196, 0x319F}, {0x31C0, 0x31E3},
+{0x3200, 0x321E}, {0x322A, 0x3247}, {0x3250, 0x3250}, {0x3260, 0x327F}, {0x328A, 0x32B0}, {0x32C0, 0x33FF}, {0x4DC0, 0x4DFF}, {0xA490, 0xA4C6}, {0xA700, 0xA716}, {0xA720, 0xA721}, {0xA789, 0xA78A},
+{0xA828, 0xA82B}, {0xA836, 0xA839}, {0xAA77, 0xAA79}, {0xAB5B, 0xAB5B}, {0xAB6A, 0xAB6B}, {0xFB29, 0xFB29}, {0xFBB2, 0xFBC1}, {0xFDFC, 0xFDFD}, {0xFE62, 0xFE62}, {0xFE64, 0xFE66}, {0xFE69, 0xFE69},
+{0xFF04, 0xFF04}, {0xFF0B, 0xFF0B}, {0xFF1C, 0xFF1E}, {0xFF3E, 0xFF3E}, {0xFF40, 0xFF40}, {0xFF5C, 0xFF5C}, {0xFF5E, 0xFF5E}, {0xFFE0, 0xFFE6}, {0xFFE8, 0xFFEE}, {0xFFFC, 0xFFFD}, {0x10137, 0x1013F},
+{0x10179, 0x10189}, {0x1018C, 0x1018E}, {0x10190, 0x1019C}, {0x101A0, 0x101A0}, {0x101D0, 0x101FC}, {0x10877, 0x10878}, {0x10AC8, 0x10AC8}, {0x1173F, 0x1173F}, {0x11FD5, 0x11FF1}, {0x16B3C, 0x16B3F},
+{0x16B45, 0x16B45}, {0x1BC9C, 0x1BC9C}, {0x1D000, 0x1D0F5}, {0x1D100, 0x1D126}, {0x1D129, 0x1D164}, {0x1D16A, 0x1D16C}, {0x1D183, 0x1D184}, {0x1D18C, 0x1D1A9}, {0x1D1AE, 0x1D1E8}, {0x1D200, 0x1D241},
+{0x1D245, 0x1D245}, {0x1D300, 0x1D356}, {0x1D6C1, 0x1D6C1}, {0x1D6DB, 0x1D6DB}, {0x1D6FB, 0x1D6FB}, {0x1D715, 0x1D715}, {0x1D735, 0x1D735}, {0x1D74F, 0x1D74F}, {0x1D76F, 0x1D76F}, {0x1D789, 0x1D789},
+{0x1D7A9, 0x1D7A9}, {0x1D7C3, 0x1D7C3}, {0x1D800, 0x1D9FF}, {0x1DA37, 0x1DA3A}, {0x1DA6D, 0x1DA74}, {0x1DA76, 0x1DA83}, {0x1DA85, 0x1DA86}, {0x1E14F, 0x1E14F}, {0x1E2FF, 0x1E2FF}, {0x1ECAC, 0x1ECAC},
+{0x1ECB0, 0x1ECB0}, {0x1ED2E, 0x1ED2E}, {0x1EEF0, 0x1EEF1}, {0x1F000, 0x1F02B}, {0x1F030, 0x1F093}, {0x1F0A0, 0x1F0AE}, {0x1F0B1, 0x1F0BF}, {0x1F0C1, 0x1F0CF}, {0x1F0D1, 0x1F0F5}, {0x1F10D, 0x1F1AD},
+{0x1F1E6, 0x1F202}, {0x1F210, 0x1F23B}, {0x1F240, 0x1F248}, {0x1F250, 0x1F251}, {0x1F260, 0x1F265}, {0x1F300, 0x1F6D7}, {0x1F6E0, 0x1F6EC}, {0x1F6F0, 0x1F6FC}, {0x1F700, 0x1F773}, {0x1F780, 0x1F7D8},
+{0x1F7E0, 0x1F7EB}, {0x1F800, 0x1F80B}, {0x1F810, 0x1F847}, {0x1F850, 0x1F859}, {0x1F860, 0x1F887}, {0x1F890, 0x1F8AD}, {0x1F8B0, 0x1F8B1}, {0x1F900, 0x1F978}, {0x1F97A, 0x1F9CB}, {0x1F9CD, 0x1FA53},
+{0x1FA60, 0x1FA6D}, {0x1FA70, 0x1FA74}, {0x1FA78, 0x1FA7A}, {0x1FA80, 0x1FA86}, {0x1FA90, 0x1FAA8}, {0x1FAB0, 0x1FAB6}, {0x1FAC0, 0x1FAC2}, {0x1FAD0, 0x1FAD6}, {0x1FB00, 0x1FB92}, {0x1FB94, 0x1FBCA},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
+{0x0, 0x8}, {0xE, 0x1B}, {0x7F, 0x84}, {0x86, 0x9F}, {0xAD, 0xAD}, {0x378, 0x379}, {0x380, 0x383}, {0x38B, 0x38B}, {0x38D, 0x38D}, {0x3A2, 0x3A2}, {0x530, 0x530}, {0x557, 0x558}, {0x58B, 0x58C},
+{0x590, 0x590}, {0x5C8, 0x5CF}, {0x5EB, 0x5EE}, {0x5F5, 0x605}, {0x61C, 0x61D}, {0x6DD, 0x6DD}, {0x70E, 0x70F}, {0x74B, 0x74C}, {0x7B2, 0x7BF}, {0x7FB, 0x7FC}, {0x82E, 0x82F}, {0x83F, 0x83F},
+{0x85C, 0x85D}, {0x85F, 0x85F}, {0x86B, 0x89F}, {0x8B5, 0x8B5}, {0x8C8, 0x8D2}, {0x8E2, 0x8E2}, {0x984, 0x984}, {0x98D, 0x98E}, {0x991, 0x992}, {0x9A9, 0x9A9}, {0x9B1, 0x9B1}, {0x9B3, 0x9B5},
+{0x9BA, 0x9BB}, {0x9C5, 0x9C6}, {0x9C9, 0x9CA}, {0x9CF, 0x9D6}, {0x9D8, 0x9DB}, {0x9DE, 0x9DE}, {0x9E4, 0x9E5}, {0x9FF, 0xA00}, {0xA04, 0xA04}, {0xA0B, 0xA0E}, {0xA11, 0xA12}, {0xA29, 0xA29},
+{0xA31, 0xA31}, {0xA34, 0xA34}, {0xA37, 0xA37}, {0xA3A, 0xA3B}, {0xA3D, 0xA3D}, {0xA43, 0xA46}, {0xA49, 0xA4A}, {0xA4E, 0xA50}, {0xA52, 0xA58}, {0xA5D, 0xA5D}, {0xA5F, 0xA65}, {0xA77, 0xA80},
+{0xA84, 0xA84}, {0xA8E, 0xA8E}, {0xA92, 0xA92}, {0xAA9, 0xAA9}, {0xAB1, 0xAB1}, {0xAB4, 0xAB4}, {0xABA, 0xABB}, {0xAC6, 0xAC6}, {0xACA, 0xACA}, {0xACE, 0xACF}, {0xAD1, 0xADF}, {0xAE4, 0xAE5},
+{0xAF2, 0xAF8}, {0xB00, 0xB00}, {0xB04, 0xB04}, {0xB0D, 0xB0E}, {0xB11, 0xB12}, {0xB29, 0xB29}, {0xB31, 0xB31}, {0xB34, 0xB34}, {0xB3A, 0xB3B}, {0xB45, 0xB46}, {0xB49, 0xB4A}, {0xB4E, 0xB54},
+{0xB58, 0xB5B}, {0xB5E, 0xB5E}, {0xB64, 0xB65}, {0xB78, 0xB81}, {0xB84, 0xB84}, {0xB8B, 0xB8D}, {0xB91, 0xB91}, {0xB96, 0xB98}, {0xB9B, 0xB9B}, {0xB9D, 0xB9D}, {0xBA0, 0xBA2}, {0xBA5, 0xBA7},
+{0xBAB, 0xBAD}, {0xBBA, 0xBBD}, {0xBC3, 0xBC5}, {0xBC9, 0xBC9}, {0xBCE, 0xBCF}, {0xBD1, 0xBD6}, {0xBD8, 0xBE5}, {0xBFB, 0xBFF}, {0xC0D, 0xC0D}, {0xC11, 0xC11}, {0xC29, 0xC29}, {0xC3A, 0xC3C},
+{0xC45, 0xC45}, {0xC49, 0xC49}, {0xC4E, 0xC54}, {0xC57, 0xC57}, {0xC5B, 0xC5F}, {0xC64, 0xC65}, {0xC70, 0xC76}, {0xC8D, 0xC8D}, {0xC91, 0xC91}, {0xCA9, 0xCA9}, {0xCB4, 0xCB4}, {0xCBA, 0xCBB},
+{0xCC5, 0xCC5}, {0xCC9, 0xCC9}, {0xCCE, 0xCD4}, {0xCD7, 0xCDD}, {0xCDF, 0xCDF}, {0xCE4, 0xCE5}, {0xCF0, 0xCF0}, {0xCF3, 0xCFF}, {0xD0D, 0xD0D}, {0xD11, 0xD11}, {0xD45, 0xD45}, {0xD49, 0xD49},
+{0xD50, 0xD53}, {0xD64, 0xD65}, {0xD80, 0xD80}, {0xD84, 0xD84}, {0xD97, 0xD99}, {0xDB2, 0xDB2}, {0xDBC, 0xDBC}, {0xDBE, 0xDBF}, {0xDC7, 0xDC9}, {0xDCB, 0xDCE}, {0xDD5, 0xDD5}, {0xDD7, 0xDD7},
+{0xDE0, 0xDE5}, {0xDF0, 0xDF1}, {0xDF5, 0xE00}, {0xE3B, 0xE3E}, {0xE5C, 0xE80}, {0xE83, 0xE83}, {0xE85, 0xE85}, {0xE8B, 0xE8B}, {0xEA4, 0xEA4}, {0xEA6, 0xEA6}, {0xEBE, 0xEBF}, {0xEC5, 0xEC5},
+{0xEC7, 0xEC7}, {0xECE, 0xECF}, {0xEDA, 0xEDB}, {0xEE0, 0xEFF}, {0xF48, 0xF48}, {0xF6D, 0xF70}, {0xF98, 0xF98}, {0xFBD, 0xFBD}, {0xFCD, 0xFCD}, {0xFDB, 0xFFF}, {0x10C6, 0x10C6}, {0x10C8, 0x10CC},
+{0x10CE, 0x10CF}, {0x1249, 0x1249}, {0x124E, 0x124F}, {0x1257, 0x1257}, {0x1259, 0x1259}, {0x125E, 0x125F}, {0x1289, 0x1289}, {0x128E, 0x128F}, {0x12B1, 0x12B1}, {0x12B6, 0x12B7}, {0x12BF, 0x12BF},
+{0x12C1, 0x12C1}, {0x12C6, 0x12C7}, {0x12D7, 0x12D7}, {0x1311, 0x1311}, {0x1316, 0x1317}, {0x135B, 0x135C}, {0x137D, 0x137F}, {0x139A, 0x139F}, {0x13F6, 0x13F7}, {0x13FE, 0x13FF}, {0x169D, 0x169F},
+{0x16F9, 0x16FF}, {0x170D, 0x170D}, {0x1715, 0x171F}, {0x1737, 0x173F}, {0x1754, 0x175F}, {0x176D, 0x176D}, {0x1771, 0x1771}, {0x1774, 0x177F}, {0x17DE, 0x17DF}, {0x17EA, 0x17EF}, {0x17FA, 0x17FF},
+{0x180E, 0x180F}, {0x181A, 0x181F}, {0x1879, 0x187F}, {0x18AB, 0x18AF}, {0x18F6, 0x18FF}, {0x191F, 0x191F}, {0x192C, 0x192F}, {0x193C, 0x193F}, {0x1941, 0x1943}, {0x196E, 0x196F}, {0x1975, 0x197F},
+{0x19AC, 0x19AF}, {0x19CA, 0x19CF}, {0x19DB, 0x19DD}, {0x1A1C, 0x1A1D}, {0x1A5F, 0x1A5F}, {0x1A7D, 0x1A7E}, {0x1A8A, 0x1A8F}, {0x1A9A, 0x1A9F}, {0x1AAE, 0x1AAF}, {0x1AC1, 0x1AFF}, {0x1B4C, 0x1B4F},
+{0x1B7D, 0x1B7F}, {0x1BF4, 0x1BFB}, {0x1C38, 0x1C3A}, {0x1C4A, 0x1C4C}, {0x1C89, 0x1C8F}, {0x1CBB, 0x1CBC}, {0x1CC8, 0x1CCF}, {0x1CFB, 0x1CFF}, {0x1DFA, 0x1DFA}, {0x1F16, 0x1F17}, {0x1F1E, 0x1F1F},
+{0x1F46, 0x1F47}, {0x1F4E, 0x1F4F}, {0x1F58, 0x1F58}, {0x1F5A, 0x1F5A}, {0x1F5C, 0x1F5C}, {0x1F5E, 0x1F5E}, {0x1F7E, 0x1F7F}, {0x1FB5, 0x1FB5}, {0x1FC5, 0x1FC5}, {0x1FD4, 0x1FD5}, {0x1FDC, 0x1FDC},
+{0x1FF0, 0x1FF1}, {0x1FF5, 0x1FF5}, {0x1FFF, 0x1FFF}, {0x200B, 0x200F}, {0x202A, 0x202E}, {0x2060, 0x206F}, {0x2072, 0x2073}, {0x208F, 0x208F}, {0x209D, 0x209F}, {0x20C0, 0x20CF}, {0x20F1, 0x20FF},
+{0x218C, 0x218F}, {0x2427, 0x243F}, {0x244B, 0x245F}, {0x2B74, 0x2B75}, {0x2B96, 0x2B96}, {0x2C2F, 0x2C2F}, {0x2C5F, 0x2C5F}, {0x2CF4, 0x2CF8}, {0x2D26, 0x2D26}, {0x2D28, 0x2D2C}, {0x2D2E, 0x2D2F},
+{0x2D68, 0x2D6E}, {0x2D71, 0x2D7E}, {0x2D97, 0x2D9F}, {0x2DA7, 0x2DA7}, {0x2DAF, 0x2DAF}, {0x2DB7, 0x2DB7}, {0x2DBF, 0x2DBF}, {0x2DC7, 0x2DC7}, {0x2DCF, 0x2DCF}, {0x2DD7, 0x2DD7}, {0x2DDF, 0x2DDF},
+{0x2E53, 0x2E7F}, {0x2E9A, 0x2E9A}, {0x2EF4, 0x2EFF}, {0x2FD6, 0x2FEF}, {0x2FFC, 0x2FFF}, {0x3040, 0x3040}, {0x3097, 0x3098}, {0x3100, 0x3104}, {0x3130, 0x3130}, {0x318F, 0x318F}, {0x31E4, 0x31EF},
+{0x321F, 0x321F}, {0x9FFD, 0x9FFF}, {0xA48D, 0xA48F}, {0xA4C7, 0xA4CF}, {0xA62C, 0xA63F}, {0xA6F8, 0xA6FF}, {0xA7C0, 0xA7C1}, {0xA7CB, 0xA7F4}, {0xA82D, 0xA82F}, {0xA83A, 0xA83F}, {0xA878, 0xA87F},
+{0xA8C6, 0xA8CD}, {0xA8DA, 0xA8DF}, {0xA954, 0xA95E}, {0xA97D, 0xA97F}, {0xA9CE, 0xA9CE}, {0xA9DA, 0xA9DD}, {0xA9FF, 0xA9FF}, {0xAA37, 0xAA3F}, {0xAA4E, 0xAA4F}, {0xAA5A, 0xAA5B}, {0xAAC3, 0xAADA},
+{0xAAF7, 0xAB00}, {0xAB07, 0xAB08}, {0xAB0F, 0xAB10}, {0xAB17, 0xAB1F}, {0xAB27, 0xAB27}, {0xAB2F, 0xAB2F}, {0xAB6C, 0xAB6F}, {0xABEE, 0xABEF}, {0xABFA, 0xABFF}, {0xD7A4, 0xD7AF}, {0xD7C7, 0xD7CA},
+{0xD7FC, 0xF8FF}, {0xFA6E, 0xFA6F}, {0xFADA, 0xFAFF}, {0xFB07, 0xFB12}, {0xFB18, 0xFB1C}, {0xFB37, 0xFB37}, {0xFB3D, 0xFB3D}, {0xFB3F, 0xFB3F}, {0xFB42, 0xFB42}, {0xFB45, 0xFB45}, {0xFBC2, 0xFBD2},
+{0xFD40, 0xFD4F}, {0xFD90, 0xFD91}, {0xFDC8, 0xFDEF}, {0xFDFE, 0xFDFF}, {0xFE1A, 0xFE1F}, {0xFE53, 0xFE53}, {0xFE67, 0xFE67}, {0xFE6C, 0xFE6F}, {0xFE75, 0xFE75}, {0xFEFD, 0xFF00}, {0xFFBF, 0xFFC1},
+{0xFFC8, 0xFFC9}, {0xFFD0, 0xFFD1}, {0xFFD8, 0xFFD9}, {0xFFDD, 0xFFDF}, {0xFFE7, 0xFFE7}, {0xFFEF, 0xFFFB}, {0xFFFE, 0xFFFF}, {0x1000C, 0x1000C}, {0x10027, 0x10027}, {0x1003B, 0x1003B},
+{0x1003E, 0x1003E}, {0x1004E, 0x1004F}, {0x1005E, 0x1007F}, {0x100FB, 0x100FF}, {0x10103, 0x10106}, {0x10134, 0x10136}, {0x1018F, 0x1018F}, {0x1019D, 0x1019F}, {0x101A1, 0x101CF}, {0x101FE, 0x1027F},
+{0x1029D, 0x1029F}, {0x102D1, 0x102DF}, {0x102FC, 0x102FF}, {0x10324, 0x1032C}, {0x1034B, 0x1034F}, {0x1037B, 0x1037F}, {0x1039E, 0x1039E}, {0x103C4, 0x103C7}, {0x103D6, 0x103FF}, {0x1049E, 0x1049F},
+{0x104AA, 0x104AF}, {0x104D4, 0x104D7}, {0x104FC, 0x104FF}, {0x10528, 0x1052F}, {0x10564, 0x1056E}, {0x10570, 0x105FF}, {0x10737, 0x1073F}, {0x10756, 0x1075F}, {0x10768, 0x107FF}, {0x10806, 0x10807},
+{0x10809, 0x10809}, {0x10836, 0x10836}, {0x10839, 0x1083B}, {0x1083D, 0x1083E}, {0x10856, 0x10856}, {0x1089F, 0x108A6}, {0x108B0, 0x108DF}, {0x108F3, 0x108F3}, {0x108F6, 0x108FA}, {0x1091C, 0x1091E},
+{0x1093A, 0x1093E}, {0x10940, 0x1097F}, {0x109B8, 0x109BB}, {0x109D0, 0x109D1}, {0x10A04, 0x10A04}, {0x10A07, 0x10A0B}, {0x10A14, 0x10A14}, {0x10A18, 0x10A18}, {0x10A36, 0x10A37}, {0x10A3B, 0x10A3E},
+{0x10A49, 0x10A4F}, {0x10A59, 0x10A5F}, {0x10AA0, 0x10ABF}, {0x10AE7, 0x10AEA}, {0x10AF7, 0x10AFF}, {0x10B36, 0x10B38}, {0x10B56, 0x10B57}, {0x10B73, 0x10B77}, {0x10B92, 0x10B98}, {0x10B9D, 0x10BA8},
+{0x10BB0, 0x10BFF}, {0x10C49, 0x10C7F}, {0x10CB3, 0x10CBF}, {0x10CF3, 0x10CF9}, {0x10D28, 0x10D2F}, {0x10D3A, 0x10E5F}, {0x10E7F, 0x10E7F}, {0x10EAA, 0x10EAA}, {0x10EAE, 0x10EAF}, {0x10EB2, 0x10EFF},
+{0x10F28, 0x10F2F}, {0x10F5A, 0x10FAF}, {0x10FCC, 0x10FDF}, {0x10FF7, 0x10FFF}, {0x1104E, 0x11051}, {0x11070, 0x1107E}, {0x110BD, 0x110BD}, {0x110C2, 0x110CF}, {0x110E9, 0x110EF}, {0x110FA, 0x110FF},
+{0x11135, 0x11135}, {0x11148, 0x1114F}, {0x11177, 0x1117F}, {0x111E0, 0x111E0}, {0x111F5, 0x111FF}, {0x11212, 0x11212}, {0x1123F, 0x1127F}, {0x11287, 0x11287}, {0x11289, 0x11289}, {0x1128E, 0x1128E},
+{0x1129E, 0x1129E}, {0x112AA, 0x112AF}, {0x112EB, 0x112EF}, {0x112FA, 0x112FF}, {0x11304, 0x11304}, {0x1130D, 0x1130E}, {0x11311, 0x11312}, {0x11329, 0x11329}, {0x11331, 0x11331}, {0x11334, 0x11334},
+{0x1133A, 0x1133A}, {0x11345, 0x11346}, {0x11349, 0x1134A}, {0x1134E, 0x1134F}, {0x11351, 0x11356}, {0x11358, 0x1135C}, {0x11364, 0x11365}, {0x1136D, 0x1136F}, {0x11375, 0x113FF}, {0x1145C, 0x1145C},
+{0x11462, 0x1147F}, {0x114C8, 0x114CF}, {0x114DA, 0x1157F}, {0x115B6, 0x115B7}, {0x115DE, 0x115FF}, {0x11645, 0x1164F}, {0x1165A, 0x1165F}, {0x1166D, 0x1167F}, {0x116B9, 0x116BF}, {0x116CA, 0x116FF},
+{0x1171B, 0x1171C}, {0x1172C, 0x1172F}, {0x11740, 0x117FF}, {0x1183C, 0x1189F}, {0x118F3, 0x118FE}, {0x11907, 0x11908}, {0x1190A, 0x1190B}, {0x11914, 0x11914}, {0x11917, 0x11917}, {0x11936, 0x11936},
+{0x11939, 0x1193A}, {0x11947, 0x1194F}, {0x1195A, 0x1199F}, {0x119A8, 0x119A9}, {0x119D8, 0x119D9}, {0x119E5, 0x119FF}, {0x11A48, 0x11A4F}, {0x11AA3, 0x11ABF}, {0x11AF9, 0x11BFF}, {0x11C09, 0x11C09},
+{0x11C37, 0x11C37}, {0x11C46, 0x11C4F}, {0x11C6D, 0x11C6F}, {0x11C90, 0x11C91}, {0x11CA8, 0x11CA8}, {0x11CB7, 0x11CFF}, {0x11D07, 0x11D07}, {0x11D0A, 0x11D0A}, {0x11D37, 0x11D39}, {0x11D3B, 0x11D3B},
+{0x11D3E, 0x11D3E}, {0x11D48, 0x11D4F}, {0x11D5A, 0x11D5F}, {0x11D66, 0x11D66}, {0x11D69, 0x11D69}, {0x11D8F, 0x11D8F}, {0x11D92, 0x11D92}, {0x11D99, 0x11D9F}, {0x11DAA, 0x11EDF}, {0x11EF9, 0x11FAF},
+{0x11FB1, 0x11FBF}, {0x11FF2, 0x11FFE}, {0x1239A, 0x123FF}, {0x1246F, 0x1246F}, {0x12475, 0x1247F}, {0x12544, 0x12FFF}, {0x1342F, 0x143FF}, {0x14647, 0x167FF}, {0x16A39, 0x16A3F}, {0x16A5F, 0x16A5F},
+{0x16A6A, 0x16A6D}, {0x16A70, 0x16ACF}, {0x16AEE, 0x16AEF}, {0x16AF6, 0x16AFF}, {0x16B46, 0x16B4F}, {0x16B5A, 0x16B5A}, {0x16B62, 0x16B62}, {0x16B78, 0x16B7C}, {0x16B90, 0x16E3F}, {0x16E9B, 0x16EFF},
+{0x16F4B, 0x16F4E}, {0x16F88, 0x16F8E}, {0x16FA0, 0x16FDF}, {0x16FE5, 0x16FEF}, {0x16FF2, 0x16FFF}, {0x187F8, 0x187FF}, {0x18CD6, 0x18CFF}, {0x18D09, 0x1AFFF}, {0x1B11F, 0x1B14F}, {0x1B153, 0x1B163},
+{0x1B168, 0x1B16F}, {0x1B2FC, 0x1BBFF}, {0x1BC6B, 0x1BC6F}, {0x1BC7D, 0x1BC7F}, {0x1BC89, 0x1BC8F}, {0x1BC9A, 0x1BC9B}, {0x1BCA0, 0x1CFFF}, {0x1D0F6, 0x1D0FF}, {0x1D127, 0x1D128}, {0x1D173, 0x1D17A},
+{0x1D1E9, 0x1D1FF}, {0x1D246, 0x1D2DF}, {0x1D2F4, 0x1D2FF}, {0x1D357, 0x1D35F}, {0x1D379, 0x1D3FF}, {0x1D455, 0x1D455}, {0x1D49D, 0x1D49D}, {0x1D4A0, 0x1D4A1}, {0x1D4A3, 0x1D4A4}, {0x1D4A7, 0x1D4A8},
+{0x1D4AD, 0x1D4AD}, {0x1D4BA, 0x1D4BA}, {0x1D4BC, 0x1D4BC}, {0x1D4C4, 0x1D4C4}, {0x1D506, 0x1D506}, {0x1D50B, 0x1D50C}, {0x1D515, 0x1D515}, {0x1D51D, 0x1D51D}, {0x1D53A, 0x1D53A}, {0x1D53F, 0x1D53F},
+{0x1D545, 0x1D545}, {0x1D547, 0x1D549}, {0x1D551, 0x1D551}, {0x1D6A6, 0x1D6A7}, {0x1D7CC, 0x1D7CD}, {0x1DA8C, 0x1DA9A}, {0x1DAA0, 0x1DAA0}, {0x1DAB0, 0x1DFFF}, {0x1E007, 0x1E007}, {0x1E019, 0x1E01A},
+{0x1E022, 0x1E022}, {0x1E025, 0x1E025}, {0x1E02B, 0x1E0FF}, {0x1E12D, 0x1E12F}, {0x1E13E, 0x1E13F}, {0x1E14A, 0x1E14D}, {0x1E150, 0x1E2BF}, {0x1E2FA, 0x1E2FE}, {0x1E300, 0x1E7FF}, {0x1E8C5, 0x1E8C6},
+{0x1E8D7, 0x1E8FF}, {0x1E94C, 0x1E94F}, {0x1E95A, 0x1E95D}, {0x1E960, 0x1EC70}, {0x1ECB5, 0x1ED00}, {0x1ED3E, 0x1EDFF}, {0x1EE04, 0x1EE04}, {0x1EE20, 0x1EE20}, {0x1EE23, 0x1EE23}, {0x1EE25, 0x1EE26},
+{0x1EE28, 0x1EE28}, {0x1EE33, 0x1EE33}, {0x1EE38, 0x1EE38}, {0x1EE3A, 0x1EE3A}, {0x1EE3C, 0x1EE41}, {0x1EE43, 0x1EE46}, {0x1EE48, 0x1EE48}, {0x1EE4A, 0x1EE4A}, {0x1EE4C, 0x1EE4C}, {0x1EE50, 0x1EE50},
+{0x1EE53, 0x1EE53}, {0x1EE55, 0x1EE56}, {0x1EE58, 0x1EE58}, {0x1EE5A, 0x1EE5A}, {0x1EE5C, 0x1EE5C}, {0x1EE5E, 0x1EE5E}, {0x1EE60, 0x1EE60}, {0x1EE63, 0x1EE63}, {0x1EE65, 0x1EE66}, {0x1EE6B, 0x1EE6B},
+{0x1EE73, 0x1EE73}, {0x1EE78, 0x1EE78}, {0x1EE7D, 0x1EE7D}, {0x1EE7F, 0x1EE7F}, {0x1EE8A, 0x1EE8A}, {0x1EE9C, 0x1EEA0}, {0x1EEA4, 0x1EEA4}, {0x1EEAA, 0x1EEAA}, {0x1EEBC, 0x1EEEF}, {0x1EEF2, 0x1EFFF},
+{0x1F02C, 0x1F02F}, {0x1F094, 0x1F09F}, {0x1F0AF, 0x1F0B0}, {0x1F0C0, 0x1F0C0}, {0x1F0D0, 0x1F0D0}, {0x1F0F6, 0x1F0FF}, {0x1F1AE, 0x1F1E5}, {0x1F203, 0x1F20F}, {0x1F23C, 0x1F23F}, {0x1F249, 0x1F24F},
+{0x1F252, 0x1F25F}, {0x1F266, 0x1F2FF}, {0x1F6D8, 0x1F6DF}, {0x1F6ED, 0x1F6EF}, {0x1F6FD, 0x1F6FF}, {0x1F774, 0x1F77F}, {0x1F7D9, 0x1F7DF}, {0x1F7EC, 0x1F7FF}, {0x1F80C, 0x1F80F}, {0x1F848, 0x1F84F},
+{0x1F85A, 0x1F85F}, {0x1F888, 0x1F88F}, {0x1F8AE, 0x1F8AF}, {0x1F8B2, 0x1F8FF}, {0x1F979, 0x1F979}, {0x1F9CC, 0x1F9CC}, {0x1FA54, 0x1FA5F}, {0x1FA6E, 0x1FA6F}, {0x1FA75, 0x1FA77}, {0x1FA7B, 0x1FA7F},
+{0x1FA87, 0x1FA8F}, {0x1FAA9, 0x1FAAF}, {0x1FAB7, 0x1FABF}, {0x1FAC3, 0x1FACF}, {0x1FAD7, 0x1FAFF}, {0x1FB93, 0x1FB93}, {0x1FBCB, 0x1FBEF}, {0x1FBFA, 0x1FFFF}, {0x2A6DE, 0x2A6FF}, {0x2B735, 0x2B73F},
+{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
+};
+
+static std::string codepoint_to_utf8(uint32_t cp) {
+    std::string result;
+    if (/* 0x00 <= cp && */ cp <= 0x7f) {
+        result.push_back(cp);
+    }
+    else if (0x80 <= cp && cp <= 0x7ff) {
+        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x800 <= cp && cp <= 0xffff) {
+        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.push_back(0xf0 | ((cp >> 18) & 0x07));
+        result.push_back(0x80 | ((cp >> 12) & 0x3f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::string codepoints_to_utf8(const std::vector<uint32_t> & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(codepoint_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
+    assert(offset < utf8.size());
+    if (!(utf8[offset + 0] & 0x80)) {
+        auto result = utf8[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x40)) {
+        throw std::invalid_argument("invalid character");
+    }
+    else if (!(utf8[offset + 0] & 0x20)) {
+        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
+        offset += 2;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x10)) {
+        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
+        offset += 3;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x08)) {
+        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
+        offset += 4;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf8(const std::string & utf8) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf8.size()) {
+        result.push_back(codepoint_from_utf8(utf8, offset));
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoint_to_utf16(uint32_t cp) {
+    std::vector<uint16_t> result;
+    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
+        result.emplace_back(cp);
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
+        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoints_to_utf16(const std::vector<uint32_t> & cps) {
+    std::vector<uint16_t> result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        auto temp = codepoint_to_utf16(cps[i]);
+        result.insert(result.end(), temp.begin(), temp.end());
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
+    assert(offset < utf16.size());
+    if (((utf16[0] >> 10) << 10) != 0xd800) {
+        auto result = utf16[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else {
+        if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
+            throw std::invalid_argument("invalid character");
+        auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
+        offset += 2;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf16.size())
+        result.push_back(codepoint_from_utf16(utf16, offset));
+    return result;
+}
+
+#define CODEPOINT_TYPE_UNIDENTIFIED 0
+#define CODEPOINT_TYPE_DIGIT 1
+#define CODEPOINT_TYPE_LETTER 2
+#define CODEPOINT_TYPE_WHITESPACE 3
+#define CODEPOINT_TYPE_ACCENT_MARK 4
+#define CODEPOINT_TYPE_PUNCTUATION 5
+#define CODEPOINT_TYPE_SYMBOL 6
+#define CODEPOINT_TYPE_CONTROL 7
+
+static std::unordered_map<uint32_t, int> codepoint_type_map() {
+    std::unordered_map<uint32_t, int> codepoint_types;
+    for (auto p : digit_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
+    }
+    for(auto p : letter_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_LETTER;
+    }
+    for(auto p : whitespace_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
+    }
+    for(auto p : accent_mark_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
+    }
+    for(auto p : punctuation_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
+    }
+    for (auto p : symbol_ranges) {
+        for (auto i = p.first; i <= p.second; ++i)
+            codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
+    }
+    for(auto p : control_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
+    }
+    return codepoint_types;
+}
+
+static int codepoint_type(uint32_t cp) {
+    static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
+    return codepoint_types[cp];
+}
+
+static int codepoint_type(const std::string & utf8) {
+    if (utf8.length() == 0)
+        return CODEPOINT_TYPE_UNIDENTIFIED;
+    size_t offset = 0;
+    return codepoint_type(codepoint_from_utf8(utf8, offset));
+}
+
+static std::unordered_map<uint8_t, std::string> bytes_to_unicode_map_bpe() {
+    std::unordered_map<uint8_t, std::string> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(ch) == map.end()) {
+            map[ch] = codepoint_to_utf8(256 + n);
+            ++n;
+        }
+    }
+    return map;
+}
+
+static std::string bytes_to_unicode_bpe(uint8_t byte) {
+    static std::unordered_map<uint8_t, std::string> map = bytes_to_unicode_map_bpe();
+    return map.at(byte);
+}
+
+static std::unordered_map<std::string, uint8_t> unicode_to_bytes_map_bpe() {
+    std::unordered_map<std::string, uint8_t> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(codepoint_to_utf8(ch)) == map.end()) {
+            map[codepoint_to_utf8(256 + n)] = ch;
+            ++n;
+        }
+    }
+    return map;
+}
+
+static uint8_t unicode_to_bytes_bpe(const std::string & utf8) {
+    static std::unordered_map<std::string, uint8_t> map = unicode_to_bytes_map_bpe();
+    return map.at(utf8);
+}
+