Merge branch 'gg/flash-attn' of https://github.com/ggerganov/llama.cpp into flash-attn-cuda

.devops/main-intel.Dockerfile

@@ -0,0 +1,26 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+ARG UBUNTU_VERSION=22.04
+
+FROM intel/hpckit:$ONEAPI_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
+    cmake --build . --config Release --target main server
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/build/bin/main /main
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

.devops/nix/nixpkgs-instances.nix

@@ -7,6 +7,18 @@
     { system, ... }:
     {
       _module.args = {
+        # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs
+        # again, the below creates several nixpkgs instances which the
+        # flake-centric CLI will be forced to evaluate e.g. on `nix flake show`.
+        #
+        # This is currently "slow" and "expensive", on a certain scale.
+        # This also isn't "right" in that this hinders dependency injection at
+        # the level of flake inputs. This might get removed in the foreseeable
+        # future.
+        #
+        # Note that you can use these expressions without Nix
+        # (`pkgs.callPackage ./devops/nix/scope.nix { }` is the entry point).
+
         pkgsCuda = import inputs.nixpkgs {
           inherit system;
           # Ensure dependencies use CUDA consistently (e.g. that openmpi, ucc,

.devops/nix/package.nix

@@ -73,6 +73,7 @@ let
     ps: [
       ps.numpy
       ps.sentencepiece
+      ps.tiktoken
       ps.torchWithoutCuda
       ps.transformers
     ]
@@ -114,14 +115,22 @@ effectiveStdenv.mkDerivation (
     pname = "llama-cpp${pnameSuffix}";
     version = llamaVersion;
 
+    # Note: none of the files discarded here are visible in the sandbox or
+    # affect the output hash. This also means they can be modified without
+    # triggering a rebuild.
     src = lib.cleanSourceWith {
       filter =
         name: type:
-        !(builtins.any (_: _) [
+        let
+          noneOf = builtins.all (x: !x);
+          baseName = baseNameOf name;
+        in
+        noneOf [
           (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
-          (name == "README.md") # Ignore *.md changes whe computing outPaths
+          (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
-          (lib.hasPrefix "." name) # Skip hidden files and directories
+          (lib.hasPrefix "." baseName) # Skip hidden files and directories
-        ]);
+          (baseName == "flake.lock")
+        ];
       src = lib.cleanSource ../../.;
     };
 
@@ -159,7 +168,7 @@ effectiveStdenv.mkDerivation (
 
     cmakeFlags =
       [
-        (cmakeBool "LLAMA_NATIVE" true)
+        (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
         (cmakeBool "BUILD_SHARED_LIBS" true)
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
@@ -216,6 +225,9 @@ effectiveStdenv.mkDerivation (
         description = "contains numpy and sentencepiece";
         buildInputs = [ llama-python ];
         inputsFrom = [ finalAttrs.finalPackage ];
+        shellHook = ''
+          addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib"
+        '';
       };
 
       shell-extra = mkShell {

.devops/nix/scope.nix

@@ -4,6 +4,10 @@
   llamaVersion ? "0.0.0",
 }:
 
+# We're using `makeScope` instead of just writing out an attrset
+# because it allows users to apply overlays later using `overrideScope'`.
+# Cf. https://noogle.dev/f/lib/makeScope
+
 lib.makeScope newScope (
   self: {
     inherit llamaVersion;

.devops/server-cuda.Dockerfile

@@ -0,0 +1,32 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/server /server
+
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -0,0 +1,25 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+ARG UBUNTU_VERSION=22.04
+
+FROM intel/hpckit:$ONEAPI_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
+    cmake --build . --config Release --target main server
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -0,0 +1,45 @@
+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 ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${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/server" ]

.devops/server.Dockerfile

@@ -0,0 +1,20 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+RUN make
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.github/workflows/build.yml

@@ -72,7 +72,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -107,7 +107,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -141,7 +141,48 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest -L main --verbose
+
+  ubuntu-22-cmake-sycl:
+    runs-on: ubuntu-22.04
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake --build . --config Release -j $(nproc)
 
   # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
   #       how to debug it.
@@ -202,7 +243,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -295,7 +336,7 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
-      SDE_VERSION: 9.21.1-2023-04-24
+      SDE_VERSION: 9.33.0-2024-01-07
 
     strategy:
       matrix:
@@ -394,19 +435,19 @@ jobs:
         if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
         run: |
           cd build
-          ctest -C Release --verbose --timeout 900
+          ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
         if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
         run: |
-          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz"
+          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
           # for some weird reason windows tar doesn't like sde tar.xz
           7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
           7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
           $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
           cd build
-          & $sde -future -- ctest -C Release --verbose --timeout 900
+          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
 
       - name: Determine tag name
         id: tag

.github/workflows/docker.yml

@@ -28,13 +28,18 @@ jobs:
         config:
           - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/server.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: "server-cuda", dockerfile: ".devops/server-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" }
+          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3

.github/workflows/nix-ci-aarch64.yml

@@ -2,13 +2,20 @@ name: Nix aarch64 builds
 
 on:
   workflow_dispatch: # allows manual triggering
+  schedule:
+    # Rebuild daily rather than on every push because QEMU is expensive (e.g.
+    # 1.5h instead of minutes with the cold cache).
+    #
+    # randint(0, 59), randint(0, 23)
+    - cron: '26 12 * * *'
+  # But also rebuild if we touched any of the Nix expressions:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+    paths: ['**/*.nix', 'flake.lock']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+    paths: ['**/*.nix', 'flake.lock']
 
 jobs:
   nix-build-aarch64:

.github/workflows/nix-ci.yml

@@ -5,10 +5,8 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
 
 jobs:
   nix-eval:

.gitignore

@@ -27,7 +27,7 @@
 lcov-report/
 gcovr-report/
 
-build*/
+build*
 out/
 tmp/
 
@@ -89,20 +89,3 @@ examples/jeopardy/results.txt
 
 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-llama
-/tests/test-tokenizer-0-falcon
-/tests/test-tokenizer-1-llama
-/tests/test-tokenizer-1-bpe
-/tests/test-rope
-/tests/test-backend-ops
-/tests/test-autorelease

CMakeLists.txt

@@ -1,5 +1,6 @@
 cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
+include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
@@ -47,6 +48,7 @@ option(BUILD_SHARED_LIBS                "build shared libraries"
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
 option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
+option(LLAMA_CCACHE                     "llama: use ccache if available"                        ON)
 
 # debug
 option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
@@ -97,24 +99,38 @@ set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
+option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
+option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
+
+# add perf arguments
+option(LLAMA_PERF                            "llama: enable perf"                               OFF)
+if (LLAMA_PERF)
+    add_definitions(-DGGML_PERF)
+endif()
+
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 
 #
 # Compile flags
 #
-
+if (LLAMA_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
     set(CMAKE_CXX_STANDARD 11)
+endif()
+
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
@@ -401,6 +417,22 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_VULKAN)
+    find_package(Vulkan)
+    if (Vulkan_FOUND)
+        message(STATUS "Vulkan found")
+
+        add_library(ggml-vulkan STATIC ggml-vulkan.cpp ggml-vulkan.h)
+        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+
+        add_compile_definitions(GGML_USE_VULKAN)
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+    else()
+        message(WARNING "Vulkan not found")
+    endif()
+endif()
+
 if (LLAMA_HIPBLAS)
     list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 
@@ -446,6 +478,32 @@ if (LLAMA_HIPBLAS)
     endif()
 endif()
 
+
+if (LLAMA_SYCL)
+    if ( NOT DEFINED ENV{ONEAPI_ROOT})
+        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
+    endif()
+    #todo: AOT
+
+    find_package(IntelSYCL REQUIRED)
+    if (LLAMA_SYCL_F16)
+        add_compile_definitions(GGML_SYCL_F16)
+    endif()
+    add_compile_definitions(GGML_USE_SYCL)
+
+    add_compile_options(-I./) #include DPCT
+    add_compile_options(-I/${SYCL_INCLUDE_DIR})
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
+
+    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
+    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -458,17 +516,24 @@ function(get_flags CCID CCVER)
             (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
             (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
         )
-            set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
+            list(APPEND C_FLAGS -Wdouble-promotion)
         endif()
     elseif (CCID STREQUAL "GNU")
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
         if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
+            list(APPEND CXX_FLAGS -Wno-format-truncation)
         endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
+            list(APPEND CXX_FLAGS -Wextra-semi)
+        endif()
+    elseif (CCID MATCHES "Intel")
+        if (NOT LLAMA_SYCL)
+            # enable max optimization level when using Intel compiler
+            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            add_link_options(-fuse-ld=lld -static-intel)
         endif()
     endif()
 
@@ -497,16 +562,18 @@ if (LLAMA_ALL_WARNINGS)
     endif()
 endif()
 
+set(CUDA_CXX_FLAGS "")
+
 if (LLAMA_CUBLAS)
     set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
     if (NOT MSVC)
-        set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
+        list(APPEND CUDA_FLAGS -Wno-pedantic)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
         set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
         if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
         endif()
 
         execute_process(
@@ -534,15 +601,10 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS)  # pass host compiler flags as a single argument
+        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-        if (NOT CUDA_CXX_FLAGS STREQUAL "")
-            set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
     endif()
 endif()
 
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
-endif()
-
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 
@@ -561,6 +623,17 @@ if (LLAMA_LTO)
     endif()
 endif()
 
+if (LLAMA_CCACHE)
+    find_program(LLAMA_CCACHE_FOUND ccache)
+    if (LLAMA_CCACHE_FOUND)
+        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
+        set(ENV{CCACHE_SLOPPINESS} time_macros)
+        message(STATUS "Using ccache")
+    else()
+        message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF")
+    endif ()
+endif()
+
 # this version of Apple ld64 is buggy
 execute_process(
     COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
@@ -594,12 +667,7 @@ if (NOT MSVC)
     endif()
 endif()
 
-function(add_compile_option_cpp ARG)
+set(ARCH_FLAGS "")
-    # Adds a compile option to C/C++ only, but not for Cuda.
-    # Use, e.g., for CPU-architecture flags.
-    add_compile_options($<$<COMPILE_LANGUAGE:CXX>:${ARG}>)
-    add_compile_options($<$<COMPILE_LANGUAGE:C>:${ARG}>)
-endfunction()
 
 if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
@@ -612,19 +680,19 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            add_compile_options(-mfp16-format=ieee)
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
-            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
             # Raspberry Pi 2
-            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            add_compile_options(-mno-unaligned-access)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
@@ -635,7 +703,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
             include(cmake/FindSIMD.cmake)
         endif ()
         if (LLAMA_AVX512)
-            add_compile_option_cpp(/arch:AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
@@ -649,49 +717,61 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            add_compile_option_cpp(/arch:AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
         elseif (LLAMA_AVX)
-            add_compile_option_cpp(/arch:AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
         endif()
     else()
         if (LLAMA_NATIVE)
-            add_compile_option_cpp(-march=native)
+            list(APPEND ARCH_FLAGS -march=native)
         endif()
         if (LLAMA_F16C)
-            add_compile_option_cpp(-mf16c)
+            list(APPEND ARCH_FLAGS -mf16c)
         endif()
         if (LLAMA_FMA)
-            add_compile_option_cpp(-mfma)
+            list(APPEND ARCH_FLAGS -mfma)
         endif()
         if (LLAMA_AVX)
-            add_compile_option_cpp(-mavx)
+            list(APPEND ARCH_FLAGS -mavx)
         endif()
         if (LLAMA_AVX2)
-            add_compile_option_cpp(-mavx2)
+            list(APPEND ARCH_FLAGS -mavx2)
         endif()
         if (LLAMA_AVX512)
-            add_compile_option_cpp(-mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512f)
-            add_compile_option_cpp(-mavx512bw)
+            list(APPEND ARCH_FLAGS -mavx512bw)
         endif()
         if (LLAMA_AVX512_VBMI)
-            add_compile_option_cpp(-mavx512vbmi)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
         endif()
         if (LLAMA_AVX512_VNNI)
-            add_compile_option_cpp(-mavx512vnni)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
     if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        add_compile_options(-mcpu=powerpc64le)
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
     else()
-        add_compile_options(-mcpu=native -mtune=native)
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
         #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
     endif()
 else()
     message(STATUS "Unknown architecture")
 endif()
 
+add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
+add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+
+if (LLAMA_CUBLAS)
+    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
+    list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+endif()
+
 if (MINGW)
     # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
@@ -771,6 +851,7 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
             ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
             ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}   ${GGML_HEADERS_SYCL}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

Makefile

@@ -9,7 +9,7 @@ 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 tests/test-rope      \
-	tests/test-backend-ops tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -448,6 +448,19 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
+ifdef LLAMA_VULKAN
+	MK_CPPFLAGS  += -DGGML_USE_VULKAN
+	MK_LDFLAGS += -lvulkan
+	OBJS    += ggml-vulkan.o
+
+ifdef LLAMA_VULKAN_CHECK_RESULTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
+endif
+
+ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_VULKAN
+
 ifdef LLAMA_HIPBLAS
 
 	ifeq ($(wildcard /opt/rocm),)
@@ -619,7 +632,7 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
@@ -748,5 +761,8 @@ tests/test-c.o: tests/test-c.c llama.h
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

README.md

@@ -10,11 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- ⚠️ Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
 - New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
-- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -63,7 +63,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
-- CUDA, Metal and OpenCL GPU backend support
+- CUDA, Metal, OpenCL, SYCL 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 mainly for educational purposes and serves
@@ -112,6 +112,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
 - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
+- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 
 
 **Bindings:**
@@ -121,13 +122,15 @@ as the main playground for developing new features for the [ggml](https://github
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - 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)
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
+- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 
 **UI:**
 
@@ -596,6 +599,15 @@ Building the program with BLAS support may lead to some performance improvements
 
   You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
+- #### SYCL
+
+  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+  llama.cpp based on SYCL is used to support Intel GPU (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+  For detailed info, please refer to [llama.cpp for SYCL](README_sycl.md).
+
+
 ### Prepare Data & Run
 
 ```bash
@@ -929,17 +941,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 * Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
-We have two Docker images available for this project:
+We have three 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. (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`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle 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:server-cuda`: Same as `server` 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`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` 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 GitHub 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).
 
@@ -965,6 +980,12 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
 ### Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -974,6 +995,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
 docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -987,6 +1009,7 @@ The resulting images, are essentially the same as the non-CUDA images:
 
 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
 
 #### Usage
 
@@ -995,6 +1018,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ### Contributing

README_sycl.md

@@ -0,0 +1,252 @@
+# llama.cpp for SYCL
+
+[Background](#background)
+
+[OS](#os)
+
+[Intel GPU](#intel-gpu)
+
+[Linux](#linux)
+
+[Environment Variable](#environment-variable)
+
+[Known Issue](#known-issue)
+
+[Todo](#todo)
+
+## Background
+
+SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
+
+oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
+
+Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
+
+To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
+
+The llama.cpp for SYCL is used to support Intel GPUs.
+
+For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
+
+## OS
+
+|OS|Status|Verified|
+|-|-|-|
+|Linux|Support|Ubuntu 22.04|
+|Windows|Ongoing| |
+
+
+## Intel GPU
+
+|Intel GPU| Status | Verified Model|
+|-|-|-|
+|Intel Data Center Max Series| Support| Max 1550|
+|Intel Data Center Flex Series| Support| Flex 170|
+|Intel Arc Series| Support| Arc 770|
+|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
+|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
+
+
+## Linux
+
+### Setup Environment
+
+1. Install Intel GPU driver.
+
+a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
+
+Note: for iGPU, please install the client GPU driver.
+
+b. Add user to group: video, render.
+
+```
+sudo usermod -aG render username
+sudo usermod -aG video username
+```
+
+Note: re-login to enable it.
+
+c. Check
+
+```
+sudo apt install clinfo
+sudo clinfo -l
+```
+
+Output (example):
+
+```
+Platform #0: Intel(R) OpenCL Graphics
+ `-- Device #0: Intel(R) Arc(TM) A770 Graphics
+
+
+Platform #0: Intel(R) OpenCL HD Graphics
+ `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
+```
+
+2. Install Intel® oneAPI Base toolkit.
+
+
+a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+
+Recommend to install to default folder: **/opt/intel/oneapi**.
+
+Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
+
+b. Check
+
+```
+source /opt/intel/oneapi/setvars.sh
+
+sycl-ls
+```
+
+There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
+
+Output (example):
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+
+```
+
+2. Build locally:
+
+```
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main only
+#cmake --build . --config Release --target main
+
+#build all binary
+cmake --build . --config Release -v
+
+```
+
+or
+
+```
+./examples/sycl/build.sh
+```
+
+Note:
+
+- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+
+### Run
+
+1. Put model file to folder **models**
+
+2. Enable oneAPI running environment
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. List device ID
+
+Run without parameter:
+
+```
+./build/bin/ls-sycl-device
+
+or
+
+./build/bin/main
+```
+
+Check the ID in startup log, like:
+
+```
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
+```
+
+|Attribute|Note|
+|-|-|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+
+4. Set device ID and execute llama.cpp
+
+Set device ID = 0 by **GGML_SYCL_DEVICE=0**
+
+```
+GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+or run by script:
+
+```
+./examples/sycl/run_llama2.sh
+```
+
+Note:
+
+- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+
+
+5. Check the device ID in output
+
+Like：
+```
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+```
+
+
+## Environment Variable
+
+#### Build
+
+|Name|Value|Function|
+|-|-|-|
+|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
+|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
+|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
+|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
+
+#### Running
+
+
+|Name|Value|Function|
+|-|-|-|
+|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
+|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
+
+## Known Issue
+
+- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+
+  Miss to enable oneAPI running environment.
+
+  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
+
+
+- Hang during startup
+
+  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
+
+  Solution: add **--no-mmap**.
+
+## Todo
+
+- Support to build in Windows.
+
+- Support multiple cards.

ci/README.md

@@ -22,4 +22,8 @@ bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
 # with CUDA support
 GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with SYCL support
+source /opt/intel/oneapi/setvars.sh
+GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```

ci/run.sh

@@ -10,6 +10,9 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with SYCL support
+# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -22,9 +25,9 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -v $OUT/*.log
+rm -f "$OUT/*.log"
-rm -v $OUT/*.exit
+rm -f "$OUT/*.exit"
-rm -v $OUT/*.md
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
@@ -40,6 +43,14 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
 fi
 
+if [ ! -z ${GG_BUILD_SYCL} ]; then
+    if [ -z ${ONEAPI_ROOT} ]; then
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
+        exit 1
+    fi
+
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON"
+fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -94,7 +105,7 @@ function gg_run_ctest_debug {
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -123,9 +134,9 @@ function gg_run_ctest_release {
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -141,6 +152,61 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+function gg_get_model {
+    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
+    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_3b ]]; then
+        echo -n "$gguf_3b"
+    elif [[ -s $gguf_7b ]]; then
+        echo -n "$gguf_7b"
+    else
+        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
+        exit 1
+    fi
+}
+
+function gg_run_ctest_with_model_debug {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-debug
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_run_ctest_with_model_release {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-release
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_sum_ctest_with_model_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+function gg_sum_ctest_with_model_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
 # open_llama_3b_v2
 
 function gg_run_open_llama_3b_v2 {
@@ -183,8 +249,6 @@ function gg_run_open_llama_3b_v2 {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
-    ./bin/test-autorelease ${model_f16}
-
     ./bin/quantize ${model_f16} ${model_q8_0} q8_0
     ./bin/quantize ${model_f16} ${model_q4_0} q4_0
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1
@@ -507,14 +571,18 @@ function gg_sum_open_llama_7b_v2 {
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
-
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    python3 -m pip install -r ${SRC}/requirements.txt
+    # Create a fresh python3 venv and enter it
-    python3 -m pip install --editable gguf-py
+    python3 -m venv "$MNT/venv"
+    source "$MNT/venv/bin/activate"
+
+    pip install -r ${SRC}/requirements.txt --disable-pip-version-check
+    pip install --editable gguf-py --disable-pip-version-check
 fi
 
 ret=0
@@ -529,6 +597,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
         else
             test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
+        test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi
 

common/common.cpp

@@ -42,6 +42,10 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL))
+#define GGML_USE_CUBLAS_SYCL
+#endif
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -203,6 +207,23 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.prompt_cache_all = true;
         } else if (arg == "--prompt-cache-ro") {
             params.prompt_cache_ro = true;
+        } else if (arg == "-bf" || arg == "--binary-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i], std::ios::binary);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            // store the external file name in params
+            params.prompt_file = argv[i];
+            std::ostringstream ss;
+            ss << file.rdbuf();
+            params.prompt = ss.str();
+            fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), argv[i]);
         } else if (arg == "-f" || arg == "--file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -582,9 +603,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.main_gpu = std::stoi(argv[i]);
-#ifndef GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the main GPU has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#endif // GGML_USE_CUBLAS_SYCL
         } else if (arg == "--split-mode" || arg == "-sm") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -601,9 +622,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifndef GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#endif // GGML_USE_CUBLAS_SYCL
+
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -626,9 +648,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#ifndef GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting a tensor split has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#endif // GGML_USE_CUBLAS_SYCL
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
@@ -653,6 +675,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             if (params.logdir.back() != DIRECTORY_SEPARATOR) {
                 params.logdir += DIRECTORY_SEPARATOR;
             }
+        } else if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.logits_file = argv[i];
         } else if (arg == "--perplexity" || arg == "--all-logits") {
             params.logits_all = true;
         } else if (arg == "--ppl-stride") {
@@ -689,6 +717,16 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.winogrande_tasks = std::stoi(argv[i]);
+        } else if (arg == "--multiple-choice") {
+            params.multiple_choice = true;
+        } else if (arg == "--multiple-choice-tasks") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.multiple_choice_tasks = std::stoi(argv[i]);
+        } else if (arg == "--kl-divergence") {
+            params.kl_divergence = true;
         } else if (arg == "--ignore-eos") {
             params.ignore_eos = true;
         } else if (arg == "--no-penalize-nl") {
@@ -888,6 +926,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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("  -bf FNAME, --binary-file FNAME\n");
+    printf("                        binary file containing multiple choice tasks.\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);
@@ -936,6 +976,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
     printf("  --winogrande          compute Winogrande score over random tasks from datafile supplied with -f\n");
     printf("  --winogrande-tasks N  number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
+    printf("  --multiple-choice     compute multiple choice score over random tasks from datafile supplied with -f\n");
+    printf("  --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
+    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base");
     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);
@@ -969,7 +1012,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
     printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
     printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
-#endif
+#endif // LLAMA_SUPPORTS_GPU_OFFLOAD
     printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
     printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
     printf("  -gan N, --grp-attn-n N\n");
@@ -1476,7 +1519,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     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");

common/common.h

@@ -91,6 +91,7 @@ struct gpt_params {
     std::string input_suffix      = "";  // string to suffix user inputs with
     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 logits_file       = "";  // file for saving *all* logits
 
     std::vector<llama_model_kv_override> kv_overrides;
 
@@ -108,6 +109,11 @@ struct gpt_params {
     bool   winogrande      = false; // compute Winogrande score over random tasks from datafile supplied in prompt
     size_t winogrande_tasks= 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
 
+    bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
+    size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
+
+    bool   kl_divergence   = false; // compute KL-divergence
+
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

common/sampling.cpp

@@ -13,6 +13,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
         // will be empty (default) if there are parse errors
         if (result->parsed_grammar.rules.empty()) {
             fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            delete result;
             return nullptr;
         }
 
@@ -129,6 +130,8 @@ static void sampler_queue(
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
     const float         temp              = params.temp;
+    const float         dynatemp_range    = params.dynatemp_range;
+    const float         dynatemp_exponent = params.dynatemp_exponent;
     const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
     const float         top_p             = params.top_p;
     const float         min_p             = params.min_p;
@@ -143,7 +146,15 @@ static void sampler_queue(
             case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
             case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
             case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't': llama_sample_temp     (ctx_main, &cur_p, temp); break;
+            case 't':
+                if (dynatemp_range > 0) {
+                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
+                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
+                    llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
+                } else {
+                    llama_sample_temp(ctx_main, &cur_p, temp);
+                }
+                break;
             default : break;
         }
     }

common/sampling.h

@@ -18,6 +18,8 @@ typedef struct llama_sampling_params {
     float       tfs_z                 = 1.00f;    // 1.0 = disabled
     float       typical_p             = 1.00f;    // 1.0 = disabled
     float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float       dynatemp_range        = 0.00f;    // 0.0 = disabled
+    float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
     int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float       penalty_repeat        = 1.10f;    // 1.0 = disabled
     float       penalty_freq          = 0.00f;    // 0.0 = disabled

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@ import re
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
 
 import numpy as np
 import torch
@@ -201,6 +201,8 @@ class Model:
             return PlamoModel
         if model_architecture == "CodeShellForCausalLM":
             return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -250,6 +252,8 @@ class Model:
             return gguf.MODEL_ARCH.PLAMO
         if arch == "CodeShellForCausalLM":
             return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -289,6 +293,58 @@ class Model:
         special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_qwen(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams["vocab_size"]
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        merges = []
+        vocab = {}
+        mergeable_ranks = tokenizer.mergeable_ranks
+        for token, rank in mergeable_ranks.items():
+            vocab[QwenModel.token_bytes_to_string(token)] = rank
+            if len(token) == 1:
+                continue
+            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+            assert len(merged) == 2
+            merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+
+        # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
+        added_vocab = tokenizer.special_tokens
+        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in (vocab | added_vocab).items()}
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                pad_token = f"[PAD{i}]".encode("utf-8")
+                tokens.append(bytearray(pad_token))
+                toktypes.append(gguf.TokenType.USER_DEFINED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.CONTROL)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
+        special_vocab.merges = merges
+        # only add special tokens when they were not already loaded from config.json
+        if len(special_vocab.special_token_ids) == 0:
+            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
+            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
+        # this one is usually not in config.json anyway
+        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_sentencepiece(self):
         from sentencepiece import SentencePieceProcessor
 
@@ -487,6 +543,7 @@ class MPTModel(Model):
             # map tensor names
             if "scales" in name:
                 new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales"))
+                if new_name is not None:
                     new_name = new_name.replace("scales", "act.scales")
             else:
                 new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
@@ -519,6 +576,83 @@ class MPTModel(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
+class OrionModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.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 self.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 self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 class BaichuanModel(Model):
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -876,6 +1010,13 @@ class PersimmonModel(Model):
 
 
 class StableLMModel(Model):
+    def set_vocab(self):
+        if (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab
+            self._set_vocab_qwen()
+
     def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
@@ -904,7 +1045,7 @@ class QwenModel(Model):
         return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
 
     @staticmethod
-    def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: Optional[int] = None) -> list[bytes]:
+    def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]:
         parts = [bytes([b]) for b in token]
         while True:
             min_idx = None
@@ -921,52 +1062,7 @@ class QwenModel(Model):
         return parts
 
     def set_vocab(self):
-        dir_model = self.dir_model
+        self._set_vocab_qwen()
-        hparams = self.hparams
-        tokens: list[bytearray] = []
-        toktypes: list[int] = []
-
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
-        vocab_size = hparams["vocab_size"]
-        assert max(tokenizer.get_vocab().values()) < vocab_size
-
-        merges = []
-        vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks
-        for token, rank in mergeable_ranks.items():
-            vocab[self.token_bytes_to_string(token)] = rank
-            if len(token) == 1:
-                continue
-            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
-            assert len(merged) == 2
-            merges.append(' '.join(map(self.token_bytes_to_string, merged)))
-
-        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.special_tokens
-
-        for i in range(vocab_size):
-            if i not in reverse_vocab:
-                pad_token = f"[PAD{i}]".encode("utf-8")
-                tokens.append(bytearray(pad_token))
-                toktypes.append(gguf.TokenType.USER_DEFINED)
-            elif reverse_vocab[i] in added_vocab:
-                tokens.append(reverse_vocab[i])
-                toktypes.append(gguf.TokenType.CONTROL)
-            else:
-                tokens.append(reverse_vocab[i])
-                toktypes.append(gguf.TokenType.NORMAL)
-
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
-        special_vocab.merges = merges
-        special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
-        special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
-        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
-        special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("Qwen")
@@ -1285,7 +1381,7 @@ def main() -> None:
 
     if args.awq_path:
         sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
-        from awq.apply_awq import add_scale_weights
+        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
         tmp_model_path = args.model / "weighted_model"
         dir_model = tmp_model_path
         if tmp_model_path.is_dir():

convert-llama-ggml-to-gguf.py

@@ -2,6 +2,7 @@
 from __future__ import annotations
 
 import argparse
+import os
 import struct
 import sys
 from enum import IntEnum
@@ -9,7 +10,6 @@ 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'))
 import gguf
@@ -371,15 +371,11 @@ def handle_metadata(cfg, hp):
         params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
     else:
         raise ValueError('Unable to load metadata')
-    vocab = convert.load_vocab(
+    vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
-        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+    vocab_factory = convert.VocabFactory(vocab_path)
-        cfg.vocabtype)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir)
-    # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
-                               load_merges = cfg.vocabtype == 'bpe',
-                               n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
-    return (params, vocab, svocab)
+    return params, vocab, special_vocab
 
 
 def handle_args():

convert-lora-to-ggml.py

@@ -5,17 +5,16 @@ import json
 import os
 import struct
 import sys
+from pathlib import Path
 from typing import Any, BinaryIO, Sequence
 
 import numpy as np
 import torch
 
-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
 
-
 NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
@@ -60,7 +59,14 @@ if __name__ == '__main__':
     input_model = os.path.join(sys.argv[1], "adapter_model.bin")
     output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
 
+    if os.path.exists(input_model):
         model = torch.load(input_model, map_location="cpu")
+    else:
+        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
+        # lazy import load_file only if lora is in safetensors format.
+        from safetensors.torch import load_file
+        model = load_file(input_model, device="cpu")
+
     arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
 
     if arch_name not in gguf.MODEL_ARCH_NAMES.values():

convert-persimmon-to-gguf.py

@@ -1,11 +1,13 @@
 #!/usr/bin/env python3
-import torch
-import os
-from pprint import pprint
-import sys
 import argparse
+import os
+import sys
 from pathlib import Path
+from pprint import pprint
+
+import torch
 from sentencepiece import SentencePieceProcessor
+
 if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
@@ -69,7 +71,7 @@ def main():
     persimmon_model = torch.load(args.ckpt_path)
     hparams = persimmon_model['args']
     pprint(hparams)
-    tensors = {}
+    tensors: dict[str, torch.Tensor] = {}
     _flatten_dict(persimmon_model['model'], tensors, None)
 
     arch = gguf.MODEL_ARCH.PERSIMMON

convert.py

@@ -17,58 +17,28 @@ import signal
 import struct
 import sys
 import time
-import warnings
 import zipfile
 from abc import ABCMeta, abstractmethod
-from argparse import ArgumentParser
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import (
+from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
-    IO,
-    TYPE_CHECKING,
-    Any,
-    Callable,
-    Iterable,
-    Literal,
-    Optional,
-    Tuple,
-    TypeVar,
-)
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
 
-try:
+if 'NO_LOCAL_GGUF' not in os.environ:
-    from transformers import AutoTokenizer
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
-except ModuleNotFoundError as e:
-    warnings.warn(f"Could not import AutoTokenizer from transformers: {e}")
-
-# If NO_LOCAL_GGUF is not set, try to import gguf from the local gguf-py directory
-if "NO_LOCAL_GGUF" not in os.environ:
-    # Use absolute path to the gguf-py directory
-    gguf_py_dir = str(Path(__file__).resolve().parent / "gguf-py")
-    print(gguf_py_dir)  # NOTE: Remove this once path is verified after changes are completed
-    if gguf_py_dir not in sys.path:
-        sys.path.insert(1, gguf_py_dir)
-
-# Import gguf module
-try:
 import gguf
-except ModuleNotFoundError as e:
-    print(f"Could not import gguf: {e}")
-    sys.exit(1)
 
-if TYPE_CHECKING:  # NOTE: This isn't necessary.
+if TYPE_CHECKING:
-    from typing import TypeAlias  # This can technically be omitted.
+    from typing import TypeAlias
 
-if hasattr(faulthandler, "register") and hasattr(signal, "SIGUSR1"):
+if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
     faulthandler.register(signal.SIGUSR1)
 
-# NOTE: n-dimensional arrays should be directly referenced
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
-NDArray: TypeAlias = "np.ndarray[Any, Any]"
 
-# Why is this here? LLAMA and GPT are technically the only compatible ARCHs.
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
@@ -78,7 +48,6 @@ DEFAULT_CONCURRENCY = 8
 #
 
 
-# TODO: Clean up and refactor data types
 @dataclass(frozen=True)
 class DataType:
     name: str
@@ -190,57 +159,37 @@ class Params:
     n_ff:           int
     n_head:         int
     n_head_kv:      int
-    f_norm_eps: Optional[float] = None
+    n_experts:      int | None = None
-    n_experts: Optional[int] = None
+    n_experts_used: int | None = None
-    n_experts_used: Optional[int] = None
+    f_norm_eps:     float | None = None
 
-    rope_scaling_type: Optional[gguf.RopeScalingType] = None
+    rope_scaling_type: gguf.RopeScalingType | None = None
-    f_rope_freq_base: Optional[float] = None
+    f_rope_freq_base: float | None = None
-    f_rope_scale: Optional[float] = None
+    f_rope_scale: float | None = None
-    n_orig_ctx: Optional[int] = None
+    n_orig_ctx: int | None = None
-    rope_finetuned: Optional[bool] = None
+    rope_finetuned: bool | None = None
 
-    ftype: Optional[GGMLFileType] = None
+    ftype: GGMLFileType | None = None
 
     # path to the directory containing the model files
-    path_model: Optional[Path] = None
+    path_model: Path | None = None
 
     @staticmethod
-    def guessed(model: LazyModel) -> "Params":
+    def guessed(model: LazyModel) -> Params:
         # try transformer naming first
-        n_vocab, n_embd = (
+        n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape
-            model["model.embed_tokens.weight"].shape
-            if "model.embed_tokens.weight" in model
-            else model["tok_embeddings.weight"].shape
-        )
 
         # try transformer naming first
         if "model.layers.0.self_attn.q_proj.weight" in model:
-            n_layer = next(
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
-                i
+        elif "model.layers.0.self_attn.W_pack.weight" in model:   # next: try baichuan naming
-                for i in itertools.count()
+            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
-                if f"model.layers.{i}.self_attn.q_proj.weight" not in model
-            )
-        elif (
-            "model.layers.0.self_attn.W_pack.weight" in model
-        ):  # next: try baichuan naming
-            n_layer = next(
-                i
-                for i in itertools.count()
-                if f"model.layers.{i}.self_attn.W_pack.weight" not in model
-            )
         else:
-            n_layer = next(
+            n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
-                i
-                for i in itertools.count()
-                if f"layers.{i}.attention.wq.weight" not in model
-            )
 
         if n_layer < 1:
-            raise Exception(
+            raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
-                "failed to guess 'n_layer'. This model is unknown or unsupported.\n"
+                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
-                "Suggestion: provide 'config.json' of the model in the same directory containing model files."
-            )
 
         n_head = n_embd // 128 # guessed
         n_mult = 256           # guessed
@@ -261,7 +210,7 @@ class Params:
         )
 
     @staticmethod
-    def load_transformers_config(model: LazyModel, config_path: Path) -> "Params":
+    def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
         rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
@@ -274,20 +223,18 @@ class Params:
                 rope_scaling_type = gguf.RopeScalingType.LINEAR
             elif typ == "yarn":
                 rope_scaling_type = gguf.RopeScalingType.YARN
-                n_orig_ctx = rope_scaling["original_max_position_embeddings"]
+                n_orig_ctx = rope_scaling['original_max_position_embeddings']
-                rope_finetuned = rope_scaling["finetuned"]
+                rope_finetuned = rope_scaling['finetuned']
             else:
-                raise NotImplementedError(f"Unknown rope scaling type: {typ}")
+                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
         elif "max_position_embeddings" in config:
             n_ctx = config["max_position_embeddings"]
         else:
-            raise Exception(
+            raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
-                "failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
+                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
-                "Suggestion: provide 'config.json' of the model in the same directory containing model files."
-            )
 
         n_experts      = None
         n_experts_used = None
@@ -317,7 +264,7 @@ class Params:
     # 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}
     @staticmethod
-    def load_torch_params(model: LazyModel, config_path: Path) -> "Params":
+    def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
         n_experts      = None
@@ -362,31 +309,31 @@ class Params:
         )
 
     @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"
 
         if hf_config_path.exists():
-            params = Params.load_transformers_config(model_plus.model, hf_config_path)
+            params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
         elif orig_config_path.exists():
-            params = Params.load_torch_params(model_plus.model, orig_config_path)
+            params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
-        elif model_plus.format != "none":
+        elif model_plus.format != 'none':
             params = Params.guessed(model_plus.model)
         else:
-            raise ValueError("Cannot guess params when model format is none")
+            raise ValueError('Cannot guess params when model format is none')
 
         params.path_model = model_plus.paths[0].parent
 
         return params
 
 
-class BpeVocab:  # GPT
+#
-    def __init__(
+# vocab
-        self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]
+#
-    ) -> None:
+
-        self.bpe_tokenizer = json.loads(
+class BpeVocab:
-            open(str(fname_tokenizer), encoding="utf-8").read()
+    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())
         self.vocab = self.bpe_tokenizer["model"]["vocab"]
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
@@ -394,26 +341,23 @@ class BpeVocab:  # GPT
             added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
         else:
             # Fall back to trying to find the added tokens in tokenizer.json
-            tokenizer_json_file = fname_tokenizer.parent / "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"])
+                    (item['content'], item['id'])
-                    for item in tokenizer_json.get("added_tokens", [])
+                    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
+                    if item['content'] not in self.bpe_tokenizer)
-                )
 
         vocab_size: int = len(self.vocab)
         expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids      = sorted(added_tokens.values())
         if expected_ids != actual_ids:
             expected_end_id = vocab_size + len(actual_ids) - 1
-            raise Exception(
+            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}")
-                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_dict    = added_tokens
@@ -442,10 +386,8 @@ class BpeVocab:  # GPT
         return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class SentencePieceVocab:  # LlaMa
+class SentencePieceVocab:
-    def __init__(
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]
-    ) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
@@ -455,16 +397,12 @@ class SentencePieceVocab:  # LlaMa
 
         vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
 
-        new_tokens = {
+        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
-            id: piece for piece, id in added_tokens.items() if id >= vocab_size
-        }
         expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
         actual_new_ids   = sorted(new_tokens.keys())
 
         if expected_new_ids != actual_new_ids:
-            raise ValueError(
+            raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
-                f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}"
-            )
 
         # Token pieces that were added to the base vocabulary.
         self.added_tokens_dict = added_tokens
@@ -512,11 +450,15 @@ class SentencePieceVocab:  # LlaMa
 
 
 class HfVocab:
-    def __init__(
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
-        self,
+        try:
-        fname_tokenizer: Path,
+            from transformers import AutoTokenizer
-        fname_added_tokens: Optional[Path] = None,
+        except ImportError as e:
-    ) -> None:
+            raise ImportError(
+                "To use HfVocab, please install the `transformers` package. "
+                "You can install it with `pip install transformers`."
+            ) from e
+
         print("fname_tokenizer:", fname_tokenizer)
         # Allow the tokenizer to default to slow or fast versions.
         # Explicitly set tokenizer to use local paths.
@@ -555,7 +497,7 @@ class HfVocab:
         self.fname_tokenizer    = fname_tokenizer
         self.fname_added_tokens = fname_added_tokens
 
-    def hf_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]:
+    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {
             id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
         }
@@ -573,11 +515,9 @@ class HfVocab:
                 token_id, self.special_ids  # Reuse already stored special IDs
             )
 
-    def get_token_type(self, token_id: int, special_ids: set) -> gguf.TokenType:
+    def get_token_type(self, token_id: int, special_ids: set[int]) -> gguf.TokenType:
         # Determine token type based on whether it's a special token
-        return (
+        return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
-            gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
-        )
 
     def get_token_score(self, token_id: int) -> float:
         # Placeholder for actual logic to determine the token's score
@@ -589,7 +529,6 @@ class HfVocab:
             if text in self.specials:
                 toktype = self.get_token_type(self.specials[text], self.special_ids)
                 score = self.get_token_score(self.specials[text])
-
             else:
                 toktype = gguf.TokenType.USER_DEFINED
                 score = -1000.0
@@ -783,7 +722,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     else:
         model = merge_sharded([mp.model for mp in models_plus])
 
-    return ModelPlus(model, paths, format, vocab)
+    return ModelPlus(model, paths, format, vocab)  # pytype: disable=wrong-arg-types
 
 
 def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
@@ -871,17 +810,13 @@ class LazyUnpickler(pickle.Unpickler):
     CLASSES: dict[tuple[str, str], Any] = {
         # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
-        ("torch._tensor", "_rebuild_from_type_v2"): getattr(
+        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
-            rebuild_from_type_v2, "__func__"
+        ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
-        ),
+        ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16),
-        ("torch._utils", "_rebuild_tensor_v2"): getattr(
+        ('torch', 'HalfStorage'): LazyStorageKind(DT_F16),
-            lazy_rebuild_tensor_v2, "__func__"
+        ('torch', 'FloatStorage'): LazyStorageKind(DT_F32),
-        ),
+        ('torch', 'IntStorage'): LazyStorageKind(DT_I32),
-        ("torch", "BFloat16Storage"): LazyStorageKind(DT_BF16),
+        ('torch', 'Tensor'): LazyTensor,
-        ("torch", "HalfStorage"): LazyStorageKind(DT_F16),
-        ("torch", "FloatStorage"): LazyStorageKind(DT_F32),
-        ("torch", "IntStorage"): LazyStorageKind(DT_I32),
-        ("torch", "Tensor"): LazyTensor,
     }
 
     def find_class(self, module: str, name: str) -> Any:
@@ -1022,12 +957,8 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
 
 
 class OutputFile:
-    def __init__(
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
-        self, fname_out: Path, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE
+        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
-    ) -> None:
-        self.gguf = gguf.GGUFWriter(
-            fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess
-        )
 
     def add_meta_arch(self, params: Params) -> None:
         name = "LLaMA"
@@ -1036,7 +967,7 @@ class OutputFile:
         if params.n_ctx == 4096:
             name = "LLaMA v2"
         elif params.path_model is not None:
-            name = str(params.path_model.parent).split("/")[-1]
+            name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
         self.gguf.add_context_length      (params.n_ctx)
@@ -1047,17 +978,17 @@ class OutputFile:
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
 
-        if params.f_norm_eps is None:
-            raise ValueError("f_norm_eps is None")
-
-        self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
-
         if params.n_experts:
             self.gguf.add_expert_count(params.n_experts)
 
         if params.n_experts_used:
             self.gguf.add_expert_used_count(params.n_experts_used)
 
+        if params.f_norm_eps:
+            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
+        else:
+            raise ValueError('f_norm_eps is None')
+
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
@@ -1089,7 +1020,7 @@ class OutputFile:
 
         return tokenizer_model
 
-    def extract_vocabulary_from_model(self, vocab: Vocab) -> Tuple[list, list, list]:
+    def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
         tokens = []
         scores = []
         toktypes = []
@@ -1124,14 +1055,10 @@ class OutputFile:
 
     def add_tensor_info(self, name: str, tensor: LazyTensor) -> None:
         n_elements = int(np.prod(tensor.shape))
-        raw_dtype = getattr(tensor.data_type, "ggml_type", None)
+        raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
-        data_type = (
+        data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
-            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(
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype)
-            name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype
-        )
 
     def write_meta(self) -> None:
         self.gguf.write_header_to_file()
@@ -1145,12 +1072,8 @@ class OutputFile:
 
     @staticmethod
     def write_vocab_only(
-        fname_out: Path,
+        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        params: Params,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
-        vocab: Vocab,
-        svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
-        pad_vocab: bool = False,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab = pad_vocab)
 
@@ -1180,14 +1103,8 @@ class OutputFile:
 
     @staticmethod
     def write_all(
-        fname_out: Path,
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
-        ftype: GGMLFileType,
+        concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
-        params: Params,
-        model: LazyModel,
-        vocab: Vocab,
-        svocab: gguf.SpecialVocab,
-        concurrency: int = DEFAULT_CONCURRENCY,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
@@ -1207,26 +1124,19 @@ class OutputFile:
         of.write_tensor_info()
 
         # tensor data
-        ndarrays_inner = bounded_parallel_map(
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
-            OutputFile.do_item, model.items(), concurrency=concurrency
-        )
         if ftype == GGMLFileType.MostlyQ8_0:
             ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize,
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                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(
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            zip(model.items(), ndarrays)
-        ):
             elapsed = time.time() - start
-            size = " x ".join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            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.name:4} | T+{int(elapsed):4}"
@@ -1363,7 +1273,7 @@ def load_some_model(path: Path) -> ModelPlus:
 class VocabFactory:
     def __init__(self, path: Path):
         self.path = path
-        self.files = {
+        self.files: dict[str, Path | None] = {
             "tokenizer.model": None,
             "vocab.json": None,
             "tokenizer.json": None,
@@ -1380,24 +1290,18 @@ class VocabFactory:
                 self.files[file] = parent_file_path
         print(f"Found vocab files: {self.files}")
 
-    def _select_file(self, vocabtype: Optional[str]) -> Path:
+    def _select_file(self, vocabtype: str | None) -> Path:
         if vocabtype in ["spm", "bpe"]:
             for file_key in self.files.keys():
-                if self.files[file_key]:
+                if (file := self.files[file_key]) is not None:
-                    return self.files[file_key]
+                    return file
             raise FileNotFoundError(f"{vocabtype} vocab not found.")
-        elif vocabtype == "hfft":
+        if vocabtype == "hfft":
             # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
             return self.path
-        else:
         raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
-    def _create_special_vocab(
+    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
-        self,
-        vocab: Vocab,
-        vocabtype: str,
-        model_parent_path: Path,
-    ) -> gguf.SpecialVocab:
         load_merges = vocabtype == "bpe"
         n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
         return gguf.SpecialVocab(
@@ -1407,13 +1311,12 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def load_vocab(
+    def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
-        self, vocabtype: str, model_parent_path: Path
-    ) -> Tuple[Vocab, gguf.SpecialVocab]:
         path = self._select_file(vocabtype)
         print(f"Loading vocab file '{path}', type '{vocabtype}'")
 
         added_tokens_path = path.parent / "added_tokens.json"
+        vocab: Vocab
         if vocabtype == "bpe":
             vocab = BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
@@ -1428,6 +1331,7 @@ class VocabFactory:
             )
         else:
             raise ValueError(f"Unsupported vocabulary type {vocabtype}")
+        # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
             vocabtype,
@@ -1436,7 +1340,7 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_output_file(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",
@@ -1446,8 +1350,7 @@ def default_output_file(model_paths: list[Path], file_type: GGMLFileType) -> Pat
     if ret in model_paths:
         sys.stderr.write(
             f"Error: Default output path ({ret}) would overwrite the input. "
-            "Please explicitly specify a path using --outfile.\n"
+            "Please explicitly specify a path using --outfile.\n")
-        )
         sys.exit(1)
     return ret
 
@@ -1457,111 +1360,34 @@ def do_dump_model(model_plus: ModelPlus) -> None:
     print(f"model_plus.format = {model_plus.format!r}")
     print(f"model_plus.vocab = {model_plus.vocab!r}")
     for name, lazy_tensor in model_plus.model.items():
-        print(
+        print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}")
-            f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}"
-        )
 
 
-def get_argument_parser() -> ArgumentParser:
+def main(args_in: list[str] | None = None) -> None:
     output_choices = ["f32", "f16"]
     if np.uint32(1) == np.uint32(1).newbyteorder("<"):
         # We currently only support Q8_0 output on little endian systems.
         output_choices.append("q8_0")
+    vocab_types = ["spm", "bpe", "hfft"]
+    parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
+    parser.add_argument("--awq-path",    type=Path,              help="Path to scale awq cache file", default=None)
+    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=output_choices, 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("--vocab-type",  choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
+    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("--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)
+    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
 
-    parser = argparse.ArgumentParser(
+    args = parser.parse_args(args_in)
-        description="Convert a LLaMa model to a GGML compatible file"
-    )
-
-    parser.add_argument(
-        "model",
-        type=Path,
-        help="Directory containing the model file or the model file itself (*.pth, *.pt, *.bin)",
-    )
-
-    parser.add_argument(
-        "--awq-path",
-        type=Path,
-        help="Path to the Activation-aware Weight Quantization cache file",
-        default=None,
-    )
-
-    parser.add_argument(
-        "--dump",
-        action="store_true",
-        help="Display the model content without converting it",
-    )
-
-    parser.add_argument(
-        "--dump-single",
-        action="store_true",
-        help="Display the content of a single model file without conversion",
-    )
-
-    parser.add_argument(
-        "--vocab-only",
-        action="store_true",
-        help="Extract and output only the vocabulary",
-    )
-
-    parser.add_argument(
-        "--outtype",
-        choices=output_choices,
-        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 the tokenizer.model, if separate from the model file",
-    )
-
-    parser.add_argument(
-        "--vocab-type",
-        choices=["spm", "bpe", "hfft"],  # hfft: Hugging Face Fast Tokenizer
-        default="spm",
-        help="The vocabulary format used to define the tokenizer model (default: spm)",
-    )
-
-    parser.add_argument(
-        "--pad-vocab",
-        action="store_true",
-        help="Add padding tokens when the model's vocabulary size exceeds the tokenizer metadata",
-    )
-
-    parser.add_argument(
-        "--outfile",
-        type=Path,
-        help="Specify the path for the output file (default is based on input)",
-    )
-
-    parser.add_argument(
-        "--ctx", type=int, help="Model training context (default is based on input)"
-    )
-
-    parser.add_argument(
-        "--concurrency",
-        type=int,
-        help=f"Concurrency used for conversion (default: {DEFAULT_CONCURRENCY})",
-        default=DEFAULT_CONCURRENCY,
-    )
-
-    parser.add_argument(
-        "--big-endian",
-        action="store_true",
-        help="Indicate that the model is executed on a big-endian machine",
-    )
-
-    return parser
-
-
-def main(argv: Optional[list[str]] = None) -> None:
-    parser = get_argument_parser()
-    args = parser.parse_args(argv)
-
     if args.awq_path:
-        sys.path.insert(1, str(Path(__file__).resolve().parent / "awq-py"))
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
-        from awq.apply_awq import add_scale_weights
+        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
-
         tmp_model_path = args.model / "weighted_model"
         if tmp_model_path.is_dir():
             print(f"{tmp_model_path} exists as a weighted model.")
@@ -1580,14 +1406,11 @@ def main(argv: Optional[list[str]] = None) -> None:
     if not args.vocab_only:
         model_plus = load_some_model(args.model)
     else:
-        model_plus = ModelPlus(
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
-            model={}, paths=[args.model / "dummy"], format="none", vocab=None
-        )
 
     if args.dump:
         do_dump_model(model_plus)
         return
-
     endianess = gguf.GGUFEndian.LITTLE
     if args.big_endian:
         endianess = gguf.GGUFEndian.BIG
@@ -1595,12 +1418,10 @@ def main(argv: Optional[list[str]] = None) -> None:
     params = Params.load(model_plus)
     if params.n_ctx == -1:
         if args.ctx is None:
-            raise Exception(
+            raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n"
-                "The model doesn't have a context size, and you didn't specify one with --ctx\n"
                             "Please specify one with --ctx:\n"
                             " - LLaMA v1: --ctx 2048\n"
-                " - LLaMA v2: --ctx 4096\n"
+                            " - LLaMA v2: --ctx 4096\n")
-            )
         params.n_ctx = args.ctx
 
     if args.outtype:
@@ -1621,42 +1442,30 @@ def main(argv: Optional[list[str]] = None) -> None:
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
         outfile = args.outfile
-        OutputFile.write_vocab_only(
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-            outfile,
+                                    endianess=endianess, pad_vocab=args.pad_vocab)
-            params,
-            vocab,
-            special_vocab,
-            endianess=endianess,
-            pad_vocab=args.pad_vocab,
-        )
         print(f"Wrote {outfile}")
         return
 
     if model_plus.vocab is not None and args.vocab_dir is None:
         vocab = model_plus.vocab
 
+    print(f"Vocab info: {vocab}")
+    print(f"Special vocab info: {special_vocab}")
+
     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_output_file(model_plus.paths, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
 
     params.ftype = ftype
     print(f"Writing {outfile}, format {ftype}")
 
-    OutputFile.write_all(
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-        outfile,
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
-        ftype,
-        params,
-        model,
-        vocab,
-        special_vocab,
-        concurrency=args.concurrency,
-        endianess=endianess,
-        pad_vocab=args.pad_vocab,
-    )
     print(f"Wrote {outfile}")
 
 
-if __name__ == "__main__":
+if __name__ == '__main__':
-    main(sys.argv[1:])  # Exclude the first element (script name) from sys.argv
+    main()

examples/CMakeLists.txt

@@ -23,6 +23,9 @@ else()
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
+    if (LLAMA_SYCL)
+        add_subdirectory(sycl)
+    endif()
     add_subdirectory(main)
     add_subdirectory(tokenize)
     add_subdirectory(parallel)

examples/batched-bench/batched-bench.cpp

@@ -104,7 +104,7 @@ int main(int argc, char ** argv) {
 
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
+    ctx_params.n_batch   = 2048;
     ctx_params.mul_mat_q = mmq;
 
     ctx_params.n_threads       = params.n_threads;

examples/finetune/finetune.cpp

@@ -1800,6 +1800,8 @@ int main(int argc, char ** argv) {
     std::vector<size_t> train_samples_begin;
     std::vector<size_t> train_samples_size;
     printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data);
+    printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str());
+    printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false");
     tokenize_file(lctx,
             params.common.fn_train_data,
             params.common.sample_start,

examples/imatrix/imatrix.cpp

@@ -26,6 +26,7 @@ struct StatParams {
     std::string ofile = "imatrix.dat";
     int         n_output_frequency = 10;
     int         verbosity = 1;
+    int         keep_every = 0;
     bool        collect_output_weight = false;
 };
 
@@ -42,6 +43,9 @@ private:
     int                                    m_last_call = 0;
     std::vector<float>                     m_src1_data;
     std::vector<int>                       m_ids; // the expert ids from ggml_mul_mat_id
+                                                  //
+    void save_imatrix(const char * file_name) const;
+    void keep_imatrix(int ncall) const;
 };
 
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
@@ -117,6 +121,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                 if (m_last_call % m_params.n_output_frequency == 0) {
                     save_imatrix();
                 }
+                if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
+                    keep_imatrix(m_last_call);
+                }
             }
         }
     } else {
@@ -143,6 +150,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             if (m_last_call % m_params.n_output_frequency == 0) {
                 save_imatrix();
             }
+            if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
+                keep_imatrix(m_last_call);
+            }
         }
     }
 
@@ -150,7 +160,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 }
 
 void IMatrixCollector::save_imatrix() const {
-    const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str();
+    save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str());
+}
+
+void IMatrixCollector::keep_imatrix(int ncall) const {
+    auto file_name = m_params.ofile;
+    if (file_name.empty()) file_name = "imatrix.dat";
+    file_name += ".at_";
+    file_name += std::to_string(ncall);
+    save_imatrix(file_name.c_str());
+}
+
+void IMatrixCollector::save_imatrix(const char * fname) const {
     std::ofstream out(fname, std::ios::binary);
     int n_entries = m_stats.size();
     out.write((const char*)&n_entries, sizeof(n_entries));
@@ -248,7 +269,7 @@ static void process_logits(
     }
 }
 
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) {
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
     const int n_ctx = llama_n_ctx(ctx);
@@ -269,10 +290,12 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
     }
 
     std::vector<float> logit_history;
-    logit_history.resize(tokens.size());
-
     std::vector<float> prob_history;
+
+    if (compute_ppl) {
+        logit_history.resize(tokens.size());
         prob_history.resize(tokens.size());
+    }
 
     const int n_chunk_max = tokens.size() / n_ctx;
 
@@ -288,12 +311,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
+    const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+
+    std::vector<float> logits;
+    if (compute_ppl && num_batches > 1) {
+        logits.reserve((size_t)n_ctx * n_vocab);
+    }
+
     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();
@@ -321,9 +349,11 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
             // restore the original token in case it was set to BOS
             tokens[batch_start] = token_org;
 
+            if (compute_ppl && num_batches > 1) {
                 const auto * batch_logits = llama_get_logits(ctx);
                 logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
             }
+        }
 
         const auto t_end = std::chrono::high_resolution_clock::now();
 
@@ -338,16 +368,22 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
+        if (compute_ppl) {
             const int first = n_ctx/2;
-        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+            const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+            process_logits(n_vocab, all_logits + 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;
 
             printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
             fflush(stdout);
+
+            logits.clear();
+        }
     }
     printf("\n");
 
+    if (compute_ppl) {
         nll2 /= count;
         nll /= count;
         const double ppl = exp(nll);
@@ -358,6 +394,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
         } else {
             printf("Unexpected negative standard deviation of log(prob)\n");
         }
+    }
 
     return true;
 }
@@ -365,6 +402,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
 
     StatParams sparams;
+    bool compute_ppl = true;
     std::vector<char*> args;
     args.push_back(argv[0]);
     int iarg = 1;
@@ -381,13 +419,22 @@ int main(int argc, char ** argv) {
         }
         else if (arg == "--verbosity") {
             sparams.verbosity = std::stoi(argv[++iarg]);
+        } else if (arg == "--no-ppl") {
+            compute_ppl = false;
+        } else if (arg == "--keep-imatrix") {
+            sparams.keep_every = std::stoi(argv[++iarg]);
         } else {
             args.push_back(argv[iarg]);
         }
     }
     if (iarg < argc) {
+        std::string arg{argv[iarg]};
+        if (arg == "--no-ppl") {
+            compute_ppl = false;
+        } else {
             args.push_back(argv[iarg]);
         }
+    }
 
     gpt_params params;
     params.n_batch = 512;
@@ -448,7 +495,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    bool OK = compute_imatrix(ctx, params);
+    bool OK = compute_imatrix(ctx, params, compute_ppl);
     if (!OK) {
         return 1;
     }

examples/infill/infill.cpp

@@ -241,7 +241,7 @@ int main(int argc, char ** argv) {
     LOG("add_bos: %d\n", add_bos);
 
     bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+    if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
         params.input_suffix.erase(0, 1);
         suff_rm_leading_spc = false;
     }

examples/llama-bench/llama-bench.cpp

@@ -562,6 +562,7 @@ struct test {
     static const int build_number;
     static const bool cuda;
     static const bool opencl;
+    static const bool vulkan;
     static const bool metal;
     static const bool gpu_blas;
     static const bool blas;
@@ -643,6 +644,9 @@ struct test {
         if (opencl) {
             return "OpenCL";
         }
+        if (vulkan) {
+            return "Vulkan";
+        }
         if (metal) {
             return "Metal";
         }
@@ -658,7 +662,7 @@ struct test {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "metal", "gpu_blas", "blas",
+            "cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "type_k", "type_v",
@@ -682,7 +686,7 @@ struct test {
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
+        if (field == "cuda" || field == "opencl"  || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
             field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
             return BOOL;
         }
@@ -710,7 +714,7 @@ struct test {
         }
         std::vector<std::string> values = {
             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),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             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), ggml_type_name(type_k), ggml_type_name(type_v),
@@ -738,6 +742,7 @@ const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
+const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();

examples/llama.android/app/build.gradle.kts

@@ -30,6 +30,7 @@ android {
         }
         externalNativeBuild {
             cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
                 cppFlags += listOf()
                 arguments += listOf()
             }

examples/llama.vim

@@ -6,7 +6,7 @@
 " Similarly, you could add an insert mode keybind with
 " inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
 "
-" g:llama_api_url and g:llama_overrides can be configured in your .vimrc
+" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc
 " let g:llama_api_url = "192.168.1.10:8080"
 " llama_overrides can also be set through buffer/window scopes. For instance
 " autocmd filetype python let b:llama_overrides = {"temp": 0.2}
@@ -82,6 +82,9 @@ func llama#doLlamaGen()
    endif
    let l:querydata.prompt = join(l:buflines, "\n")
    let l:curlcommand = copy(s:curlcommand)
+   if exists("g:llama_api_key")
+       call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key])
+   endif
    let l:curlcommand[2] = json_encode(l:querydata)
    let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
 endfunction

examples/llava/MobileVLM-README.md

@@ -0,0 +1,131 @@
+# MobileVLM
+
+Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants.
+
+for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM)
+
+The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
+
+## Usage
+Build with cmake or run `make llava-cli` to build it.
+
+After building, run: `./llava-cli` to see the usage. For example:
+
+```sh
+./llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
+    --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \
+    --image path/to/an/image.jpg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:"
+```
+
+## Model conversion
+
+- Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally:
+
+```sh
+git clone https://huggingface.co/mtgv/MobileVLM-1.7B
+
+git clone https://huggingface.co/openai/clip-vit-large-patch14-336
+```
+
+2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+
+```sh
+python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
+```
+
+3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF:
+
+```sh
+python ./examples/llava/convert-image-encoder-to-gguf \
+    -m path/to/clip-vit-large-patch14-336 \
+    --llava-projector path/to/MobileVLM-1.7B/llava.projector \
+    --output-dir path/to/MobileVLM-1.7B \
+    --projector-type ldp
+```
+
+4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+
+```sh
+python ./convert.py path/to/MobileVLM-1.7B
+```
+
+5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
+```sh
+./quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
+```
+
+Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.
+
+## Android compile and run
+### compile
+refer to `examples/llava/android/build_64.sh`
+```sh
+mkdir examples/llava/android/build_64
+cd examples/llava/android/build_64
+../build_64.sh
+```
+### run on Android
+refer to `android/adb_run.sh`, modify resources' `name` and `path`
+
+## some result on Android with `Snapdragon 888` chip
+### case 1
+**input**
+```sh
+/data/local/tmp/llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    -t 4 \
+    --image /data/local/tmp/demo.jpg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:"
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in 21148.71 ms by CLIP (  146.87 ms per image patch)
+ Susan Wise Bauer
+llama_print_timings:        load time =   23574.72 ms
+llama_print_timings:      sample time =       1.24 ms /     6 runs   (    0.21 ms per token,  4850.44 tokens per second)
+llama_print_timings: prompt eval time =   12460.15 ms /   246 tokens (   50.65 ms per token,    19.74 tokens per second)
+llama_print_timings:        eval time =     424.86 ms /     6 runs   (   70.81 ms per token,    14.12 tokens per second)
+llama_print_timings:       total time =   34731.93 ms
+```
+### case 2
+**input**
+```sh
+/data/local/tmp/llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    -t 4 \
+    --image /data/local/tmp/cat.jpeg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
+```
+
+**output**
+```sh
+encode_image_with_clip: image encoded in 21149.51 ms by CLIP (  146.87 ms per image patch)
+ The image depicts a cat sitting in the grass near some tall green plants.
+llama_print_timings:        load time =   23257.32 ms
+llama_print_timings:      sample time =       5.25 ms /    18 runs   (    0.29 ms per token,  3430.53 tokens per second)
+llama_print_timings: prompt eval time =   11900.73 ms /   232 tokens (   51.30 ms per token,    19.49 tokens per second)
+llama_print_timings:        eval time =    1279.03 ms /    18 runs   (   71.06 ms per token,    14.07 tokens per second)
+llama_print_timings:       total time =   34570.79 ms
+```
+
+## Minor shortcomings
+The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost.
+
+## TODO
+
+- [ ] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid`
+- [ ] Optimize LDP projector performance
+
+      - Optimize the structure definition to avoid unnecessary memory rearrangements, to reduce the use of `ggml_permute_cpy`;
+      - Optimize operator implementation (ARM CPU/NVIDIA GPU): such as depthwise conv, hardswish, hardsigmoid, etc.
+- [ ] run MobileVLM on `Jetson Orin`
+- [ ] Support more model variants, such as `MobileVLM-3B`.
+
+
+## contributor
+```sh
+zhangjidong05, yangyang260, huyiming03, chenxiaotao03
+```

examples/llava/android/adb_run.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+
+model_dir="/Users/cxt/model/llm/mobileVLM/MobileVLM-1.7B_processed"
+projector_name="mmproj-model-f16.gguf"
+llama_name="ggml-model-q4_k.gguf"
+img_dir="/Users/cxt/model/llm"
+img_name="demo.jpg"
+prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:"
+# img_name="cat.jpeg"
+# prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
+
+program_dir="build_64/bin"
+binName="llava-cli"
+n_threads=4
+
+
+deviceDir="/data/local/tmp"
+saveDir="output"
+if [ ! -d ${saveDir} ]; then
+    mkdir ${saveDir}
+fi
+
+
+function android_run() {
+    # # copy resource into device
+    # adb push ${model_dir}/${projector_name} ${deviceDir}/${projector_name}
+    # adb push ${model_dir}/${llama_name} ${deviceDir}/${llama_name}
+    adb push ${img_dir}/${img_name} ${deviceDir}/${img_name}
+    # copy program into device
+    adb push ${program_dir}/${binName} ${deviceDir}/${binName}
+    adb shell "chmod 0777 ${deviceDir}/${binName}"
+
+    # run
+    adb shell "echo cd ${deviceDir} ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 > ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt"
+    adb shell "cd ${deviceDir}; pwd; ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 >> ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt 2>&1"
+    adb pull ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt ${saveDir}
+}
+
+android_run
+
+echo "android_run is Done!"

examples/llava/android/build_64.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+cmake ../../../../ \
+-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+-DCMAKE_BUILD_TYPE=Release \
+-DANDROID_ABI="arm64-v8a" \
+-DANDROID_PLATFORM=android-23 $1
+
+make -j4

examples/llava/clip.cpp

@@ -2,17 +2,6 @@
 // so there might be still unnecessary artifacts hanging around
 // I'll gradually clean and extend it
 
-#include <cassert>
-#include <cmath>
-#include <cstdlib>
-#include <cstring>
-#include <fstream>
-#include <iostream>
-#include <map>
-#include <regex>
-#include <stdexcept>
-#include <vector>
-
 #include "clip.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
@@ -29,6 +18,19 @@
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
+#include <cassert>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <regex>
+#include <stdexcept>
+#include <vector>
+#include <sstream>
+#include <cinttypes>
+
 static std::string format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;
@@ -67,6 +69,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_PATCH_SIZE "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN "clip.vision.image_mean"
 #define KEY_IMAGE_STD "clip.vision.image_std"
+#define KEY_PROJ_TYPE "clip.projector_type"
 
 //
 // tensor name constants
@@ -89,6 +92,22 @@ static std::string format(const char * fmt, ...) {
 #define TN_TEXT_PROJ "text_projection.weight"
 #define TN_VIS_PROJ "visual_projection.weight"
 #define TN_LLAVA_PROJ "mm.%d.%s"
+#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP,           "mlp"     },
+    { PROJECTOR_TYPE_LDP,          "ldp"    },
+};
+
 
 //
 // utilities to get data from a gguf file
@@ -129,6 +148,91 @@ static std::string get_ftype(int ftype) {
     return ggml_type_name(static_cast<ggml_type>(ftype));
 }
 
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+
+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 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") {
+    size_t tensor_size = ggml_nbytes(tensor);
+    printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+            prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
+            tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
+}
+
+static projector_type clip_projector_type_from_string(const std::string & name) {
+    for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT
+        if (kv.second == name) {
+            return kv.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
 //
 // image data
 //
@@ -201,10 +305,44 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
-    struct ggml_tensor * mm_0_w;
+    struct ggml_tensor * mm_0_w = NULL;
-    struct ggml_tensor * mm_0_b;
+    struct ggml_tensor * mm_0_b = NULL;
-    struct ggml_tensor * mm_2_w;
+    struct ggml_tensor * mm_2_w = NULL;
-    struct ggml_tensor * mm_2_b;
+    struct ggml_tensor * mm_2_b = NULL;
+
+    // Yi type models with mlp+normalization projection
+    struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
+    struct ggml_tensor * mm_1_b = NULL;
+    struct ggml_tensor * mm_3_w = NULL;
+    struct ggml_tensor * mm_3_b = NULL;
+    struct ggml_tensor * mm_4_w = NULL;
+    struct ggml_tensor * mm_4_b = NULL;
+
+    // MobileVLM projection
+    struct ggml_tensor * mm_model_mlp_1_w;
+    struct ggml_tensor * mm_model_mlp_1_b;
+    struct ggml_tensor * mm_model_mlp_3_w;
+    struct ggml_tensor * mm_model_mlp_3_b;
+    struct ggml_tensor * mm_model_block_1_block_0_0_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_w;
+    struct ggml_tensor * mm_model_block_1_block_0_1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_1_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_1_block_2_0_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_w;
+    struct ggml_tensor * mm_model_block_1_block_2_1_b;
+    struct ggml_tensor * mm_model_block_2_block_0_0_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_w;
+    struct ggml_tensor * mm_model_block_2_block_0_1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc1_b;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_w;
+    struct ggml_tensor * mm_model_block_2_block_1_fc2_b;
+    struct ggml_tensor * mm_model_block_2_block_2_0_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_w;
+    struct ggml_tensor * mm_model_block_2_block_2_1_b;
 };
 
 struct clip_ctx {
@@ -213,6 +351,7 @@ struct clip_ctx {
     bool has_llava_projector = false;
 
     struct clip_vision_model vision_model;
+    projector_type proj_type = PROJECTOR_TYPE_MLP;
 
     float image_mean[3];
     float image_std[3];
@@ -330,6 +469,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     // pre-layernorm
     {
         embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "pre_ln");
 
         embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
@@ -430,9 +570,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             free(patches_data);
         }
 
+        // shape [1, 576, 1024]
+        // ne is whcn, ne = [1024, 576, 1, 1]
         embeddings = ggml_get_rows(ctx0, embeddings, patches);
 
-        // mm projection 0
+        // print_tensor_info(embeddings, "embeddings");
+
+        // llava projector
+        if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
             embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
 
@@ -440,6 +585,141 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+
+        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projector
+            int n_patch = 24;
+            struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings);
+            mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b);
+            mlp_1 = ggml_gelu(ctx0, mlp_1);
+            struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1);
+            mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b);
+            // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1]
+
+            // block 1
+            struct ggml_tensor * block_1 = nullptr;
+            {
+                // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
+                mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
+                mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                // stride = 1, padding = 1, bias is nullptr
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
+
+                // layer norm
+                // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+
+                // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+                // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+                // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1]
+                // residual
+                block_1 = ggml_add(ctx0, mlp_3, block_1);
+            }
+
+            // block_2
+            {
+                // stride = 2
+                block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1);
+
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // layer norm
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3));
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3));
+                // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1]
+                // hardswish
+                struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1);
+
+                // not sure the parameters is right for globalAvgPooling
+                block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0);
+                // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                // pointwise conv
+                block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b);
+                block_1 = ggml_relu(ctx0, block_1);
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1);
+                block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b);
+                block_1 = ggml_hardsigmoid(ctx0, block_1);
+
+                // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1]
+                block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]);
+                block_1 = ggml_mul(ctx0, block_1_hw, block_1);
+
+                int w = block_1->ne[0], h = block_1->ne[1];
+                block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]);
+                block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3));
+                // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1]
+                block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1);
+                block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]);
+
+
+                // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1]
+                block_1 = ggml_norm(ctx0, block_1, eps);
+                block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b);
+                block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]);
+                // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1]
+            }
+            embeddings = block_1;
+        }
+        else {
+            GGML_ASSERT(false);
+        }
     }
 
     // build the graph
@@ -485,16 +765,47 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         printf("\n");
     }
     const int n_tensors = gguf_get_n_tensors(ctx);
+
     // kv
-    if (verbosity >= 3) {
     const int n_kv = gguf_get_n_kv(ctx);
+    printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+        __func__, n_kv, n_tensors, fname);
+    {
+        std::map<enum ggml_type, uint32_t> n_type;
 
-        for (int i = 0; i < n_kv; ++i) {
+        for (int i = 0; i < n_tensors; i++) {
-            const char * key = gguf_get_key(ctx, i);
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
 
-            printf("%s: kv[%d]: key = %s\n", __func__, i, key);
+            n_type[type]++;
+        }
+
+        printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        for (int i = 0; i < n_kv; i++) {
+            const char * name           = gguf_get_key(ctx, i);
+            const enum gguf_type type   = gguf_get_kv_type(ctx, i);
+            const std::string type_name =
+                type == GGUF_TYPE_ARRAY
+                ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i))
+                : gguf_type_name(type);
+
+            std::string value          = gguf_kv_to_str(ctx, i);
+            const size_t MAX_VALUE_LEN = 40;
+            if (value.size() > MAX_VALUE_LEN) {
+                value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str());
+            }
+            replace_all(value, "\n", "\\n");
+
+            printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+        }
+
+        // print type counts
+        for (auto & kv : n_type) {
+            if (kv.second == 0) {
+                continue;
+            }
+
+            printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
         }
-        printf("\n");
     }
 
     // data
@@ -503,12 +814,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         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);
+            enum ggml_type type = gguf_get_tensor_type(ctx, i);
             struct ggml_tensor * cur = ggml_get_tensor(meta, name);
             size_t tensor_size = ggml_nbytes(cur);
             buffer_size += tensor_size;
             if (verbosity >= 3) {
-                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i,
+                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
-                       ggml_n_dims(cur), cur->name, tensor_size, offset);
+                       __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
             }
         }
     }
@@ -517,6 +829,23 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
     clip_ctx * new_clip = new clip_ctx;
 
+    // update projector type
+    {
+        int idx = gguf_find_key(ctx, KEY_PROJ_TYPE);
+        if (idx != -1) {
+            const std::string proj_type = gguf_get_val_str(ctx, idx);
+            new_clip->proj_type = clip_projector_type_from_string(proj_type);
+        }
+        else {
+            new_clip->proj_type = PROJECTOR_TYPE_MLP;
+        }
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+            if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
+                new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
+            }
+        }
+    }
+
 #ifdef GGML_USE_CUBLAS
     new_clip->backend = ggml_backend_cuda_init(0);
     printf("%s: CLIP using CUDA backend\n", __func__);
@@ -661,10 +990,63 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
         vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+
+        // LLaVA projection
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
             vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+            try {
+                // Yi-type llava
+                vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
+                vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // missing in Yi-type llava
                 vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
                 vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
+                vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
+                vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
+            } catch (std::runtime_error & e) {  }
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
+            // MobileVLM projection
+            vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
+            vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias"));
+            vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight"));
+            vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias"));
+            vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
+            vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
+            vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
+            vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight"));
+            vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias"));
+            vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight"));
+            vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias"));
+            vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
+            vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
+            vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
+            vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
+            vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
+            vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
+            vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight"));
+            vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias"));
+            vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight"));
+            vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias"));
+            vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
+            vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
+            vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
+        }
+        else {
+            std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
+            throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+        }
 
         vision_model.layers.resize(hparams.n_layer);
         for (int il = 0; il < hparams.n_layer; ++il) {
@@ -949,7 +1331,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         ".*weight",
     };
 
-    std::vector<uint8_t> read_data(512);
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
     std::vector<int64_t> hist_all(1 << 4, 0);
@@ -1100,13 +1481,27 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 }
 
 int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+        return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
+    }
+    else if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
+    } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        return ctx->vision_model.mm_3_b->ne[0];
+    }
+    else {
+        std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
+        throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
+    }
 }
 
 int clip_n_patches(const struct clip_ctx * ctx) {
     auto & params = ctx->vision_model.hparams;
-
+    int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
-    return (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+        n_patches /= 4;
+    }
+    return n_patches;
 }
 
 size_t clip_embd_nbytes(const struct clip_ctx * ctx) {

examples/llava/convert-image-encoder-to-gguf.py

@@ -81,6 +81,7 @@ ap.add_argument("--vision-only", action="store_true", required=False,
 ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
                 help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
 ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
 ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
@@ -174,6 +175,8 @@ elif args.vision_only and not has_llava_projector:
     fout.add_description("vision-only CLIP model")
 elif has_llava_projector:
     fout.add_description("image encoder for LLaVA")
+    # add projector type
+    fout.add_string("clip.projector_type", args.projector_type)
 else:
     fout.add_description("two-tower CLIP model")
 
@@ -218,7 +221,8 @@ if has_llava_projector:
     projector = torch.load(args.llava_projector)
     for name, data in projector.items():
         name = get_tensor_name(name)
-        if data.ndim == 2:
+        # pw and dw conv ndim==4
+        if data.ndim == 2 or data.ndim == 4:
             data = data.squeeze().numpy().astype(np.float16)
         else:
             data = data.squeeze().numpy().astype(np.float32)

examples/llava/llava-cli.cpp

@@ -148,10 +148,35 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
 
-    // llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
+    std::string system_prompt, user_prompt;
-    eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos);
+    size_t image_pos = prompt.find("<image>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
+        // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string
+        size_t pos = 0;
+        while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) {
+            user_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+        while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) {
+            system_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+
+        printf("system_prompt: %s\n", system_prompt.c_str());
+        printf("user_prompt: %s\n", user_prompt.c_str());
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
+        user_prompt = prompt + "\nASSISTANT:";
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
     llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
 
     // generate the response
 
@@ -162,6 +187,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
         if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
 
         printf("%s", tmp);
         fflush(stdout);

examples/perplexity/perplexity.cpp

@@ -112,6 +112,43 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to
     return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
 }
 
+static inline int nearest_int(float fval) {
+    //assert(fval <= 4194303.f);
+    float val = fval + 12582912.f;
+    int i; memcpy(&i, &val, sizeof(int));
+    return (i & 0x007fffff) - 0x00400000;
+}
+
+static double log_softmax(int n_vocab, const float * logits, uint16_t * log_prob, int tok) {
+    float max_logit = logits[0];
+    float min_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+        min_logit = std::min(min_logit, logits[i]);
+    }
+    min_logit = std::max(min_logit, max_logit - 16);
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    const float log_sum_exp = log(sum_exp);
+    const float min_log_prob = min_logit - max_logit - log_sum_exp;
+    const float scale = (max_logit - min_logit)/65535.f;
+    float * d = (float *)log_prob;
+    d[0] = scale;
+    d[1] = min_log_prob;
+    log_prob += 4;
+    if (scale) {
+        const float inv_scale = 1/scale;
+        for (int i = 0; i < n_vocab; ++i) {
+            log_prob[i] = logits[i] > min_logit ? nearest_int(inv_scale*(logits[i] - min_logit)) : 0;
+        }
+    } else {
+        std::memset(log_prob, 0, n_vocab*sizeof(uint16_t));
+    }
+    return max_logit + log_sum_exp - logits[tok];
+}
+
 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
@@ -147,6 +184,130 @@ static void process_logits(
     }
 }
 
+static void process_logits(std::ostream& out, int n_vocab, const float * logits, const int * tokens, int n_token,
+        std::vector<std::thread> & workers, std::vector<uint16_t> & log_probs, double & nll, double & nll2) {
+    std::mutex mutex;
+    const int nv = 2*((n_vocab + 1)/2) + 4;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &log_probs, &nll, &nll2, n_vocab, logits, tokens, n_token, nv] () {
+        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 double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]);
+            local_nll += v;
+            local_nll2 += v*v;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+    out.write((const char *)log_probs.data(), n_token*nv*sizeof(uint16_t));
+}
+
+struct kl_divergence_result {
+    double sum_nll  = 0;
+    double sum_nll2 = 0;
+    double sum_kld  = 0;
+    double sum_kld2 = 0;
+    double sum_nll_diff  = 0;
+    double sum_nll_diff2 = 0;
+    size_t n_same_top = 0;
+    size_t count = 0;
+};
+
+static double log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) {
+    float max_logit = logits[0];
+    int imax = 0;
+    for (int i = 1; i < n_vocab; ++i) {
+        if (logits[i] > max_logit) {
+            max_logit = logits[i];
+            imax = i;
+        }
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    const float log_sum_exp = log(sum_exp);
+    const float * d = (const float *)base_log_prob;
+    const float scale = d[0];
+    const float min_log_prob = d[1];
+    base_log_prob += 4;
+    float nll = max_logit + log_sum_exp - logits[tok];
+    kld.sum_nll  += nll;
+    kld.sum_nll2 += nll*nll;
+    nll += (scale*base_log_prob[tok] + min_log_prob);
+    kld.sum_nll_diff  += nll;
+    kld.sum_nll_diff2 += nll*nll;
+    max_logit += log_sum_exp;
+    double sum = 0;
+    int imax_base = -1;
+    float p_log_base_max = 0;
+    for (int i = 0; i < n_vocab; ++i) {
+        const float p_log_base = scale*base_log_prob[i] + min_log_prob;
+        if (i == 0 || p_log_base > p_log_base_max) {
+            p_log_base_max = p_log_base;
+            imax_base = i;
+        }
+        if (p_log_base > -16.f) {
+            const float p_base = expf(p_log_base);
+            sum += p_base * (p_log_base - logits[i] + max_logit);
+        }
+    }
+    kld.sum_kld  += sum;
+    kld.sum_kld2 += sum*sum;
+    ++kld.count;
+    if (imax == imax_base) ++kld.n_same_top;
+    return sum;
+}
+
+static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token,
+        std::vector<std::thread> & workers, const std::vector<uint16_t> & base_log_probs, kl_divergence_result & kld,
+        float * kld_values) {
+    std::mutex mutex;
+    const int nv = 2*((n_vocab + 1)/2) + 4;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv, kld_values] () {
+        kl_divergence_result local_kld;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                kld.sum_nll  += local_kld.sum_nll;
+                kld.sum_nll2 += local_kld.sum_nll2;
+                kld.sum_kld  += local_kld.sum_kld;
+                kld.sum_kld2 += local_kld.sum_kld2;
+                kld.sum_nll_diff  += local_kld.sum_nll_diff;
+                kld.sum_nll_diff2 += local_kld.sum_nll_diff2;
+                kld.n_same_top += local_kld.n_same_top;
+                kld.count += local_kld.count;
+                break;
+            }
+            lock.unlock();
+            double v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld);
+            kld_values[i] = (float)v;
+        }
+    };
+    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`
@@ -294,6 +455,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
     const int n_ctx = llama_n_ctx(ctx);
 
+    std::ofstream logits_stream;
+    if (!params.logits_file.empty()) {
+        logits_stream.open(params.logits_file.c_str());
+        if (!logits_stream.is_open()) {
+            fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str());
+            return {};
+        }
+        fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
+        logits_stream.write("_logits_", 8);
+        logits_stream.write((const char *)&n_ctx, sizeof(n_ctx));
+    }
+
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
@@ -336,6 +509,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
+    std::vector<uint16_t> log_probs;
+    if (!params.logits_file.empty()) {
+        logits_stream.write((const char *)&n_vocab, sizeof(n_vocab));
+        logits_stream.write((const char *)&n_chunk, sizeof(n_chunk));
+        logits_stream.write((const char *)tokens.data(), n_chunk*n_ctx*sizeof(tokens[0]));
+        const int nv = 2*((n_vocab + 1)/2) + 4;
+        log_probs.resize(n_ctx * nv);
+    }
+
     for (int i = 0; i < n_chunk; ++i) {
         const int start =     i * n_ctx;
         const int end   = start + n_ctx;
@@ -398,8 +580,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         // process the entire prompt.
         const int first = n_ctx/2;
         const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+        if (!params.logits_file.empty()) {
+            process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                    workers, log_probs, nll, nll2);
+        } else {
             process_logits(n_vocab, all_logits + 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;
 
         // perplexity is e^(average negative log-likelihood)
@@ -458,23 +645,24 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
     return true;
 }
 
+#define K_TOKEN_CHUNK 4
+
 static void compute_logprobs(const float * batch_logits, int n_vocab, std::vector<std::thread>& workers,
         const std::vector<std::pair<size_t, llama_token>>& eval_pairs, std::vector<float>& eval_results) {
-    constexpr int k_token_chunk = 4;
     if (eval_results.size() != eval_pairs.size()) {
         eval_results.resize(eval_pairs.size());
     }
     if (eval_pairs.empty()) return;
 
-    size_t max_threads = std::min((eval_pairs.size() + k_token_chunk - 1)/k_token_chunk, workers.size());
+    size_t max_threads = std::min((eval_pairs.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK, workers.size());
 
     std::atomic<int> counter(0);
     auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () {
-        float local_logprobs[k_token_chunk];
+        float local_logprobs[K_TOKEN_CHUNK];
         while (true) {
-            size_t first = counter.fetch_add(k_token_chunk, std::memory_order_relaxed);
+            size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed);
             if (first >= eval_results.size()) break;
-            size_t last = std::min(first + k_token_chunk, eval_results.size());
+            size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size());
             for (size_t i = first; i < last; ++i) {
                 auto logits = batch_logits + eval_pairs[i].first * n_vocab;
                 float max_logit = logits[0];
@@ -497,7 +685,6 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto
     for (size_t it = 0; it < max_threads; ++it) {
         workers[it].join();
     }
-
 }
 
 static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
@@ -540,14 +727,14 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     // This is needed as usual for LLaMA models
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
 
+    // The tasks should be randomized so the score stabilizes quickly.
+    bool randomize_tasks = true;
+
     // Number of tasks to use when computing the score
     if (params.hellaswag_tasks < hs_task_count) {
         hs_task_count = params.hellaswag_tasks;
     }
 
-    // The tasks should be randomized so the score stabilizes quickly.
-    bool randomize_tasks = true;
-
     // The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
     std::mt19937 rng(1);
 
@@ -1031,6 +1218,566 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma);
 }
 
+static bool deserialize_string(std::istream & in, std::string & str) {
+    uint32_t size;
+    if (!in.read((char *)&size, sizeof(size)).fail()) {
+        str.resize(size);
+        if (!in.read((char *)&str[0], size).fail()) return true;
+    }
+    return false;
+}
+
+struct multiple_choice_answers {
+    std::vector<std::string> answers;
+    std::vector<int>         labels;
+    bool deserialize(std::istream& in) {
+        uint32_t n;
+        in.read((char *)&n, sizeof(n));
+        if (in.fail() || n > 100) return false; // 100 as max. number of answers should be good enough for any practical purpose
+        answers.resize(n);
+        labels.resize(n);
+        for (auto& a : answers) {
+            if (!deserialize_string(in, a)) return false;
+        }
+        in.read((char *)labels.data(), n*sizeof(int));
+        return !in.fail();
+    }
+};
+
+struct multiple_choice_task {
+    std::string question;         // the question (or context that needs to be continued)
+    multiple_choice_answers mc1;  // possible answers (continuations) with a single correct answer
+    multiple_choice_answers mc2;  // possible answers (continuations) with multiple correct answers - not handled yet
+    bool deserialize(std::istream& in) {
+        if (!deserialize_string(in, question)) return false;
+        return mc1.deserialize(in) && mc2.deserialize(in);
+    }
+
+    // For evaluation
+    size_t i_batch;         // starting index in the llama_batch
+    size_t common_prefix;   // max number of initial tokens that are the same in all sentences
+    size_t required_tokens; // needed number of tokens to evaluate all answers
+    std::vector<std::vector<llama_token>> seq_tokens;
+    std::vector<float> log_probs;
+};
+
+static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos, multiple_choice_task& task, bool log_error) {
+    if (task.question.empty() || task.mc1.answers.empty()) {
+        if (log_error) {
+            printf("%s: found bad task with empty question and/or answers\n", __func__);
+        }
+        return false;
+    }
+    task.seq_tokens.reserve(task.mc1.answers.size());
+    for (auto& answer : task.mc1.answers) {
+        if (answer.empty()) {
+            if (log_error) {
+                printf("%s: found empty answer\n", __func__);
+            }
+            return false;
+        }
+        task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, add_bos));
+    }
+    auto min_len = task.seq_tokens.front().size();
+    for (auto& seq : task.seq_tokens) {
+        min_len = std::min(min_len, seq.size());
+    }
+    task.common_prefix = 0;
+    for (size_t k = 0; k < min_len; ++k) {
+        auto token = task.seq_tokens[0][k];
+        bool all_same = true;
+        for (size_t i = 1; i < task.seq_tokens.size(); ++i) {
+            if (task.seq_tokens[i][k] != token) {
+                all_same = false;
+                break;
+            }
+        }
+        if (!all_same) {
+            break;
+        }
+        ++task.common_prefix;
+    }
+    task.required_tokens = task.common_prefix;
+    for (auto& seq : task.seq_tokens) {
+        task.required_tokens += seq.size() - task.common_prefix;
+    }
+    return true;
+}
+
+//
+// Calculates score for multiple choice tasks with single correct answer from prompt.
+// Commonly used LLM evaluation metrics of this type are
+//   * ARC
+//   * HellaSwag
+//   * MMLU
+//   * TruthfulQA
+//
+// Validation datasets for these 4 tests can be found at
+//     https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp
+// The data for these datasets was extracted from
+//     git@hf.co:datasets/allenai/ai2_arc
+//     https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
+//     git@hf.co:datasets/Stevross/mmlu
+//     https://huggingface.co/datasets/truthful_qa
+//
+static void multiple_choice_score(llama_context * ctx, const gpt_params & params) {
+
+    std::istringstream strstream(params.prompt);
+    uint32_t n_task;
+    strstream.read((char *)&n_task, sizeof(n_task));
+    if (strstream.fail() || n_task == 0) {
+        printf("%s: no tasks\n", __func__);
+        return;
+    }
+    printf("%s: there are %u tasks in prompt\n", __func__, n_task);
+    std::vector<uint32_t> task_pos(n_task);
+    strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
+    if (strstream.fail()) {
+        printf("%s: failed to raad task positions from prompt\n", __func__);
+        return;
+    }
+
+    std::vector<multiple_choice_task> tasks;
+    if (params.multiple_choice_tasks == 0 || params.multiple_choice_tasks >= (size_t)n_task) {
+        // Use all tasks
+        tasks.resize(n_task);
+        printf("%s: reading tasks", __func__);
+        int n_dot = n_task/100;
+        int i = 0;
+        for (auto& task : tasks) {
+            ++i;
+            if (!task.deserialize(strstream)) {
+                printf("%s: failed to read task %d of %u\n", __func__, i, n_task);
+                return;
+            }
+            if (i%n_dot == 0) printf(".");
+        }
+        printf("done\n");
+    }
+    else {
+        printf("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task);
+        std::mt19937 rng(1);
+        std::vector<int> aux(n_task);
+        for (uint32_t i = 0; i < n_task; ++i) aux[i] = i;
+        float scale = 1.f/(1.f + (float)std::mt19937::max());
+        tasks.resize(params.multiple_choice_tasks);
+        for (auto& task : tasks) {
+            int j = (int)(scale * rng() * aux.size());
+            int idx = aux[j];
+            aux[j] = aux.back();
+            aux.pop_back();
+            strstream.seekg(task_pos[idx], std::ios::beg);
+            if (!task.deserialize(strstream)) {
+                printf("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]);
+                return;
+            }
+        }
+        n_task = params.multiple_choice_tasks;
+    }
+
+    // This is needed as usual for LLaMA models
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+
+    printf("%s: preparing task data", __func__);
+    fflush(stdout);
+    if (n_task > 500) {
+        printf("...");
+        fflush(stdout);
+        std::atomic<int> counter(0);
+        std::atomic<int> n_bad(0);
+        auto prepare = [&counter, &n_bad, &tasks, ctx, add_bos] () {
+            int num_tasks = tasks.size();
+            int n_bad_local = 0;
+            while (true) {
+                int first = counter.fetch_add(K_TOKEN_CHUNK);
+                if (first >= num_tasks) {
+                    if (n_bad_local > 0) n_bad += n_bad_local;
+                    break;
+                }
+                int last = std::min(first + K_TOKEN_CHUNK, num_tasks);
+                for (int i = first; i < last; ++i) {
+                    if (!multiple_choice_prepare_one_task(ctx, add_bos, tasks[i], false)) ++n_bad_local;
+                }
+            }
+        };
+        size_t max_thread = std::thread::hardware_concurrency();
+        max_thread = std::min(max_thread, (tasks.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK);
+        std::vector<std::thread> workers(max_thread-1);
+        for (auto& w : workers) w = std::thread(prepare);
+        prepare();
+        for (auto& w : workers) w.join();
+        printf("done\n");
+        fflush(stdout);
+        int nbad = n_bad;
+        if (nbad > 0) {
+            printf("%s: found %d malformed tasks\n", __func__, nbad);
+            return;
+        }
+    } else {
+        int n_dot = n_task/100;
+        int i_task = 0;
+        for (auto& task : tasks) {
+            ++i_task;
+            if (!multiple_choice_prepare_one_task(ctx, add_bos, task, true)) {
+                return;
+            }
+            if (i_task%n_dot == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+        printf("done\n");
+    }
+
+    printf("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size());
+
+    printf("\ntask\tacc_norm\n");
+
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_ctx   = llama_n_ctx(ctx);
+    const int n_batch = params.n_batch;
+
+    const int max_tasks_per_batch = 32;
+    const int max_seq = 4*max_tasks_per_batch;
+
+    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+
+    std::vector<float> tok_logits(n_vocab);
+    std::vector<float> batch_logits(n_vocab*n_ctx);
+
+    std::vector<std::pair<size_t, llama_token>> eval_pairs;
+    std::vector<float> eval_results;
+    std::vector<std::thread> workers(std::thread::hardware_concurrency());
+    std::vector<int> batch_indeces;
+
+    int n_done = 0;
+    int n_correct = 0;
+    int n_tot_answers = 0;
+
+    for (size_t i0 = 0; i0 < tasks.size(); i0++) {
+        int n_cur = 0;
+
+        size_t i1 = i0;
+        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+
+        llama_batch_clear(batch);
+
+        // batch as much tasks as possible into the available context
+        // each task has 4 unique seuqnce ids - one for each ending
+        // the common prefix is shared among the 4 sequences to save tokens
+        // we extract logits only from the last common token and from all ending tokens of each sequence
+        int s0 = 0;
+        while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
+            auto& cur_task = tasks[i1];
+
+            int num_answers = cur_task.seq_tokens.size();
+            if (s0 + num_answers > max_seq) {
+                break;
+            }
+
+            if (int(batch_indeces.size()) != num_answers) {
+                batch_indeces.resize(num_answers);
+            }
+            for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s;
+
+            for (size_t i = 0; i < cur_task.common_prefix; ++i) {
+                //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
+                llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
+            }
+            batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+
+            for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
+                for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) {
+                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true);
+                }
+            }
+
+            s0 += num_answers;
+
+            cur_task.i_batch = i_batch;
+            i_batch += cur_task.required_tokens;
+
+            n_cur += cur_task.required_tokens;
+            if (++i1 == tasks.size()) {
+                break;
+            }
+        }
+
+        if (i0 == i1) {
+            fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0);
+            return;
+        }
+
+        llama_kv_cache_clear(ctx);
+
+        // decode all tasks [i0, i1)
+        if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
+            fprintf(stderr, "%s: llama_decode() failed\n", __func__);
+            return;
+        }
+
+        // Compute log-probs in parallel
+        // First we collect all tasks
+        eval_pairs.clear();
+        for (size_t i = i0; i < i1; ++i) {
+            auto& cur_task = tasks[i];
+            size_t li = cur_task.common_prefix;
+            for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
+                for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
+                    eval_pairs.push_back(std::make_pair(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]));
+                }
+                ++li;
+            }
+        }
+        // Then we do the actual calculation
+        compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
+
+        size_t ir = 0;
+
+        // compute the logprobs for each ending of the decoded tasks
+        for (size_t i = i0; i < i1; ++i) {
+            auto & cur_task = tasks[i];
+            //printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str());
+            //for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) {
+            //    if (cur_task.mc1.labels[j] == 1) {
+            //        printf("%d", j+1);
+            //    }
+            //}
+            //printf("\n    common_prefix: %zu\n", cur_task.common_prefix);
+
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float));
+
+            const auto first_probs = softmax(tok_logits);
+
+            cur_task.log_probs.resize(cur_task.seq_tokens.size());
+            for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
+                size_t count = 1;
+                float  log_prob  = std::log(first_probs[cur_task.seq_tokens[s][cur_task.common_prefix]]);
+                for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
+                    //printf("        %zu  %g\n", ir, eval_results[ir]);
+                    ++count;
+                    log_prob += eval_results[ir++];
+                }
+                cur_task.log_probs[s] = log_prob / count;
+                //printf("        Final: %g\n", log_prob / count);
+                //printf("    <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count);
+            }
+
+            // Find the ending with maximum logprob
+            size_t logprob_max_idx = 0;
+            float  logprob_max_val = cur_task.log_probs[0];
+            for (size_t s = 1; s < cur_task.log_probs.size(); s++) {
+                if (cur_task.log_probs[s] > logprob_max_val) {
+                    logprob_max_val = cur_task.log_probs[s];
+                    logprob_max_idx = s;
+                }
+            }
+
+            n_tot_answers += cur_task.log_probs.size();
+            if (cur_task.mc1.labels[logprob_max_idx] == 1) {
+                ++n_correct;
+            }
+            ++n_done;
+
+            // Print the accumulated accuracy mean x 100
+            printf("%d\t%.8lf\n", n_done, 100.*n_correct/n_done);
+            fflush(stdout);
+        }
+
+        i0 = i1 - 1;
+    }
+
+    llama_batch_free(batch);
+
+    if (n_done < 100) return;
+
+    float p = 1.f*n_correct/n_done;
+    float sigma = sqrt(p*(1-p)/(n_done-1));
+    printf("\n Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
+    p = 1.f*n_done/n_tot_answers;
+    sigma = sqrt(p*(1-p)/(n_done-1));
+    printf("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
+
+    printf("\n");
+}
+
+static void kl_divergence(llama_context * ctx, const gpt_params & params) {
+    if (params.logits_file.empty()) {
+        fprintf(stderr, "%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
+        return;
+    }
+    std::ifstream in(params.logits_file.c_str(), std::ios::binary);
+    if (!in) {
+        fprintf(stderr, "%s: failed to open %s\n", __func__, params.logits_file.c_str());
+        return;
+    }
+    {
+        char check[9]; check[8] = 0;
+        in.read(check, 8);
+        if (in.fail() || strncmp("_logits_", check, 8) != 0) {
+            fprintf(stderr, "%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str());
+            return;
+        }
+    }
+
+    uint32_t n_ctx;
+    in.read((char *)&n_ctx, sizeof(n_ctx));
+    if (n_ctx > llama_n_ctx(ctx)) {
+        fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
+                __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
+    }
+
+    int n_vocab, n_chunk;
+    in.read((char *)&n_vocab, sizeof(n_vocab));
+    in.read((char *)&n_chunk, sizeof(n_chunk));
+    if (in.fail()) {
+        fprintf(stderr, "%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
+        return;
+    }
+    if (n_vocab != llama_n_vocab(llama_get_model(ctx))) {
+        fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
+    }
+
+    std::vector<llama_token> tokens(n_ctx * n_chunk);
+    if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
+        fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
+        return;
+    }
+
+    const int n_batch = params.n_batch;
+    const int num_batches = (n_ctx + n_batch - 1)/n_batch;
+    const int nv = 2*((n_vocab + 1)/2) + 4;
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+
+    std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
+    std::vector<float> kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
+    std::vector<float> logits;
+    if (num_batches > 1) {
+        logits.reserve(n_ctx * n_vocab);
+    }
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
+    auto mean_and_uncertainty = [] (double sum, double sum2, size_t count) {
+        if (count < 1) {
+            return std::make_pair(0., 0.);
+        }
+        double f = sum/count;
+        double df = sum2/count - f*f;
+        df = df > 0 && count > 10 ? sqrt(df/(count-1)) : 0.;
+        return std::make_pair(f, df);
+    };
+
+    kl_divergence_result kld;
+    auto kld_ptr = kld_values.data();
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) {
+            fprintf(stderr, "%s: failed reading log-probs for chunk %d\n", __func__, i);
+            return;
+        }
+
+        // clear the KV cache
+        llama_kv_cache_clear(ctx);
+
+        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);
+
+            // 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 (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+            }
+
+            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;
+            }
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            if (num_batches > 1) {
+                const auto * batch_logits = llama_get_logits(ctx);
+                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+            }
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+
+            printf("\nchunk        PPL          ln(PPL(Q)/PPL(base))          KL-Divergence           Same top\n");
+        }
+
+        const int first = n_ctx/2;
+        const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+        process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                workers, log_probs_uint16, kld, kld_ptr);
+        kld_ptr += n_ctx - 1 - first;
+
+        auto ppl           = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
+        auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count);
+        auto kl_div        = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
+        auto p_top = 1.*kld.n_same_top/kld.count;
+        auto d_p_top = sqrt(p_top*(1 - p_top)/(kld.count - 1));
+
+        printf("%4d    %10.4lf    %10.5lf ± %10.5f    %10.5f ± %10.5lf    %.5f ± %.5f\n", i+1, exp(ppl.first),
+                log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second,
+                p_top, d_p_top);
+
+        fflush(stdout);
+
+        logits.clear();
+    }
+    printf("\n");
+
+    if (kld.count < 100) return; // we do not wish to do statistics on so few values
+
+    std::sort(kld_values.begin(), kld_values.end());
+
+    printf("===== KL-divergence statistics\n");
+    auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
+    printf("Average: %10.6f ±%10.6lf\n", kl_div.first, kl_div.second);
+    auto kld_median = kld_values.size()%2 == 0 ? 0.5f*(kld_values[kld_values.size()/2] + kld_values[kld_values.size()/2-1])
+                                               : kld_values[kld_values.size()/2];
+    printf("Median : %10.6f\n", kld_median);
+
+    auto percentile = [&kld_values] (float fraction) {
+        if (fraction <= 0) return kld_values.front();
+        if (fraction >= 1) return kld_values.back();
+        float p = fraction*(kld_values.size() - 1);
+        size_t ip = size_t(p); p -= ip;
+        return (1 - p)*kld_values[ip] + p*kld_values[std::min(ip+1, kld_values.size()-1)];
+    };
+
+    printf("Maximum: %10.6f\n", kld_values.back());
+    printf("KLD_99 : %10.6f\n", percentile(0.99f));
+    printf("KLD_95 : %10.6f\n", percentile(0.95f));
+    printf("KLD_90 : %10.6f\n", percentile(0.90f));
+
+    printf("Minimum: %10.6f\n", kld_values.front());
+    printf("KLD_01 : %10.6f\n", percentile(0.01f));
+    printf("KLD_05 : %10.6f\n", percentile(0.05f));
+    printf("KLD_10 : %10.6f\n", percentile(0.10f));
+
+}
 
 int main(int argc, char ** argv) {
     gpt_params params;
@@ -1091,6 +1838,10 @@ int main(int argc, char ** argv) {
         hellaswag_score(ctx, params);
     } else if (params.winogrande) {
         winogrande_score(ctx, params);
+    } else if (params.multiple_choice) {
+        multiple_choice_score(ctx, params);
+    } else if (params.kl_divergence) {
+        kl_divergence(ctx, params);
     } else {
         results = perplexity(ctx, params);
     }

examples/pydantic-models-to-grammar-examples.py

@@ -1,14 +1,14 @@
 # Function calling example using pydantic models.
 import datetime
+import importlib
 import json
 from enum import Enum
-from typing import Union, Optional
+from typing import Optional, Union
 
 import requests
 from pydantic import BaseModel, Field
-
+from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model,
-import importlib
+                                        create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation)
-from pydantic_models_to_grammar import generate_gbnf_grammar_and_documentation, convert_dictionary_to_pydantic_model, add_run_method_to_dynamic_model, create_dynamic_model_from_function
 
 
 # Function to get completion on the llama.cpp server with grammar.
@@ -35,7 +35,7 @@ class SendMessageToUser(BaseModel):
         print(self.message)
 
 
-# Enum for the calculator function.
+# Enum for the calculator tool.
 class MathOperation(Enum):
     ADD = "add"
     SUBTRACT = "subtract"
@@ -43,7 +43,7 @@ class MathOperation(Enum):
     DIVIDE = "divide"
 
 
-# Very simple calculator tool for the agent.
+# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt.
 class Calculator(BaseModel):
     """
     Perform a math operation on two numbers.
@@ -148,37 +148,6 @@ def get_current_datetime(output_format: Optional[str] = None):
     return datetime.datetime.now().strftime(output_format)
 
 
-# Enum for the calculator tool.
-class MathOperation(Enum):
-    ADD = "add"
-    SUBTRACT = "subtract"
-    MULTIPLY = "multiply"
-    DIVIDE = "divide"
-
-
-
-# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt.
-class Calculator(BaseModel):
-    """
-    Perform a math operation on two numbers.
-    """
-    number_one: Union[int, float] = Field(..., description="First number.")
-    operation: MathOperation = Field(..., description="Math operation to perform.")
-    number_two: Union[int, float] = Field(..., description="Second number.")
-
-    def run(self):
-        if self.operation == MathOperation.ADD:
-            return self.number_one + self.number_two
-        elif self.operation == MathOperation.SUBTRACT:
-            return self.number_one - self.number_two
-        elif self.operation == MathOperation.MULTIPLY:
-            return self.number_one * self.number_two
-        elif self.operation == MathOperation.DIVIDE:
-            return self.number_one / self.number_two
-        else:
-            raise ValueError("Unknown operation.")
-
-
 # Example function to get the weather
 def get_current_weather(location, unit):
     """Get the current weather in a given location"""

examples/pydantic_models_to_grammar.py

@@ -1,15 +1,21 @@
+from __future__ import annotations
+
 import inspect
 import json
+import re
 from copy import copy
-from inspect import isclass, getdoc
+from enum import Enum
-from types import NoneType
+from inspect import getdoc, isclass
+from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints
 
 from docstring_parser import parse
-from pydantic import BaseModel, create_model, Field
+from pydantic import BaseModel, Field, create_model
-from typing import Any, Type, List, get_args, get_origin, Tuple, Union, Optional, _GenericAlias
+
-from enum import Enum
+if TYPE_CHECKING:
-from typing import get_type_hints, Callable
+    from types import GenericAlias
-import re
+else:
+    # python 3.8 compat
+    from typing import _GenericAlias as GenericAlias
 
 
 class PydanticDataType(Enum):
@@ -43,7 +49,7 @@ class PydanticDataType(Enum):
     SET = "set"
 
 
-def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str:
+def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str:
     if isclass(pydantic_type) and issubclass(pydantic_type, str):
         return PydanticDataType.STRING.value
     elif isclass(pydantic_type) and issubclass(pydantic_type, bool):
@@ -57,22 +63,22 @@ def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str:
 
     elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel):
         return format_model_and_field_name(pydantic_type.__name__)
-    elif get_origin(pydantic_type) == list:
+    elif get_origin(pydantic_type) is list:
         element_type = get_args(pydantic_type)[0]
         return f"{map_pydantic_type_to_gbnf(element_type)}-list"
-    elif get_origin(pydantic_type) == set:
+    elif get_origin(pydantic_type) is set:
         element_type = get_args(pydantic_type)[0]
         return f"{map_pydantic_type_to_gbnf(element_type)}-set"
-    elif get_origin(pydantic_type) == Union:
+    elif get_origin(pydantic_type) is Union:
         union_types = get_args(pydantic_type)
         union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types]
         return f"union-{'-or-'.join(union_rules)}"
-    elif get_origin(pydantic_type) == Optional:
+    elif get_origin(pydantic_type) is Optional:
         element_type = get_args(pydantic_type)[0]
         return f"optional-{map_pydantic_type_to_gbnf(element_type)}"
     elif isclass(pydantic_type):
         return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}"
-    elif get_origin(pydantic_type) == dict:
+    elif get_origin(pydantic_type) is dict:
         key_type, value_type = get_args(pydantic_type)
         return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}"
     else:
@@ -106,7 +112,6 @@ def get_members_structure(cls, rule_name):
         return f"{cls.__name__.lower()} ::= " + " | ".join(members)
     if cls.__annotations__ and cls.__annotations__ != {}:
         result = f'{rule_name} ::= "{{"'
-        type_list_rules = []
         # Modify this comprehension
         members = [
             f'  "\\"{name}\\"" ":"  {map_pydantic_type_to_gbnf(param_type)}'
@@ -116,17 +121,15 @@ def get_members_structure(cls, rule_name):
 
         result += '"," '.join(members)
         result += '  "}"'
-        return result, type_list_rules
+        return result
-    elif rule_name == "custom-class-any":
+    if rule_name == "custom-class-any":
         result = f"{rule_name} ::= "
         result += "value"
-        type_list_rules = []
+        return result
-        return result, type_list_rules
+
-    else:
     init_signature = inspect.signature(cls.__init__)
     parameters = init_signature.parameters
     result = f'{rule_name} ::=  "{{"'
-        type_list_rules = []
     # Modify this comprehension too
     members = [
         f'  "\\"{name}\\"" ":"  {map_pydantic_type_to_gbnf(param.annotation)}'
@@ -136,7 +139,7 @@ def get_members_structure(cls, rule_name):
 
     result += '", "'.join(members)
     result += '  "}"'
-        return result, type_list_rules
+    return result
 
 
 def regex_to_gbnf(regex_pattern: str) -> str:
@@ -269,7 +272,7 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None
 
 def generate_gbnf_rule_for_type(
     model_name, field_name, field_type, is_optional, processed_models, created_rules, field_info=None
-) -> Tuple[str, list]:
+) -> tuple[str, list[str]]:
     """
     Generate GBNF rule for a given field type.
 
@@ -283,7 +286,7 @@ def generate_gbnf_rule_for_type(
     :param field_info: Additional information about the field (optional).
 
     :return: Tuple containing the GBNF type and a list of additional rules.
-    :rtype: Tuple[str, list]
+    :rtype: tuple[str, list]
     """
     rules = []
 
@@ -321,8 +324,7 @@ def generate_gbnf_rule_for_type(
         gbnf_type, rules = model_name + "-" + field_name, rules
 
     elif gbnf_type.startswith("custom-class-"):
-        nested_model_rules, field_types = get_members_structure(field_type, gbnf_type)
+        rules.append(get_members_structure(field_type, gbnf_type))
-        rules.append(nested_model_rules)
     elif gbnf_type.startswith("custom-dict-"):
         key_type, value_type = get_args(field_type)
 
@@ -341,14 +343,14 @@ def generate_gbnf_rule_for_type(
         union_rules = []
 
         for union_type in union_types:
-            if isinstance(union_type, _GenericAlias):
+            if isinstance(union_type, GenericAlias):
                 union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type(
                     model_name, field_name, union_type, False, processed_models, created_rules
                 )
                 union_rules.append(union_gbnf_type)
                 rules.extend(union_rules_list)
 
-            elif not issubclass(union_type, NoneType):
+            elif not issubclass(union_type, type(None)):
                 union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type(
                     model_name, field_name, union_type, False, processed_models, created_rules
                 )
@@ -424,14 +426,10 @@ def generate_gbnf_rule_for_type(
     else:
         gbnf_type, rules = gbnf_type, []
 
-    if gbnf_type not in created_rules:
-        return gbnf_type, rules
-    else:
-        if gbnf_type in created_rules:
     return gbnf_type, rules
 
 
-def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created_rules: dict) -> (list, bool, bool):
+def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[BaseModel]], created_rules: dict[str, list[str]]) -> tuple[list[str], bool]:
     """
 
     Generate GBnF Grammar
@@ -452,7 +450,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created
     ```
     """
     if model in processed_models:
-        return []
+        return [], False
 
     processed_models.add(model)
     model_name = format_model_and_field_name(model.__name__)
@@ -518,7 +516,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created
 
 
 def generate_gbnf_grammar_from_pydantic_models(
-    models: List[Type[BaseModel]], outer_object_name: str = None, outer_object_content: str = None,
+    models: list[type[BaseModel]], outer_object_name: str | None = None, outer_object_content: str | None = None,
     list_of_outputs: bool = False
 ) -> str:
     """
@@ -528,7 +526,7 @@ def generate_gbnf_grammar_from_pydantic_models(
     * grammar.
 
     Args:
-        models (List[Type[BaseModel]]): A list of Pydantic models to generate the grammar from.
+        models (list[type[BaseModel]]): A list of Pydantic models to generate the grammar from.
         outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling.
         outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling.
         list_of_outputs (str, optional): Allows a list of output objects
@@ -543,9 +541,9 @@ def generate_gbnf_grammar_from_pydantic_models(
         # root ::= UserModel | PostModel
         # ...
     """
-    processed_models = set()
+    processed_models: set[type[BaseModel]] = set()
     all_rules = []
-    created_rules = {}
+    created_rules: dict[str, list[str]] = {}
     if outer_object_name is None:
         for model in models:
             model_rules, _ = generate_gbnf_grammar(model, processed_models, created_rules)
@@ -608,7 +606,7 @@ def get_primitive_grammar(grammar):
     Returns:
         str: GBNF primitive grammar string.
     """
-    type_list = []
+    type_list: list[type[object]] = []
     if "string-list" in grammar:
         type_list.append(str)
     if "boolean-list" in grammar:
@@ -666,14 +664,14 @@ triple-quotes ::= "'''" """
 
 
 def generate_markdown_documentation(
-    pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields",
+    pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields",
     documentation_with_field_description=True
 ) -> str:
     """
     Generate markdown documentation for a list of Pydantic models.
 
     Args:
-        pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes.
+        pydantic_models (list[type[BaseModel]]): list of Pydantic model classes.
         model_prefix (str): Prefix for the model section.
         fields_prefix (str): Prefix for the fields section.
         documentation_with_field_description (bool): Include field descriptions in the documentation.
@@ -731,7 +729,7 @@ def generate_markdown_documentation(
 
 
 def generate_field_markdown(
-    field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1,
+    field_name: str, field_type: type[Any], model: type[BaseModel], depth=1,
     documentation_with_field_description=True
 ) -> str:
     """
@@ -739,8 +737,8 @@ def generate_field_markdown(
 
     Args:
         field_name (str): Name of the field.
-        field_type (Type[Any]): Type of the field.
+        field_type (type[Any]): Type of the field.
-        model (Type[BaseModel]): Pydantic model class.
+        model (type[BaseModel]): Pydantic model class.
         depth (int): Indentation depth in the documentation.
         documentation_with_field_description (bool): Include field descriptions in the documentation.
 
@@ -798,7 +796,7 @@ def generate_field_markdown(
     return field_text
 
 
-def format_json_example(example: dict, depth: int) -> str:
+def format_json_example(example: dict[str, Any], depth: int) -> str:
     """
     Format a JSON example into a readable string with indentation.
 
@@ -819,14 +817,14 @@ def format_json_example(example: dict, depth: int) -> str:
 
 
 def generate_text_documentation(
-    pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields",
+    pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields",
     documentation_with_field_description=True
 ) -> str:
     """
     Generate text documentation for a list of Pydantic models.
 
     Args:
-        pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes.
+        pydantic_models (list[type[BaseModel]]): List of Pydantic model classes.
         model_prefix (str): Prefix for the model section.
         fields_prefix (str): Prefix for the fields section.
         documentation_with_field_description (bool): Include field descriptions in the documentation.
@@ -885,7 +883,7 @@ def generate_text_documentation(
 
 
 def generate_field_text(
-    field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1,
+    field_name: str, field_type: type[Any], model: type[BaseModel], depth=1,
     documentation_with_field_description=True
 ) -> str:
     """
@@ -893,8 +891,8 @@ def generate_field_text(
 
     Args:
         field_name (str): Name of the field.
-        field_type (Type[Any]): Type of the field.
+        field_type (type[Any]): Type of the field.
-        model (Type[BaseModel]): Pydantic model class.
+        model (type[BaseModel]): Pydantic model class.
         depth (int): Indentation depth in the documentation.
         documentation_with_field_description (bool): Include field descriptions in the documentation.
 
@@ -1017,8 +1015,8 @@ def generate_and_save_gbnf_grammar_and_documentation(
     pydantic_model_list,
     grammar_file_path="./generated_grammar.gbnf",
     documentation_file_path="./generated_grammar_documentation.md",
-    outer_object_name: str = None,
+    outer_object_name: str | None = None,
-    outer_object_content: str = None,
+    outer_object_content: str | None = None,
     model_prefix: str = "Output Model",
     fields_prefix: str = "Output Fields",
     list_of_outputs: bool = False,
@@ -1053,8 +1051,8 @@ def generate_and_save_gbnf_grammar_and_documentation(
 
 def generate_gbnf_grammar_and_documentation(
     pydantic_model_list,
-    outer_object_name: str = None,
+    outer_object_name: str | None = None,
-    outer_object_content: str = None,
+    outer_object_content: str | None = None,
     model_prefix: str = "Output Model",
     fields_prefix: str = "Output Fields",
     list_of_outputs: bool = False,
@@ -1086,9 +1084,9 @@ def generate_gbnf_grammar_and_documentation(
 
 
 def generate_gbnf_grammar_and_documentation_from_dictionaries(
-    dictionaries: List[dict],
+    dictionaries: list[dict[str, Any]],
-    outer_object_name: str = None,
+    outer_object_name: str | None = None,
-    outer_object_content: str = None,
+    outer_object_content: str | None = None,
     model_prefix: str = "Output Model",
     fields_prefix: str = "Output Fields",
     list_of_outputs: bool = False,
@@ -1098,7 +1096,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries(
     Generate GBNF grammar and documentation from a list of dictionaries.
 
     Args:
-        dictionaries (List[dict]): List of dictionaries representing Pydantic models.
+        dictionaries (list[dict]): List of dictionaries representing Pydantic models.
         outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling.
         outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling.
         model_prefix (str): Prefix for the model section in the documentation.
@@ -1120,7 +1118,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries(
     return grammar, documentation
 
 
-def create_dynamic_model_from_function(func: Callable):
+def create_dynamic_model_from_function(func: Callable[..., Any]):
     """
     Creates a dynamic Pydantic model from a given function's type hints and adds the function as a 'run' method.
 
@@ -1135,6 +1133,7 @@ def create_dynamic_model_from_function(func: Callable):
     sig = inspect.signature(func)
 
     # Parse the docstring
+    assert func.__doc__ is not None
     docstring = parse(func.__doc__)
 
     dynamic_fields = {}
@@ -1157,7 +1156,6 @@ def create_dynamic_model_from_function(func: Callable):
                 f"Parameter '{param.name}' in function '{func.__name__}' lacks a description in the docstring")
 
         # Add parameter details to the schema
-        param_doc = next((d for d in docstring.params if d.arg_name == param.name), None)
         param_docs.append((param.name, param_doc))
         if param.default == inspect.Parameter.empty:
             default_value = ...
@@ -1166,10 +1164,10 @@ def create_dynamic_model_from_function(func: Callable):
         dynamic_fields[param.name] = (
             param.annotation if param.annotation != inspect.Parameter.empty else str, default_value)
     # Creating the dynamic model
-    dynamic_model = create_model(f"{func.__name__}", **dynamic_fields)
+    dynamic_model = create_model(f"{func.__name__}", **dynamic_fields)  # type: ignore[call-overload]
 
-    for param_doc in param_docs:
+    for name, param_doc in param_docs:
-        dynamic_model.model_fields[param_doc[0]].description = param_doc[1].description
+        dynamic_model.model_fields[name].description = param_doc.description
 
     dynamic_model.__doc__ = docstring.short_description
 
@@ -1182,16 +1180,16 @@ def create_dynamic_model_from_function(func: Callable):
     return dynamic_model
 
 
-def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable):
+def add_run_method_to_dynamic_model(model: type[BaseModel], func: Callable[..., Any]):
     """
     Add a 'run' method to a dynamic Pydantic model, using the provided function.
 
     Args:
-        model (Type[BaseModel]): Dynamic Pydantic model class.
+        model (type[BaseModel]): Dynamic Pydantic model class.
         func (Callable): Function to be added as a 'run' method to the model.
 
     Returns:
-        Type[BaseModel]: Pydantic model class with the added 'run' method.
+        type[BaseModel]: Pydantic model class with the added 'run' method.
     """
 
     def run_method_wrapper(self):
@@ -1204,15 +1202,15 @@ def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable):
     return model
 
 
-def create_dynamic_models_from_dictionaries(dictionaries: List[dict]):
+def create_dynamic_models_from_dictionaries(dictionaries: list[dict[str, Any]]):
     """
     Create a list of dynamic Pydantic model classes from a list of dictionaries.
 
     Args:
-        dictionaries (List[dict]): List of dictionaries representing model structures.
+        dictionaries (list[dict]): List of dictionaries representing model structures.
 
     Returns:
-        List[Type[BaseModel]]: List of generated dynamic Pydantic model classes.
+        list[type[BaseModel]]: List of generated dynamic Pydantic model classes.
     """
     dynamic_models = []
     for func in dictionaries:
@@ -1249,7 +1247,7 @@ def list_to_enum(enum_name, values):
     return Enum(enum_name, {value: value for value in values})
 
 
-def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "CustomModel") -> Type[BaseModel]:
+def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: str = "CustomModel") -> type[Any]:
     """
     Convert a dictionary to a Pydantic model class.
 
@@ -1258,9 +1256,9 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu
         model_name (str): Name of the generated Pydantic model.
 
     Returns:
-        Type[BaseModel]: Generated Pydantic model class.
+        type[BaseModel]: Generated Pydantic model class.
     """
-    fields = {}
+    fields: dict[str, Any] = {}
 
     if "properties" in dictionary:
         for field_name, field_data in dictionary.get("properties", {}).items():
@@ -1277,7 +1275,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu
                     if items != {}:
                         array = {"properties": items}
                         array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items")
-                        fields[field_name] = (List[array_type], ...)
+                        fields[field_name] = (List[array_type], ...)  # type: ignore[valid-type]
                     else:
                         fields[field_name] = (list, ...)
                 elif field_type == "object":

examples/quantize/quantize.cpp

@@ -26,6 +26,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
+    { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
     { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },

examples/server/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

examples/server/README.md

@@ -30,7 +30,8 @@ Command line options:
 -   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
 -   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 -   `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
-
+-   `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
+-   `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
 ## Build
 
 server is build alongside everything else from the root of the project
@@ -65,6 +66,14 @@ server.exe -m models\7B\ggml-model.gguf -c 2048
 The above command will start a server that by default listens on `127.0.0.1:8080`.
 You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url.
 
+### Docker:
+```bash
+docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+
+# or, with CUDA:
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+```
+
 ## Testing with CURL
 
 Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.

examples/server/oai.hpp

@@ -0,0 +1,208 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+#include "utils.hpp"
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
+using json = nlohmann::json;
+
+inline static json oaicompat_completion_params_parse(
+    const json &body /* openai api json semantics */)
+{
+    json llama_params;
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
+    llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
+    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
+    llama_params["stream"]            = json_value(body, "stream", false);
+    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}

examples/server/utils.hpp

@@ -0,0 +1,508 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+
+#include "../llava/clip.h"
+
+using json = nlohmann::json;
+
+extern bool server_verbose;
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+#if SERVER_VERBOSE != 1
+#define LOG_VERBOSE(MSG, ...)
+#else
+#define LOG_VERBOSE(MSG, ...)                                            \
+    do                                                                   \
+    {                                                                    \
+        if (server_verbose)                                              \
+        {                                                                \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while (0)
+#endif
+
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+
+//
+// parallel
+//
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// TODO: can become bool if we can't find use of more states
+enum slot_state
+{
+    IDLE,
+    PROCESSING,
+};
+
+enum slot_command
+{
+    NONE,
+    LOAD_PROMPT,
+    RELEASE,
+};
+
+struct slot_params
+{
+    bool stream       = true;
+    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+
+    uint32_t seed      = -1; // RNG seed
+    int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t  n_predict = -1; // new tokens to predict
+
+    std::vector<std::string> antiprompt;
+
+    json input_prefix;
+    json input_suffix;
+};
+
+struct slot_image
+{
+    int32_t id;
+
+    bool request_encode_image = false;
+    float * image_embedding = nullptr;
+    int32_t image_tokens = 0;
+
+    clip_image_u8 * img_data;
+
+    std::string prefix_prompt; // before of this image
+};
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
+{
+    nlohmann::ordered_json log
+    {
+        {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
+    };
+
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+//
+// server utils
+//
+
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
+inline std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    return chatml_msgs.str();
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_all_task_finished;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        return id++;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when the batch of tasks is finished
+    void on_all_tasks_finished(std::function<void(void)> callback) {
+        callback_all_task_finished = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // Start the main loop. This call is blocking
+    [[noreturn]]
+    void start_loop() {
+        while (true) {
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("callback_all_task_finished", {});
+                // process and update all the multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is finished
+                callback_all_task_finished();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    condition_tasks.wait(lock, [&]{
+                        return !queue_tasks.empty();
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    void add_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    void remove_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+            LOG_VERBOSE("condition_results unblock", {});
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
+        }
+
+        // should never reach here
+    }
+
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
+    }
+
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
+
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {});
+                queue_results.push_back(result);
+                condition_results.notify_one();
+                return;
+            }
+        }
+    }
+};
+
+//
+// base64 utils (TODO: move to common in the future)
+//
+
+static const std::string base64_chars =
+             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+             "abcdefghijklmnopqrstuvwxyz"
+             "0123456789+/";
+
+static inline bool is_base64(uint8_t c)
+{
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
+    int i = 0;
+    int j = 0;
+    int in_ = 0;
+
+    int in_len = encoded_string.size();
+
+    uint8_t char_array_4[4];
+    uint8_t char_array_3[3];
+
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
+        char_array_4[i++] = encoded_string[in_]; in_++;
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
+                char_array_4[i] = base64_chars.find(char_array_4[i]);
+            }
+
+            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+            for (i = 0; (i < 3); i++)
+            {
+                ret.push_back(char_array_3[i]);
+            }
+            i = 0;
+        }
+    }
+
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
+            char_array_4[j] = 0;
+        }
+
+        for (j = 0; j <4; j++)
+        {
+            char_array_4[j] = base64_chars.find(char_array_4[j]);
+        }
+
+        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+        for (j = 0; (j < i - 1); j++)
+        {
+            ret.push_back(char_array_3[j]);
+        }
+    }
+
+    return ret;
+}
+
+//
+// random string / id
+//
+
+static std::string random_string()
+{
+    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
+
+    std::random_device rd;
+    std::mt19937 generator(rd());
+
+    std::string result(32, ' ');
+
+    for (int i = 0; i < 32; ++i) {
+        result[i] = str[generator() % str.size()];
+    }
+
+    return result;
+}
+
+static std::string gen_chatcmplid()
+{
+    std::stringstream chatcmplid;
+    chatcmplid << "chatcmpl-" << random_string();
+    return chatcmplid.str();
+}

examples/sycl/CMakeLists.txt

@@ -0,0 +1,9 @@
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+set(TARGET ls-sycl-device)
+add_executable(${TARGET} ls-sycl-device.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/README.md

@@ -0,0 +1,47 @@
+# llama.cpp/example/sycl
+
+This example program provide the tools for llama.cpp for SYCL on Intel GPU.
+
+## Tool
+
+|Tool Name| Function|Status|
+|-|-|-|
+|ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
+
+### ls-sycl-device
+
+List all SYCL devices with ID, compute capability, max work group size, ect.
+
+1. Build the llama.cpp for SYCL for all targets.
+
+2. Enable oneAPI running environment
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. Execute
+
+```
+./build/bin/ls-sycl-device
+```
+
+Check the ID in startup log, like:
+
+```
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
+```
+
+|Attribute|Note|
+|-|-|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|

examples/sycl/build.sh

@@ -0,0 +1,20 @@
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main only
+#cmake --build . --config Release --target main
+
+#build all binary
+cmake --build . --config Release -v

examples/sycl/ls-sycl-device.cpp

@@ -0,0 +1,11 @@
+/*MIT license
+  Copyright (C) 2024 Intel Corporation
+  SPDX-License-Identifier: MIT
+*/
+
+#include "ggml-sycl.h"
+
+int main(int argc, char ** argv) {
+    ggml_backend_sycl_print_sycl_devices();
+    return 0;
+}

examples/sycl/run-llama2.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+source /opt/intel/oneapi/setvars.sh
+
+if [ $# -gt 0 ]; then
+    export GGML_SYCL_DEVICE=$1
+else
+    export GGML_SYCL_DEVICE=0
+fi
+echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
+#export GGML_SYCL_DEBUG=1
+./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+#./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 5 -e -ngl 33 -t 1 -s 0
+

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1705133751,
+        "lastModified": 1706191920,
-        "narHash": "sha256-rCIsyE80jgiOU78gCWN3A0wE0tR2GI5nH6MlS+HaaSQ=",
+        "narHash": "sha256-eLihrZAPZX0R6RyM5fYAWeKVNuQPYjAkCUBr+JNvtdE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9b19f5e77dd906cb52dade0b7bd280339d2a1f3d",
+        "rev": "ae5c332cbb5827f6b1f02572496b141021de335f",
         "type": "github"
       },
       "original": {

flake.nix

@@ -1,3 +1,17 @@
+# The flake interface to llama.cpp's Nix expressions. The flake is used as a
+# more discoverable entry-point, as well as a way to pin the dependencies and
+# expose default outputs, including the outputs built by the CI.
+
+# For more serious applications involving some kind of customization  you may
+# want to consider consuming the overlay, or instantiating `llamaPackages`
+# directly:
+#
+# ```nix
+# pkgs.callPackage ${llama-cpp-root}/.devops/nix/scope.nix { }`
+# ```
+
+# Cf. https://jade.fyi/blog/flakes-arent-real/ for a more detailed exposition
+# of the relation between Nix and the Nix Flakes.
 {
   description = "Port of Facebook's LLaMA model in C/C++";
 

ggml-alloc.c

@@ -109,8 +109,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
         } else {
-            fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+            fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, largest block available %zu)\n",
-                    __func__, size, max_avail);
+                    __func__, tensor->name, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
             return;
         }
@@ -335,7 +335,9 @@ bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) {
 }
 
 size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) {
-    return alloc->max_size;
+    // FIXME: changes in the tensor sizes compared to the measure graph may cause allocations to fail
+    // to avoid this, we add a 10% margin to the buffer size
+    return alloc->max_size + alloc->max_size/10;
 }
 
 // graph allocator
@@ -776,38 +778,26 @@ size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph)
 }
 
 // utils
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
-    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
 
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
+static bool alloc_tensor_range(struct ggml_context * ctx,
-
+        struct ggml_tensor * first, struct ggml_tensor * last,
-    size_t nbytes = 0;
+        ggml_backend_buffer_type_t buft, size_t size,
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
-        if (t->data == NULL && t->view_src == NULL) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
-            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
-        }
-    }
-
-    if (nbytes == 0) {
-        // all the tensors in the context are already allocated
-#ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
-#endif
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
     if (buffer == NULL) {
-        // failed to allocate buffer
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
-        return NULL;
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free(*buffers[i]);
+        }
+        free(buffers);
+        return false;
     }
 
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
         if (t->data == NULL) {
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(tallocr, t);
@@ -824,6 +814,76 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     ggml_tallocr_free(tallocr);
 
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            // tensor is too large to fit in a single buffer
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+        // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
     return buffer;
 }
 

ggml-backend-impl.h

@@ -19,6 +19,7 @@ extern "C" {
         const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
         size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
         // check if tensor data is in host memory
@@ -63,6 +64,11 @@ extern "C" {
     // do not use directly, use ggml_backend_tensor_copy instead
     bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
     //
     // Backend
     //

ggml-backend.c

@@ -27,10 +27,20 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_alignment(buft);
 }
 
+size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    // get_max_size is optional, defaults to SIZE_MAX
+    if (buft->iface.get_max_size) {
+        return buft->iface.get_max_size(buft);
+    }
+    return SIZE_MAX;
+}
+
 GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
-        return buft->iface.get_alloc_size(buft, tensor);
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
     }
     return ggml_nbytes(tensor);
 }
@@ -55,8 +65,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
                size_t                          size) {
     ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
 
-    GGML_ASSERT(iface.get_base != NULL);
-
     (*buffer) = (struct ggml_backend_buffer) {
         /* .interface = */ iface,
         /* .buft      = */ buft,
@@ -106,6 +114,10 @@ size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
 }
 
+size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
+}
+
 size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }
@@ -120,6 +132,11 @@ bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
 
 void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
     buffer->usage = usage;
+
+    // FIXME: add a generic callback to the buffer interface
+    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
+        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    }
 }
 
 ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
@@ -169,6 +186,10 @@ size_t ggml_backend_get_alignment(ggml_backend_t backend) {
     return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
 }
 
+size_t ggml_backend_get_max_size(ggml_backend_t backend) {
+    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
+}
+
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
@@ -337,11 +358,21 @@ GGML_CALL static void ggml_backend_registry_init(void) {
     ggml_backend_cuda_reg_devices();
 #endif
 
+#ifdef GGML_USE_SYCL
+    extern void ggml_backend_sycl_reg_devices(void);
+    ggml_backend_sycl_reg_devices();
+#endif
+
 #ifdef GGML_USE_METAL
     extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
     extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
     ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
 #endif
+
+#ifdef GGML_USE_VULKAN
+    extern GGML_CALL int ggml_backend_vk_reg_devices(void);
+    ggml_backend_vk_reg_devices();
+#endif
 }
 
 GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
@@ -545,6 +576,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
@@ -600,6 +632,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
             /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
@@ -756,6 +789,80 @@ GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, v
     GGML_UNUSED(user_data);
 }
 
+// multi-buffer buffer
+
+struct ggml_backend_multi_buffer_context {
+    ggml_backend_buffer_t * buffers;
+    size_t n_buffers;
+};
+
+typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
+
+GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_free(ctx->buffers[i]);
+    }
+
+    free(ctx->buffers);
+    free(ctx);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_clear(ctx->buffers[i], value);
+    }
+}
+
+static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
+    static struct ggml_backend_buffer_i multi_backend_buffer_i = {
+        /* .get_name        = */ ggml_backend_multi_buffer_get_name,
+        /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+        /* .get_base        = */ NULL,
+        /* .init_tensor     = */ NULL,
+        /* .set_tensor      = */ NULL,
+        /* .get_tensor      = */ NULL,
+        /* .cpy_tensor      = */ NULL,
+        /* .clear           = */ ggml_backend_multi_buffer_clear,
+        /* .reset           = */ NULL,
+    };
+
+    return multi_backend_buffer_i;
+}
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+    ctx->n_buffers = n_buffers;
+    ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
+
+    size_t total_size = 0;
+    for (size_t i = 0; i < n_buffers; i++) {
+        ctx->buffers[i] = buffers[i];
+        total_size += ggml_backend_buffer_get_size(buffers[i]);
+    }
+
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
+}
+
+GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+}
+
+GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
+    }
+}
+
 
 // scheduler
 
@@ -1191,6 +1298,24 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
                 ggml_tallocr_t src_allocr = node_allocr(src);
                 GGML_ASSERT(src_allocr != NULL); // all inputs should be assigned by now
                 if (src_allocr != node_allocr) {
+                    // create a copy of the input in the split's backend
+                    size_t id = hash_id(src);
+                    if (sched->node_copies[id][cur_backend_id] == NULL) {
+                        ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
+                        struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                        ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
+
+                        sched->node_copies[id][cur_backend_id] = tensor_copy;
+                        node_allocr(tensor_copy) = cur_allocr;
+                        SET_CAUSE(tensor_copy, "4.cpy");
+
+                        int n_inputs = sched->splits[cur_split].n_inputs++;
+                        GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                        sched->splits[cur_split].inputs[n_inputs] = src;
+                    }
+                    node->src[j] = sched->node_copies[id][cur_backend_id];
+
+#if 0
                     // check if the input is already in the split
                     bool found = false;
                     for (int k = 0; k < sched->splits[cur_split].n_inputs; k++) {
@@ -1206,19 +1331,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g
                         GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
                         sched->splits[cur_split].inputs[n_inputs] = src;
                     }
-
+#endif
-                    // create a copy of the input in the split's backend
-                    size_t id = hash_id(src);
-                    if (sched->node_copies[id][cur_backend_id] == NULL) {
-                        ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
-                        struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                        ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-
-                        sched->node_copies[id][cur_backend_id] = tensor_copy;
-                        node_allocr(tensor_copy) = cur_allocr;
-                        SET_CAUSE(tensor_copy, "4.cpy");
-                    }
-                    node->src[j] = sched->node_copies[id][cur_backend_id];
                 }
             }
         }

ggml-backend.h

@@ -20,6 +20,7 @@ extern "C" {
     GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
     GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
     GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
     GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
     GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
     GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
@@ -36,6 +37,7 @@ extern "C" {
     GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
     GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
     GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
     GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
@@ -54,6 +56,7 @@ extern "C" {
     GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
     GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
 
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

ggml-cuda.cu

@@ -13,6 +13,10 @@
 #include <map>
 #include <array>
 
+// stringize macro for converting __CUDA_ARCH_LIST__ (list of integers) to string
+#define STRINGIZE_IMPL(...) #__VA_ARGS__
+#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
+
 #if defined(GGML_USE_HIPBLAS)
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
@@ -585,13 +589,28 @@ static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0,
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
 [[noreturn]]
-static __device__ void bad_arch() {
+static __device__ void no_device_code(
-    printf("ERROR: ggml-cuda was compiled without support for the current GPU architecture.\n");
+    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
+           file_name, line, function_name, arch);
+    (void) arch_list;
+#else
+    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
+           file_name, line, function_name, arch, arch_list);
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
     __trap();
 
-    (void) bad_arch; // suppress unused function warning
+    (void) no_device_code; // suppress unused function warning
 }
 
+#ifdef __CUDA_ARCH__
+#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
+#else
+#define NO_DEVICE_CODE GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#endif // __CUDA_ARCH__
+
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
@@ -618,7 +637,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
     return a;
 #else
     (void) a;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 }
 
@@ -647,7 +666,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
     return x;
 #else
     (void) x;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
 }
 
@@ -2444,7 +2463,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
     }
 #else
     (void) vx; (void) y; (void) k;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_PASCAL
 }
 
@@ -2475,7 +2494,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp
     // second part effectively subtracts 8 from each quant value
     return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2512,7 +2531,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
     // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
     return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2547,7 +2566,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp
     // second part effectively subtracts 16 from each quant value
     return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2592,7 +2611,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
     return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2613,7 +2632,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_imp
 
     return d8_0*d8_1 * sumi;
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2643,7 +2662,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
     // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
     return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2678,7 +2697,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
 
     return dm2f.x*sumf_d - dm2f.y*sumf_m;
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2715,7 +2734,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
 
     return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2755,7 +2774,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
 
     return d3 * sumf;
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2780,7 +2799,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
 
     return d3*d8 * sumi;
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2813,7 +2832,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2846,7 +2865,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2886,7 +2905,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
     return dm5f.x*sumf_d - dm5f.y*sumf_m;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2919,7 +2938,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
     return dm4f.x*sumf_d - dm4f.y*sumf_m;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2949,7 +2968,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
 
     return d*sumf;
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -2980,7 +2999,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 }
 
@@ -3846,7 +3865,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     return dall * sumf_d - dmin * sumf_m;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
 #endif
@@ -4029,7 +4048,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     return d * sumf_d;
 
 #else
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= MIN_CC_DP4A
 
 #endif
@@ -4287,7 +4306,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
         q8 += 8;
         aux32 >>= 7;
     }
-    const float d = (float)bq2->d * (0.5f + aux32) * (float)bq8_1[ib32].ds.x * 0.25f;
+    const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f;
     return d * sumi;
 #else
     // iqs is 0...15
@@ -4298,7 +4317,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     const uint8_t  * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
     const uint8_t  * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
     const uint32_t aux32 = q2[2] | (q2[3] << 16);
-    const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * (float)bq8_1[ib32].ds.x * 0.25f;
+    const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f;
     const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127];
     const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127];
     const int8_t * q8 = bq8_1[ib32].qs + 16*il;
@@ -4343,7 +4362,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
         }
         q8 += 8;
     }
-    const float d = (float)bq2->d * (float)bq8_1[ib32].ds.x * 0.25f;
+    const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
 #else
     assert(false);
@@ -4524,7 +4543,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_0_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4593,7 +4612,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_1_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4660,7 +4679,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_0_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4727,7 +4746,7 @@ mul_mat_q5_1(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_1_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4794,7 +4813,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q8_0_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4861,7 +4880,7 @@ mul_mat_q2_K(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q2_K_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4930,7 +4949,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q3_K_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -4999,7 +5018,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q4_K_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -5066,7 +5085,7 @@ mul_mat_q5_K(
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q5_K_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -5135,7 +5154,7 @@ template <bool need_check> static __global__ void
         (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
 #else
     (void) vec_dot_q6_K_q8_1_mul_mat;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
@@ -5858,7 +5877,7 @@ static __global__ void soft_max_f16(const float * x, const float * y, float * ds
     }
 #else
     (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale;
-    bad_arch();
+    NO_DEVICE_CODE;
 #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
 }
 
@@ -10225,8 +10244,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     // TODO: mmq/mmv support
 #endif
 
-    const int64_t nb11 = src1->nb[1];
+    const size_t nb11 = src1->nb[1];
-    const int64_t nb1  =  dst->nb[1];
+    const size_t nb1  =  dst->nb[1];
 
     const struct ggml_tensor * ids = src0;
     const int32_t id = ((int32_t *) dst->op_params)[0];
@@ -10887,15 +10906,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         // initialize padding to 0 to avoid possible NaN values
-        int64_t row_low = 0;
+        size_t original_size = ggml_nbytes(tensor);
-        int64_t row_high = ggml_nrows(tensor);
-        int64_t nrows_split = row_high - row_low;
-
-        size_t original_size = ggml_nbytes_split(tensor, nrows_split);
         size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
         if (padded_size > original_size && tensor->view_src == nullptr) {
-            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
+            CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
         }
     }
 }
@@ -10998,12 +11013,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend
 }
 
 GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
-    int64_t row_low = 0;
+    size_t size = ggml_nbytes(tensor);
-    int64_t row_high = ggml_nrows(tensor);
-    int64_t nrows_split = row_high - row_low;
-
-    size_t size = ggml_nbytes_split(tensor, nrows_split);
-
     int64_t ne0 = tensor->ne[0];
 
     if (ggml_is_quantized(tensor->type)) {
@@ -11032,6 +11042,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cuda_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
     /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
@@ -11307,6 +11318,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface
     /* .get_name         = */ ggml_backend_cuda_split_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
     /* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend,
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
@@ -11386,6 +11398,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .get_name         = */ ggml_backend_cuda_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,

ggml-metal.m

@@ -24,19 +24,7 @@
 
 #define UNUSED(x) (void)(x)
 
-#define GGML_METAL_MAX_KERNELS 256
-
-struct ggml_metal_buffer {
-    const char * name;
-
-    void   * data;
-    size_t   size;
-
-    id<MTLBuffer> metal;
-};
-
 struct ggml_metal_kernel {
-    id<MTLFunction>             function;
     id<MTLComputePipelineState> pipeline;
 };
 
@@ -149,7 +137,10 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@@ -171,14 +162,10 @@ struct ggml_metal_context {
 
     id<MTLDevice>       device;
     id<MTLCommandQueue> queue;
-    id<MTLLibrary>      library;
 
     dispatch_queue_t d_queue;
 
-    int n_buffers;
+    struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT];
-    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-
-    struct ggml_metal_kernel kernels[GGML_METAL_MAX_KERNELS];
 
     bool support_simdgroup_reduction;
     bool support_simdgroup_mm;
@@ -245,26 +232,24 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     // Show all the Metal device instances in the system
     NSArray * devices = MTLCopyAllDevices();
     for (id<MTLDevice> device in devices) {
-        NSString * s = [device name];
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
-        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
     }
     [devices release]; // since it was created by a *Copy* C method
 #endif
 
     // Pick and show default Metal device
     id<MTLDevice> device = MTLCreateSystemDefaultDevice();
-    NSString * s = [device name];
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
-    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->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
-    ctx->n_buffers = 0;
-
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
+    id<MTLLibrary> metal_library;
+
     // load library
     {
         NSBundle * bundle = nil;
@@ -279,7 +264,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:libPath];
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
-            ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+            metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
@@ -321,7 +306,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
                 //[options setFastMathEnabled:false];
 
-                ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+                metal_library = [ctx->device newLibraryWithSource:src options:options error:&error];
                 if (error) {
                     GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                     return NULL;
@@ -386,8 +371,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     {
         NSError * error = nil;
 
-        for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) {
+        for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
-            ctx->kernels[i].function = nil;
             ctx->kernels[i].pipeline = nil;
         }
 
@@ -399,13 +383,15 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(e, name, supported) \
         if (supported) { \
             struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
-            kernel->function = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+            id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
-            kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:kernel->function error:&error]; \
+            kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
+            [metal_function release]; \
             GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
                     (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
                     (int) kernel->pipeline.threadExecutionWidth); \
             if (error) { \
                 GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+                [metal_library release]; \
                 return NULL; \
             } \
         } else { \
@@ -522,7 +508,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,    flash_attn_ext_f16_h64,  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,    flash_attn_ext_f16_h80,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,    flash_attn_ext_f16_h96,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,   flash_attn_ext_f16_h112, true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,   flash_attn_ext_f16_h128, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,   flash_attn_ext_f16_h256, true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,            true);
@@ -537,27 +526,17 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,                true);
     }
 
+    [metal_library release];
     return ctx;
 }
 
 static void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
-    for (int i = 0; i < ctx->n_buffers; ++i) {
+    for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
-        [ctx->buffers[i].metal release];
-    }
-
-    for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) {
-        if (ctx->kernels[i].pipeline) {
         [ctx->kernels[i].pipeline release];
     }
 
-        if (ctx->kernels[i].function) {
-            [ctx->kernels[i].function release];
-        }
-    }
-
-    [ctx->library release];
     [ctx->queue release];
     [ctx->device release];
 
@@ -589,15 +568,13 @@ struct ggml_backend_metal_buffer_context {
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // 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) {
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
     //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);
 
     ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
 
-    // compatibility with ggml-backend
-    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
     struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
 
     // find the view that contains the tensor fully
@@ -619,25 +596,6 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
     return nil;
 }
 
-    // find the view that contains the tensor fully
-    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;
-
-            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
-
-            return ctx->buffers[i].metal;
-        }
-    }
-
-    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
-
-    return nil;
-}
-
 static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_UNARY:
@@ -681,7 +639,8 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
             return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
-            return ctx->support_simdgroup_reduction;
+            return ctx->support_simdgroup_reduction &&
+                (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
@@ -826,9 +785,9 @@ static bool ggml_metal_graph_compute(
             const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
             const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
 
-            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
-            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
-            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
             //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
             //if (src0) {
@@ -1610,7 +1569,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
                             }
@@ -1755,7 +1714,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
                             }
@@ -2189,9 +2148,9 @@ static bool ggml_metal_graph_compute(
                         size_t offs_src3 = 0;
 
                         GGML_ASSERT(src2);
-                        id<MTLBuffer> id_src2 = ggml_metal_get_buffer(ctx, src2, &offs_src2);
+                        id<MTLBuffer> id_src2 = ggml_metal_get_buffer(src2, &offs_src2);
 
-                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(ctx, src3, &offs_src3) : nil;
+                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
 
                         const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
                         const int64_t  ne31 = src3 ? src3->ne[1] : 0;
@@ -2213,7 +2172,10 @@ static bool ggml_metal_graph_compute(
                         switch (ne00) {
                             case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
                             case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
+                            case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
+                            case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
                             case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
+                            case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
                             default:
                                 {
                                     GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
@@ -2253,14 +2215,17 @@ static bool ggml_metal_graph_compute(
                         [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:26];
                         [encoder setBytes:&scale   length:sizeof(   float) atIndex:27];
 
-                        // for small batches use more simdgroups (needs more tests, to confirm if it's worth it)
-                        const int64_t nsg = ne01 < 4 ? 12 : 4;  // simdgroups per threadgroup (a.k.a. warps)
-
                         const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
-                        const int64_t ncpsg = 32; // cache values per simdgroup (does not work for other values)
+                        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
 
-                      //const size_t smem = nqptg*(nhptg*ne00 + nsg*(nhptg*ne00 + 256))*(sizeof(float)/2);
+                        GGML_ASSERT(nqptg % 8  == 0);
-                        const size_t smem = nqptg*(ne00 + nsg*(ne00 + 1*ncpsg))*(sizeof(float)/2);
+                        GGML_ASSERT(ncpsg % 32 == 0);
+
+                        // simdgroups per threadgroup (a.k.a. warps)
+                        // for small batches use more simdgroups (needs more tests, to confirm if it's worth it)
+                        const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)) : 4;
+
+                        const size_t smem = nqptg*(ne00 + nsg*(ncpsg + nqptg))*(sizeof(float)/2);
 
                         //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
                         GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
@@ -2465,10 +2430,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe
     UNUSED(buft);
 }
 
-static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
+static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
+#ifndef GGML_METAL_NDEBUG
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
                 device.currentAllocatedSize / 1024.0 / 1024.0,
                 device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
@@ -2478,10 +2446,15 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
             GGML_METAL_LOG_INFO("\n");
         }
     } else {
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0);
     }
+#endif
 #endif
     UNUSED(device);
+    UNUSED(size_aligned);
 }
 
 GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -2515,8 +2488,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
         return NULL;
     }
 
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+    ggml_backend_metal_log_allocated_size(device, size_aligned);
-    ggml_backend_metal_log_allocated_size(device);
 
     return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
@@ -2544,6 +2516,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_name         = */ ggml_backend_metal_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // TODO: return device.maxBufferLength
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
@@ -2592,7 +2565,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
             return false;
         }
 
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+        ggml_backend_metal_log_allocated_size(device, size_aligned);
 
         ++ctx->n_buffers;
     } else {
@@ -2615,7 +2588,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            ggml_backend_metal_log_allocated_size(device, size_step_aligned);
+
             if (i + size_step < size) {
                 GGML_METAL_LOG_INFO("\n");
             }
@@ -2624,8 +2598,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
         }
     }
 
-    ggml_backend_metal_log_allocated_size(device);
-
     return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
 }
 

ggml-metal.metal

@@ -1995,6 +1995,7 @@ typedef void (flash_attn_ext_f16_t)(
         uint  tiisg[[thread_index_in_simdgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]]);
 
+// ref: https://arxiv.org/pdf/2307.08691.pdf
 template<int64_t D, int64_t Q, int64_t C> // head size, queries per threadgroup, cache items per threadgroup
 kernel void kernel_flash_attn_ext_f16(
         device const  char * q,
@@ -2038,39 +2039,45 @@ kernel void kernel_flash_attn_ext_f16(
     const int64_t iq1 = tgpig[0]*Q;
 
     const int64_t D4 = D/4;
-    const int64_t N4 = N_SIMDWIDTH;
-    const int64_t L4 = (D4 + N4 - 1)/N4;
     const int64_t D8 = D/8;
+    const int64_t Q8 = Q/8;
+    const int64_t NW = N_SIMDWIDTH;
+    const int64_t SH = (C + Q); // shared memory per simdgroup in (half)
 
-    const int64_t T  = D + nsg*(D + 1*C); // shared memory size per query in half
+    const int64_t T  = D + nsg*SH; // shared memory size per query in (half)
-    const int64_t T4 = T/4;               // shared memory size per query in half4
+    const int64_t T4 = T/4;        // shared memory size per query in (half4)
 
-    threadgroup half  * pq  = (threadgroup half  *) (shared +                   0*D);
+    threadgroup half  * sq  = (threadgroup half  *) (shared +            0*D); // holds the query data
-    threadgroup half4 * pq4 = (threadgroup half4 *) (shared +                   0*D);
+    threadgroup half4 * sq4 = (threadgroup half4 *) (shared +            0*D); // same as above but in half4
-    threadgroup half  * ps  = (threadgroup half  *) (shared + sgitg*(D + 1*C) + 1*D);
+    threadgroup half  * ss  = (threadgroup half  *) (shared + sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
-    threadgroup half4 * ps4 = (threadgroup half4 *) (shared + sgitg*(D + 1*C) + 1*D);
+
-    threadgroup half  * ss  = (threadgroup half  *) (shared + sgitg*(D + 1*C) + 2*D);
+    // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
+    simdgroup_half8x8 lo[Q8][D8];
 
-    for (int64_t i = 0; i < L4; ++i) {
     // load heads from Q to shared memory
     for (int64_t j = sgitg; j < Q; j += nsg) {
         device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03));
+
+        for (int64_t i = tiisg; i < D4; i += NW) {
             if (iq1 + j < ne01) {
-                pq4[j*T4 + N4*i + tiisg] = (half4) q4[N4*i + tiisg];
+                sq4[j*T4 + i] = (half4) q4[i];
             } else {
-                pq4[j*T4 + N4*i + tiisg] = 0.0h;
+                sq4[j*T4 + i] = 0.0h;
+            }
         }
     }
 
-        // zero out shared memory
+    // zero out lo
-        for (int64_t j = 0; j < Q; ++j) {
+    for (int64_t j = 0; j < Q8; ++j) {
-            ps4[j*T4 + N4*i + tiisg] = 0.0h;
+        for (int64_t i = 0; i < D8; ++i) {
+            lo[j][i] = make_filled_simdgroup_matrix<half, 8>(0.0h);
         }
     }
 
-    if (tiisg < C) {
+    // zero out shared memory SH
     for (int64_t j = 0; j < Q; ++j) {
-            ss[j*T + 0 + tiisg] = 0.0h;
+        for (int64_t i = tiisg; i < SH; i += NW) {
+            ss[j*T + i] = 0.0h;
         }
     }
 
@@ -2103,79 +2110,68 @@ kernel void kernel_flash_attn_ext_f16(
         const int64_t iv2 = iq2 / rv2;
         const int64_t iv3 = iq3 / rv3;
 
-        simdgroup_half8x8 mq[D8];
+        // load the queries from shared memory into local memory
+        simdgroup_half8x8 mq[Q8][D8];
 
+        for (int64_t j = 0; j < Q8; ++j) {
             for (int64_t i = 0; i < D8; ++i) {
-            simdgroup_load(mq[i], pq + i*8, T);
+                simdgroup_load(mq[j][i], sq + 8*j*T + i*8, T);
+            }
         }
 
-        // TODO: this can be improved
-        device const float * mp[Q];
-
-        {
         const int64_t ir = iq3*ne02*ne01 + iq2*ne01 + iq1;
 
-            for (int64_t j = 0; j < Q; ++j) {
+        // pointer to the mask
-                if (iq1 + j < ne01) {
+        device const float * mp = (device const float *) (mask + (ir%ne31)*nb31);
-                    mp[j] = (device const float *) (mask + ((ir + j)%ne31)*nb31);
-                } else {
-                    mp[j] = nullptr;
-                }
-            }
-        }
 
         // prepare diagonal scale matrix
-        simdgroup_half8x8 mscale(scale);
+        simdgroup_float8x8 mscale(scale);
-
-        for (int64_t iic = C*sgitg; iic < ne11; iic += C*nsg) {
-            // skip -INF blocks
-            // TODO: double-check this
-            {
-                float smc = -INFINITY;
-
-                for (int64_t j = 0; j < Q; ++j) {
-                    const float mc = mp[j] ? mp[j][iic + tiisg] : -INFINITY;
-                    smc = simd_max(max(smc, mc));
-                }
-
-                if (smc == -INFINITY) {
-                    continue;
-                }
-            }
 
+        // loop over the KV cache
+        // each simdgroup handles blocks of Q rows and C columns
+        for (int64_t ic = C*sgitg; ic < ne11; ic += C*nsg) {
             // Q*K^T
             {
-                simdgroup_half8x8 mk;
-
                 for (int cc = 0; cc < C/8; ++cc) {
-                    simdgroup_half8x8 mqk = make_filled_simdgroup_matrix<half, Q>(0.h);
+                    simdgroup_half8x8 mqk[Q8];
+                    for (int64_t j = 0; j < Q8; ++j) {
+                        mqk[j] = make_filled_simdgroup_matrix<half, 8>(0.h);
+                    }
 
-                    device const half * pk = (device const half *) ((device const char *) k + ((iic + 8*cc)*nb11 + ik2*nb12 + ik3*nb13));
+                    device const half * pk = (device const half *) ((device const char *) k + ((ic + 8*cc)*nb11 + ik2*nb12 + ik3*nb13));
 
                     for (int64_t i = 0; i < D8; ++i) {
-                        simdgroup_load(mk, pk + i*8, nb11/sizeof(half), 0, true);
+                        simdgroup_half8x8 mk;
+                        simdgroup_load(mk, pk + i*8, nb11/sizeof(half), 0, true); // transpose
 
-                        simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
+                        for (int64_t j = 0; j < Q8; ++j) {
+                            simdgroup_multiply_accumulate(mqk[j], mq[j][i], mk, mqk[j]);
+                        }
                     }
 
                     // mqk = mqk*scale + mask
+                    for (int64_t j = 0; j < Q8; ++j) {
                         simdgroup_float8x8 mm;
-                    simdgroup_load(mm, mp[0] + iic + 8*cc, nb31/sizeof(float), 0, false);
+                        simdgroup_load(mm, mp + 8*j*(nb31/sizeof(float)) + ic + 8*cc, nb31/sizeof(float), 0, false);
-                    simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
+                        simdgroup_multiply_accumulate(mqk[j], mqk[j], mscale, mm);
 
-                    simdgroup_store(mqk, ss + 8*cc, T, 0, false);
+                        simdgroup_store(mqk[j], ss + 8*j*T + 8*cc, T, 0, false);
                     }
                 }
+            }
+
+            // used to detect blocks full of -INF
+            half smax = -INFINITY;
 
             // online softmax
+            if (C == 32) {
                 for (int64_t j = 0; j < Q; ++j) {
                     const int64_t p = tiisg;
 
-                //const half s = ss[j*T + p]*scale + (mp[j][iic + p]);
+                    const half m = M[j];
                     const half s = ss[j*T + p];
 
-                half m = M[j];
+                    smax = simd_max(max(smax, s));
-
                     M[j] = simd_max(max(M[j], s));
 
                     const half ms = m == -INFINITY ? 0.0h : exp(m - M[j]);
@@ -2183,42 +2179,83 @@ kernel void kernel_flash_attn_ext_f16(
 
                     S[j] = S[j]*ms + simd_sum(vs);
 
-                for (int64_t i = 0; i < L4; ++i) {
+                    // create a QxQ diagonal matrix for rescaling the output
-                    ps4[j*T4 + N4*i + tiisg] *= ms;
+                    if (p == j) {
+                        ss[j*T + C + j] = ms;
                     }
 
+                    // the P matrix from the paper (Q rows, C columns)
                     ss[j*T + p] = vs;
                 }
+            } else {
+                for (int64_t j = 0; j < Q; ++j) {
+                    const half m = M[j];
 
-            simdgroup_barrier(mem_flags::mem_none);
+                    for (int64_t p = tiisg; p < C; p += NW) {
+                        const half s = ss[j*T + p];
 
-            // (Q*K^T)*V
+                        smax = simd_max(max(smax, s));
-            {
+                        M[j] = simd_max(max(M[j], s));
-                simdgroup_half8x8 mv;
-
-                simdgroup_half8x8 mp[C/8];
-                for (int cc = 0; cc < C/8; ++cc) {
-                    simdgroup_load(mp[cc], ss + 8*cc, T, 0, false);
                     }
 
+                    const half ms = m == -INFINITY ? 0.0h : exp(m - M[j]);
+
+                    S[j] = S[j]*ms;
+
+                    // create a QxQ diagonal matrix for rescaling the output
+                    if (tiisg == j) {
+                        ss[j*T + C + j] = ms;
+                    }
+
+                    for (int64_t p = tiisg; p < C; p += NW) {
+                        const half s = ss[j*T + p];
+
+                        const half vs = s == -INFINITY ? 0.0h : exp(s - M[j]);
+
+                        S[j] = S[j] + simd_sum(vs);
+
+                        // the P matrix from the paper (Q rows, C columns)
+                        ss[j*T + p] = vs;
+                    }
+                }
+            }
+
+            // skip -INF blocks
+            if (smax == -INFINITY) {
+                continue;
+            }
+
+            // O = diag(ms)*O
+            for (int64_t j = 0; j < Q8; ++j) {
+                simdgroup_half8x8 mm;
+                simdgroup_load(mm, ss + 8*j*T + C + 8*j, T, 0, false);
+
                 for (int64_t i = 0; i < D8; ++i) {
-                    simdgroup_half8x8 mqkv;
+                    simdgroup_multiply(lo[j][i], mm, lo[j][i]);
-
+                }
-                    simdgroup_load(mqkv, ps + i*8, T, 0, false);
+            }
 
+            // O = O + (Q*K^T)*V
+            {
                 for (int cc = 0; cc < C/8; ++cc) {
-                        device const half * pv = (device const half *) ((device const char *) v + ((iic + 8*cc)*nb21 + iv2*nb22 + iv3*nb23));
+                    device const half * pv = (device const half *) ((device const char *) v + ((ic + 8*cc)*nb21 + iv2*nb22 + iv3*nb23));
 
-                        simdgroup_load(mv, pv + i*8, nb21/sizeof(half), 0, false);
+                    for (int64_t i = 0; i < D8; ++i) {
+                        simdgroup_half8x8 mk;
+                        simdgroup_load(mk, pv + i*8, nb21/sizeof(half), 0, false);
 
-                        simdgroup_multiply_accumulate(mqkv, mp[cc], mv, mqkv);
+                        for (int64_t j = 0; j < Q8; ++j) {
+                            simdgroup_half8x8 mv;
+                            simdgroup_load(mv, ss + 8*j*T + 8*cc, T, 0, false);
+
+                            simdgroup_multiply_accumulate(lo[j][i], mv, mk, lo[j][i]);
+                        }
                     }
-
-                    simdgroup_store(mqkv, ps + i*8, T, 0, false);
                 }
             }
         }
 
+        // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
         for (int64_t j = 0; j < Q; ++j) {
             if (tiisg == 0) {
                 ss[j*T + 0] = S[j];
@@ -2227,91 +2264,86 @@ kernel void kernel_flash_attn_ext_f16(
         }
     }
 
-    threadgroup_barrier(mem_flags::mem_threadgroup);
+    // reduce the warps sequentially
-
-    // reduce the warps
-#if 1
-    if (sgitg == 0) {
-        half S = { 0.0h };
-        half M = { -INFINITY };
-
     for (int64_t sg = 1; sg < nsg; ++sg) {
-            for (int64_t j = 0; j < Q; ++j) {
-                const half S0 = ss[j*T +                0];
-                const half S1 = ss[j*T + sg*(D + 1*C) + 0];
-
-                const half M0 = ss[j*T +                1];
-                const half M1 = ss[j*T + sg*(D + 1*C) + 1];
-
-                M = max(M0, M1);
-
-                const half ms0 = exp(M0 - M);
-                const half ms1 = exp(M1 - M);
-
-                S = S0*ms0 + S1*ms1;
-
-                if (tiisg == 0) {
-                    ss[j*T + 0] = S;
-                    ss[j*T + 1] = M;
-                }
-
-                for (int64_t i = 0; i < L4; ++i) {
-                    ps4[j*T4 + N4*i + tiisg] = ps4[j*T4 + N4*i + tiisg]*ms0 + ps4[j*T4 + sg*(D + 1*C)/4 + N4*i + tiisg]*ms1;
-                }
-            }
-        }
-    }
-#else
-    // parallel reduce
-    // NOTE: this is significantly slower than the serial version above, likely due to the small number of warps
-    {
         half S = { 0.0h };
         half M = { -INFINITY };
 
-        for (int64_t sg = nsg/2; sg > 0; sg /= 2) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
-            if (sgitg >= sg) {
+
-                continue;
+        // each simdgroup stores its output to shared memory, reusing sq
+        if (sgitg == sg) {
+            for (int64_t j = 0; j < Q8; ++j) {
+                for (int64_t i = 0; i < D8; ++i) {
+                    simdgroup_store(lo[j][i], sq + 8*j*T + i*8, T, 0, false);
                 }
-
-            for (int64_t j = 0; j < Q; ++j) {
-                const half S0 = ss[j*T +                0];
-                const half S1 = ss[j*T + sg*(D + 1*C) + 0];
-
-                const half M0 = ss[j*T +                1];
-                const half M1 = ss[j*T + sg*(D + 1*C) + 1];
-
-                M = max(M0, M1);
-
-                const half ms0 = exp(M0 - M);
-                const half ms1 = exp(M1 - M);
-
-                S = S0*ms0 + S1*ms1;
-
-                if (tiisg == 0) {
-                    ss[j*T + 0] = S;
-                    ss[j*T + 1] = M;
-                }
-
-                for (int64_t i = 0; i < L4; ++i) {
-                    ps4[j*T4 + N4*i + tiisg] = ps4[j*T4 + N4*i + tiisg]*ms0 + ps4[j*T4 + sg*(D + 1*C)/4 + N4*i + tiisg]*ms1;
             }
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        }
-    }
-#endif
 
-    simdgroup_barrier(mem_flags::mem_threadgroup);
+        // the first simdgroup accumulates the results from the other simdgroups
+        if (sgitg == 0) {
+            for (int64_t j = 0; j < Q; ++j) {
+                const half S0 = ss[j*T +         0];
+                const half S1 = ss[j*T + sg*SH + 0];
+
+                const half M0 = ss[j*T +         1];
+                const half M1 = ss[j*T + sg*SH + 1];
+
+                M = max(M0, M1);
+
+                const half ms0 = M0 == -INFINITY ? 0.0h : exp(M0 - M);
+                const half ms1 = M1 == -INFINITY ? 0.0h : exp(M1 - M);
+
+                S = S0*ms0 + S1*ms1;
+
+                if (tiisg == 0) {
+                    ss[j*T + 0] = S;
+                    ss[j*T + 1] = M;
+
+                    ss[j*T + C + j        ] = ms0;
+                    ss[j*T + C + j + sg*SH] = ms1;
+                }
+            }
+
+            // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
+            for (int64_t j = 0; j < Q8; ++j) {
+                simdgroup_half8x8 t;
+                simdgroup_half8x8 ms0;
+                simdgroup_half8x8 ms1;
+
+                simdgroup_load(ms0, ss + 8*j*T + C + 8*j,         T, 0, false);
+                simdgroup_load(ms1, ss + 8*j*T + C + 8*j + sg*SH, T, 0, false);
+
+                for (int64_t i = 0; i < D8; ++i) {
+                    simdgroup_load    (t, sq + 8*j*T + i*8, T, 0, false);
+                    simdgroup_multiply(t, ms1, t);
+
+                    simdgroup_multiply_accumulate(lo[j][i], ms0, lo[j][i], t);
+                }
+            }
+        }
+    }
+
+    // store result to shared memory (reuse sq)
+    if (sgitg == 0) {
+        for (int64_t j = 0; j < Q8; ++j) {
+            for (int64_t i = 0; i < D8; ++i) {
+                simdgroup_store(lo[j][i], sq + 8*j*T + i*8, T, 0, false);
+            }
+        }
+    }
 
     device float4 * dst4 = (device float4 *) dst;
 
+    // final rescale with 1/S and store to global memory
     if (sgitg == 0) {
         for (int64_t j = 0; j < Q && iq1 + j < ne01; ++j) {
             const half S = ss[j*T + 0];
 
-            for (int64_t i = 0; i < L4; ++i) {
+            for (int64_t i = tiisg; i < D4; i += NW) {
-                dst4[(iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + N4*i + tiisg] = (float4) ps4[j*T4 + N4*i + tiisg]/S;
+                dst4[(iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) sq4[j*T4 + i]/S;
             }
         }
     }
@@ -2319,7 +2351,10 @@ kernel void kernel_flash_attn_ext_f16(
 
 template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<64,  8, 32>;
 template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<80,  8, 32>;
+template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96,  8, 32>;
+template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112, 8, 32>;
 template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128, 8, 32>;
+template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256, 8, 32>;
 
 kernel void kernel_cpy_f16_f16(
         device  const half * src0,

ggml-opencl.cpp

@@ -714,7 +714,6 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx,
         dst[row] = tmp[0];
     }
 }
-
 );
 
 
@@ -784,6 +783,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
         dst[row] = tmp[0];
     }
 }
+
 );
 
 
@@ -799,6 +799,18 @@ __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y
 }
 );
 
+std::string add_template = MULTILINE_QUOTE(
+__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) {
+    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
+
+    if (i >= get_global_size(0)) {
+        return;
+    }
+
+    dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky];
+}
+);
+
 #define CL_CHECK(err)                                               \
     do {                                                            \
         cl_int err_ = (err);                                        \
@@ -878,6 +890,7 @@ static std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
+    src << add_template << '\n';
 
     return src.str();
 }
@@ -893,6 +906,7 @@ static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl,
 static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
 static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
 static cl_kernel mul_f32_cl;
+static cl_kernel add_f32_cl;
 static bool fp16_support;
 
 static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
@@ -1100,9 +1114,10 @@ void ggml_cl_init(void) {
     char *ext_buffer = (char *)alloca(ext_str_size + 1);
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
     ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
+    // Disabled due to faulty outputs
     // Check if ext_buffer contains cl_khr_fp16
-    fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
+    fp16_support = false;  // strstr(ext_buffer, "cl_khr_fp16") != NULL;
-    fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
+    // fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
 
     cl_context_properties properties[] = {
         (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
@@ -1150,6 +1165,8 @@ void ggml_cl_init(void) {
 
     // mul kernel
     CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
+
+    CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err));
 }
 
 static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
@@ -1458,6 +1475,70 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
     ggml_cl_mul_f32(src0, src1, dst);
 }
 
+static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+    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 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];
+    size_t x_size;
+    size_t d_size;
+
+    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
+    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
+
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            cl_event ev;
+
+            // copy src0 to device
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
+
+            const int64_t i13 = i03%ne13;
+            const int64_t i12 = i02%ne12;
+            const int i1 = i13*ne12*ne11 + i12*ne11;
+
+            cl_int x_offset = 0;
+            cl_int y_offset = i1*ne10;
+            cl_int d_offset = 0;
+
+            size_t global = ne00 * ne01;
+            cl_int ky = ne10 * ne11;
+
+            CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky));
+            CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
+
+            CL_CHECK(clReleaseEvent(ev));
+            CL_CHECK(clFinish(queue));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
+        }
+    }
+    ggml_cl_pool_free(d_X, x_size);
+    ggml_cl_pool_free(d_D, d_size);
+}
+
+void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cl_add_f32(src0, src1, dst);
+}
+
 static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -2055,6 +2136,7 @@ static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_opencl_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // TODO: return from device info
     /* .get_alloc_size   = */ NULL,
     /* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
@@ -2111,6 +2193,7 @@ ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
             /* .get_name         = */ ggml_backend_opencl_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,

ggml-opencl.h

@@ -10,6 +10,7 @@ extern "C" {
 GGML_API void ggml_cl_init(void);
 
 GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
 GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);

ggml-sycl.h

@@ -0,0 +1,27 @@
+/*MIT license
+  Copyright (C) 2024 Intel Corporation
+  SPDX-License-Identifier: MIT
+*/
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_SYCL_MAX_DEVICES       16
+#define GGML_SYCL_NAME "SYCL"
+
+GGML_API void   ggml_init_sycl(void);
+GGML_API bool   ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-vulkan.h

@@ -0,0 +1,34 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_VK_NAME "Vulkan"
+
+GGML_API void ggml_vk_init(void);
+
+GGML_API void ggml_vk_preallocate_buffers_graph(struct ggml_tensor * node);
+GGML_API void ggml_vk_preallocate_buffers(void);
+GGML_API void ggml_vk_build_graph(struct ggml_tensor * node, bool last_node);
+GGML_API bool ggml_vk_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#ifdef GGML_VULKAN_CHECK_RESULTS
+void ggml_vk_check_results_1(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#endif
+GGML_API void ggml_vk_graph_cleanup(void);
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(void);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml.c

@@ -248,6 +248,10 @@ inline static void * ggml_aligned_malloc(size_t size) {
 #include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
+#elif defined(GGML_USE_VULKAN)
+#include "ggml-vulkan.h"
+#elif defined(GGML_USE_SYCL)
+#include "ggml-sycl.h"
 #endif
 
 // floating point type used to accumulate sums
@@ -1344,12 +1348,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        y[i] += GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
     }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
-        y[i] += GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
     }
 #endif
 }
@@ -1478,6 +1482,9 @@ 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; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+// TODO: optimize performance
+inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
@@ -1838,9 +1845,11 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "GELU",
     "GELU_QUICK",
     "SILU",
+    "HARDSWISH",
+    "HARDSIGMOID",
 };
 
-static_assert(GGML_UNARY_OP_COUNT == 10, "GGML_UNARY_OP_COUNT != 10");
+static_assert(GGML_UNARY_OP_COUNT == 12, "GGML_UNARY_OP_COUNT != 12");
 
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -2350,6 +2359,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         ggml_init_cublas();
 #elif defined(GGML_USE_CLBLAST)
         ggml_cl_init();
+#elif defined(GGML_USE_VULKAN)
+        ggml_vk_init();
+#elif defined(GGML_USE_SYCL)
+        ggml_init_sycl();
 #endif
 
         ggml_setup_op_has_task_pass();
@@ -4007,6 +4020,20 @@ struct ggml_tensor * ggml_silu_back(
     return result;
 }
 
+// ggml hardswish
+struct ggml_tensor * ggml_hardswish(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSWISH);
+}
+
+// ggml hardsigmoid
+struct ggml_tensor * ggml_hardsigmoid(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID);
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
@@ -5408,6 +5435,31 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
     return result;
 }
 
+// ggml_conv_depthwise
+struct ggml_tensor * ggml_conv_depthwise_2d(
+    struct ggml_context * ctx,
+    struct ggml_tensor * a,
+    struct ggml_tensor * b,
+    int                  s0,
+    int                  s1,
+    int                  p0,
+    int                  p1,
+    int                  d0,
+    int                  d1) {
+    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a,
+                                        ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
+                                        s0, s1, p0, p1, d0, d1, true); // [N * IC, OH, OW, KH * KW]
+
+    struct ggml_tensor * result =
+        ggml_mul_mat(ctx,
+                ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1),                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
+                ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3])); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+
+    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+
+    return result;
+}
 // ggml_conv_2d
 
 // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
@@ -7278,6 +7330,17 @@ static void ggml_compute_forward_add_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
+#ifdef GGML_USE_CLBLAST
+    if (src1->backend == GGML_BACKEND_GPU) {
+        // TODO: OpenCL kernel support full broadcast
+        GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
+        if (ith == 0) {
+            ggml_cl_add(src0, src1, dst);
+        }
+        return;
+    }
+#endif
+
     const int nr  = ggml_nrows(src0);
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -7558,7 +7621,12 @@ static void ggml_compute_forward_add(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
+                if (src1->type == GGML_TYPE_F32) {
                     ggml_compute_forward_add_f32(params, src0, src1, dst);
+                }
+                else {
+                    GGML_ASSERT(false);
+                }
             } break;
         case GGML_TYPE_F16:
             {
@@ -7879,6 +7947,9 @@ static void ggml_compute_forward_acc_f32(
     bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
 
     if (!inplace && (params->type == GGML_TASK_INIT)) {
+        if (params->ith != 0) {
+            return;
+        }
         // memcpy needs to be synchronized across threads to avoid race conditions.
         // => do it in INIT phase
         memcpy(
@@ -8067,7 +8138,7 @@ static void ggml_compute_forward_mul_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-#ifdef GGML_USE_CLBLAST
+#if defined(GGML_USE_CLBLAST)
     if (src1->backend == GGML_BACKEND_GPU) {
         // TODO: OpenCL kernel support full broadcast
         GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
@@ -9448,6 +9519,87 @@ static void ggml_compute_forward_silu_back(
     }
 }
 
+
+static void ggml_compute_forward_hardswish_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    assert(params->ith == 0);
+    assert(ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    assert(dst->nb[0]  == sizeof(float));
+    assert(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardswish_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+static void ggml_compute_forward_hardswish(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_hardswish_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+static void ggml_compute_forward_hardsigmoid_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    assert(params->ith == 0);
+    assert(ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    assert(dst->nb[0]  == sizeof(float));
+    assert(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardsigmoid_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_hardsigmoid(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * src0,
+        struct ggml_tensor * dst) {
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_hardsigmoid_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+
 // ggml_compute_forward_norm
 
 static void ggml_compute_forward_norm_f32(
@@ -9940,11 +10092,30 @@ static void ggml_compute_forward_mul_mat(
 
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(dst)) {
-        if (params->ith != 0) {
+        const int64_t ne_plane      = ne01*ne00;
-            return;
+        const size_t  desired_wsize = ne13*ne12*ne_plane*sizeof(float);
-        }
+        UNUSED(desired_wsize);
 
         if (params->type == GGML_TASK_INIT) {
+            if (type != GGML_TYPE_F32) {
+                assert(params->wsize >= desired_wsize);
+                // parallelize by src0 rows
+                for (int64_t i13 = 0; i13 < ne13; i13++) {
+                    for (int64_t i12 = 0; i12 < ne12; i12++) {
+                        // broadcast src0 into src1 across 2nd,3rd dimension
+                        const int64_t i03 = i13/r3;
+                        const int64_t i02 = i12/r2;
+
+                        const void           *       x        = (char *)  src0->data    + i02*nb02          + i03*nb03;
+                              float          * const wdata    = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane;
+                              ggml_to_float_t  const to_float = type_traits[type].to_float;
+
+                        for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
+                            to_float((const char *) x + i01*nb01, wdata + i01*ne00, ne00);
+                        }
+                    }
+                }
+            }
             return;
         }
 
@@ -9952,9 +10123,14 @@ static void ggml_compute_forward_mul_mat(
             return;
         }
 
+        // perform sgemm, parallelization controlled by blas lib
+        if (ith != 0) {
+            return;
+        }
+
+        //const int64_t tgemm0 = ggml_perf_time_us();
         for (int64_t i13 = 0; i13 < ne13; i13++) {
             for (int64_t i12 = 0; i12 < ne12; i12++) {
-                // broadcast src0 into src1 across 2nd,3rd dimension
                 const int64_t i03 = i13/r3;
                 const int64_t i02 = i12/r2;
 
@@ -9963,17 +10139,7 @@ static void ggml_compute_forward_mul_mat(
                       float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
 
                 if (type != GGML_TYPE_F32) {
-                            float * const wdata    = params->wdata;
+                    x = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane;
-                    ggml_to_float_t const to_float = type_traits[type].to_float;
-
-                    size_t id = 0;
-                    for (int64_t i01 = 0; i01 < ne01; ++i01) {
-                        to_float((const char *) x + i01*nb01, wdata + id, ne00);
-                        id += ne00;
-                    }
-
-                    assert(id*sizeof(float) <= params->wsize);
-                    x = wdata;
                 }
 
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
@@ -9983,6 +10149,7 @@ static void ggml_compute_forward_mul_mat(
                          0.0f,    d, ne01);
             }
         }
+        //printf("cblas_sgemm = %.3f ms, %lld flops\n", (ggml_perf_time_us() - tgemm0)/1000.0, ne13*ne12*ne1*ne01*ne10*2);
 
         //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
 
@@ -9991,6 +10158,9 @@ static void ggml_compute_forward_mul_mat(
 #endif
 
     if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         if (src1->type != vec_dot_type) {
             char * wdata = params->wdata;
             const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@@ -10155,6 +10325,9 @@ static void ggml_compute_forward_mul_mat_id(
     #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
 
    if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         char * wdata = params->wdata;
         if (src1->type != vec_dot_type) {
             const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@@ -10340,6 +10513,9 @@ static void ggml_compute_forward_out_prod_f32(
             return;
         }
 #endif
+        if (ith != 0) {
+            return;
+        }
         ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
         return;
     }
@@ -10523,6 +10699,9 @@ static void ggml_compute_forward_out_prod_q_f32(
     // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
 
     if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
         return;
     }
@@ -10707,6 +10886,9 @@ static void ggml_compute_forward_set_f32(
     bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
 
     if (!inplace && (params->type == GGML_TASK_INIT)) {
+        if (params->ith != 0) {
+            return;
+        }
         // memcpy needs to be synchronized across threads to avoid race conditions.
         // => do it in INIT phase
         memcpy(
@@ -11031,6 +11213,9 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
     // ggml_compute_forward_dup_same_cont(params, opt0, dst);
 
     if (params->type == GGML_TASK_INIT) {
+        if (params->ith != 0) {
+            return;
+        }
         memset(dst->data, 0, ggml_nbytes(dst));
     }
 
@@ -11065,6 +11250,9 @@ static void ggml_compute_forward_get_rows_back_f32(
     // ggml_compute_forward_dup_same_cont(params, opt0, dst);
 
     if (params->type == GGML_TASK_INIT) {
+        if (params->ith != 0) {
+            return;
+        }
         memset(dst->data, 0, ggml_nbytes(dst));
     }
 
@@ -11202,6 +11390,9 @@ static void ggml_compute_forward_diag_mask_f32(
     GGML_ASSERT(n_past >= 0);
 
     if (!inplace && (params->type == GGML_TASK_INIT)) {
+        if (ith != 0) {
+            return;
+        }
         // memcpy needs to be synchronized across threads to avoid race conditions.
         // => do it in INIT phase
         GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
@@ -12172,6 +12363,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         memset(params->wdata, 0, params->wsize);
 
         // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
@@ -12266,6 +12460,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         memset(params->wdata, 0, params->wsize);
 
         // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
@@ -12464,6 +12661,7 @@ static void ggml_compute_forward_im2col(
     }
 }
 
+
 // ggml_compute_forward_conv_transpose_2d
 
 static void ggml_compute_forward_conv_transpose_2d(
@@ -12489,6 +12687,9 @@ static void ggml_compute_forward_conv_transpose_2d(
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
+        if (ith != 0) {
+            return;
+        }
         memset(params->wdata, 0, params->wsize);
 
         // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
@@ -13353,11 +13554,9 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
     const int64_t D = neq0;
     const int64_t N = neq1;
-    const int64_t P = nek1 - N;
 
     GGML_ASSERT(ne0 == D);
     GGML_ASSERT(ne2 == N);
-    GGML_ASSERT(P >= 0);
 
     GGML_ASSERT(nbq0 == sizeof(float));
     GGML_ASSERT(nbk0 == sizeof(ggml_fp16_t));
@@ -13368,7 +13567,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(nev0 == D);
 
     GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nek1 == N + P);
     GGML_ASSERT(nev0 == D);
 
     // dst cannot be transposed or permuted
@@ -13407,8 +13605,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     float scale = 1.0f;
     memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
 
-    //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
-
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -14228,6 +14424,14 @@ static void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_silu(params, src0, dst);
             } break;
+        case GGML_UNARY_OP_HARDSWISH:
+            {
+                ggml_compute_forward_hardswish(params, src0, dst);
+            } break;
+        case GGML_UNARY_OP_HARDSIGMOID:
+            {
+                ggml_compute_forward_hardsigmoid(params, src0, dst);
+            } break;
         default:
             {
                 GGML_ASSERT(false);
@@ -14291,6 +14495,9 @@ static void ggml_compute_forward_add_rel_pos_f32(
 
     const bool inplace = (bool) ((int32_t *) dst->op_params)[0];
     if (!inplace && params->type == GGML_TASK_INIT) {
+        if (params->ith != 0) {
+            return;
+        }
         memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst));
         return;
     }
@@ -14806,8 +15013,26 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
     GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
     GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
+#elif defined(GGML_USE_VULKAN)
+    const bool skip_cpu = ggml_vk_compute_forward(params, tensor);
+#ifdef GGML_VULKAN_CHECK_RESULTS
+    if (skip_cpu) {
+        ggml_vk_check_results_1(params, tensor);
+    }
+#endif
+    if (skip_cpu) {
+        return;
+    }
+    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
+    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
 #endif // GGML_USE_CUBLAS
 
+#ifdef GGML_USE_SYCL
+    bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
+    if (skip_cpu) {
+        return;
+    }
+#endif // GGML_USE_SYCL
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -16591,6 +16816,7 @@ struct ggml_compute_state_shared {
     // synchronization primitives
     atomic_int n_active;  // num active threads
     atomic_int node_n;    // active graph node
+    atomic_int node_task; // active graph node task phase
 
     bool (*abort_callback)(void * data); // abort ggml_graph_compute when true
     void * abort_callback_data;
@@ -16646,6 +16872,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSWISH: // to opt for multiple threads
+                case GGML_UNARY_OP_HARDSIGMOID: // to opt for multiple threads
                     {
                         n_tasks = 1;
                     } break;
@@ -16722,7 +16950,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+                n_tasks = MIN(n_threads, ggml_nrows(node->src[0]));
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
@@ -16837,6 +17065,34 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
     return n_tasks;
 }
 
+static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_compute_state * state, const bool do_yield) {
+    // wait for other threads to finish
+    const int last_node_n = * node_n;
+
+    while (true) {
+        if (do_yield) {
+            sched_yield();
+        }
+
+        * node_n = atomic_load(&state->shared->node_n);
+        if (* node_n != last_node_n) break;
+    }
+}
+
+static void ggml_graph_compute_thread_sync_task(int * task_phase, struct ggml_compute_state * state, const bool do_yield) {
+    // wait for other threads to finish
+    const int last_task_phase = * task_phase;
+
+    while (true) {
+        if (do_yield) {
+            sched_yield();
+        }
+
+        * task_phase = atomic_load(&state->shared->node_task);
+        if (* task_phase != last_task_phase) break;
+    }
+}
+
 static thread_ret_t ggml_graph_compute_thread(void * data) {
     struct ggml_compute_state * state = (struct ggml_compute_state *) data;
 
@@ -16848,6 +17104,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     set_numa_thread_affinity(state->ith, n_threads);
 
     int node_n     = -1;
+    int task_phase = GGML_TASK_FINALIZE;
 
     while (true) {
         if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
@@ -16879,7 +17136,6 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             // distribute new work or execute it direct if 1T
             while (++node_n < cgraph->n_nodes) {
                 GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
-
                 struct ggml_tensor * node = cgraph->nodes[node_n];
                 const int n_tasks = ggml_get_n_tasks(node, n_threads);
 
@@ -16888,13 +17144,13 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
                 params.nth = n_tasks;
 
+                if (n_tasks == 1) {
                     /* INIT */
                     if (GGML_OP_HAS_INIT[node->op]) {
                         params.type = GGML_TASK_INIT;
                         ggml_compute_forward(&params, node);
                     }
 
-                if (n_tasks == 1) {
                     // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1,
                     // they do something more efficient than spinning (?)
                     params.type = GGML_TASK_COMPUTE;
@@ -16915,38 +17171,24 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
                 }
             }
 
+            task_phase = GGML_TASK_INIT;
             atomic_store(&state->shared->n_active,  n_threads);
             atomic_store(&state->shared->node_n,    node_n);
+            atomic_store(&state->shared->node_task, task_phase);
         } else {
-            // wait for other threads to finish
+            ggml_graph_compute_thread_sync_node(&node_n,     state, false);
-            const int last = node_n;
+            ggml_graph_compute_thread_sync_task(&task_phase, state, false);
-
-            const bool do_yield = last < 0 || cgraph->nodes[last]->op == GGML_OP_MUL_MAT;
-
-            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
-                // UPD:  adding the do_yield flag seems to resolve the issue universally
-                if (do_yield) {
-                    sched_yield();
-                }
-
-                node_n = atomic_load(&state->shared->node_n);
-                if (node_n != last) break;
-            };
         }
 
         // check if we should stop
         if (node_n >= cgraph->n_nodes) break;
 
-        /* COMPUTE */
+        /* INIT & COMPUTE */
         struct ggml_tensor * node = cgraph->nodes[node_n];
         const int n_tasks = ggml_get_n_tasks(node, n_threads);
 
         struct ggml_compute_params params = {
-            /*.type  =*/ GGML_TASK_COMPUTE,
+            /*.type  =*/ GGML_TASK_INIT,
             /*.ith   =*/ state->ith,
             /*.nth   =*/ n_tasks,
             /*.wsize =*/ cplan->work_size,
@@ -16954,10 +17196,41 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         };
 
         if (state->ith < n_tasks) {
+            if (GGML_OP_HAS_INIT[node->op]) {
                 ggml_compute_forward(&params, node);
             }
         }
 
+        if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
+            task_phase = GGML_TASK_COMPUTE;
+            atomic_store(&state->shared->n_active,  n_threads);
+            atomic_store(&state->shared->node_task, task_phase);
+        }
+        else {
+            // 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
+            // UPD:  adding the do_yield flag seems to resolve the issue universally
+            const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT;
+            ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield);
+        }
+
+        if (state->ith < n_tasks) {
+            params.type = GGML_TASK_COMPUTE;
+            ggml_compute_forward(&params, node);
+        }
+
+        if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
+            task_phase = GGML_TASK_FINALIZE;
+            atomic_store(&state->shared->n_active,  n_threads);
+            atomic_store(&state->shared->node_task, task_phase);
+        }
+        else {
+            ggml_graph_compute_thread_sync_task(&task_phase, state, false);
+        }
+    }
+
     return GGML_EXIT_SUCCESS;
 }
 
@@ -17012,8 +17285,11 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                     if (ggml_compute_forward_mul_mat_use_blas(node)) {
                         if (node->src[0]->type != GGML_TYPE_F32) {
-                            // here we need memory just for single 2D matrix from src0
+                            // here we need memory for fully dequantized matrix from src0
-                            cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]);
+                            // take into account that src0 can be broadcasted into src1[2,3]
+                            cur = ggml_type_size(GGML_TYPE_F32)
+                                * node->src[0]->ne[0]*node->src[0]->ne[1]
+                                * node->src[1]->ne[2]*node->src[1]->ne[3];
                         }
                     } else
 #endif
@@ -17163,6 +17439,17 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         }
     }
 
+#ifdef GGML_USE_VULKAN
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_vk_preallocate_buffers_graph(cgraph->nodes[i]);
+    }
+    ggml_vk_preallocate_buffers();
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_vk_build_graph(cgraph->nodes[i], i == cgraph->n_nodes - 1);
+    }
+#endif
+
     const int n_threads = cplan->n_threads;
 
     struct ggml_compute_state_shared state_shared = {
@@ -17173,6 +17460,7 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         /*.n_threads               =*/ n_threads,
         /*.n_active                =*/ n_threads,
         /*.node_n                  =*/ -1,
+        /*.node_task               =*/ GGML_TASK_FINALIZE,
         /*.abort_callback          =*/ NULL,
         /*.abort_callback_data     =*/ NULL,
     };
@@ -17213,6 +17501,10 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         }
     }
 
+#ifdef GGML_USE_VULKAN
+    ggml_vk_graph_cleanup();
+#endif
+
     // performance stats (graph)
     {
         int64_t perf_cycles_cur  = ggml_perf_cycles()  - perf_start_cycles;
@@ -20347,7 +20639,7 @@ int ggml_cpu_has_wasm_simd(void) {
 }
 
 int ggml_cpu_has_blas(void) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_VULKAN) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_SYCL)
     return 1;
 #else
     return 0;
@@ -20370,8 +20662,24 @@ int ggml_cpu_has_clblast(void) {
 #endif
 }
 
+int ggml_cpu_has_vulkan(void) {
+#if defined(GGML_USE_VULKAN)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sycl(void) {
+#if defined(GGML_USE_SYCL)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_gpublas(void) {
-    return ggml_cpu_has_cublas() || ggml_cpu_has_clblast();
+    return ggml_cpu_has_cublas() || ggml_cpu_has_clblast() || ggml_cpu_has_vulkan() || ggml_cpu_has_sycl();
 }
 
 int ggml_cpu_has_sse3(void) {

ggml.h

@@ -490,6 +490,8 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
+        GGML_UNARY_OP_HARDSWISH,
+        GGML_UNARY_OP_HARDSIGMOID,
 
         GGML_UNARY_OP_COUNT,
     };
@@ -1033,6 +1035,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // hardswish(x) = x * relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardswish(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // hardsigmoid(x) = relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardsigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1484,6 +1496,17 @@ extern "C" {
             int                  d1,
             bool                 is_2D);
 
+    GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                  s0,
+            int                  s1,
+            int                  p0,
+            int                  p1,
+            int                  d0,
+            int                  d1);
+
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2258,9 +2281,11 @@ extern "C" {
     GGML_API int ggml_cpu_has_blas       (void);
     GGML_API int ggml_cpu_has_cublas     (void);
     GGML_API int ggml_cpu_has_clblast    (void);
+    GGML_API int ggml_cpu_has_vulkan     (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_sycl       (void);
     GGML_API int ggml_cpu_has_vsx        (void);
 
     //

gguf-py/gguf/constants.py

@@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum):
     PHI2      = auto()
     PLAMO     = auto()
     CODESHELL = auto()
+    ORION     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PHI2:           "phi2",
     MODEL_ARCH.PLAMO:          "plamo",
     MODEL_ARCH.CODESHELL:      "codeshell",
+    MODEL_ARCH.ORION:          "orion",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
-    ]
+    ],
+    MODEL_ARCH.ORION: [
+        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,
+    ],
     # TODO
 }
 
@@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #

gguf-py/gguf/gguf_reader.py

@@ -107,7 +107,7 @@ class GGUFReader:
         offs, tensors_fields = self._build_tensors_fields(offs, tensor_count)
         new_align = self.fields.get('general.alignment')
         if new_align is not None:
-            if new_align.types != [GGUFValueType.UINT64]:
+            if new_align.types != [GGUFValueType.UINT32]:
                 raise ValueError('Bad type for general.alignment field')
             self.alignment = new_align.parts[-1][0]
         padding = offs % self.alignment

llama.h

@@ -6,6 +6,9 @@
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
 #define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#elif defined(GGML_USE_SYCL)
+#include "ggml-sycl.h"
+#define LLAMA_MAX_DEVICES GGML_SYCL_MAX_DEVICES
 #else
 #define LLAMA_MAX_DEVICES 1
 #endif // GGML_USE_CUBLAS
@@ -46,7 +49,7 @@
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION 4
 
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
@@ -107,6 +110,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -774,6 +778,14 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
+    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
+    LLAMA_API void llama_sample_entropy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates_p,
+                           float   min_temp,
+                           float   max_temp,
+                           float   exponent_val);
+
     LLAMA_API void llama_sample_temp(
             struct llama_context * ctx,
           llama_token_data_array * candidates,

mypy.ini

@@ -4,3 +4,4 @@ allow_untyped_calls = true
 allow_untyped_defs = true
 allow_incomplete_defs = true
 disable_error_code = import-untyped
+warn_return_any = false

pocs/vdot/vdot.cpp

@@ -243,7 +243,6 @@ int main(int argc, char** argv) {
     if (useQ4_1) q41.resize(n4);
     else q40.resize(n4);
     std::vector<block_q8_0> q8(n8);
-    std::vector<int64_t> H(16, 0);
     double sumt = 0, sumt2 = 0, maxt = 0;
     double sumqt = 0, sumqt2 = 0, maxqt = 0;
     double sum = 0, sumq = 0, exactSum = 0;

scripts/ci-run.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -euo pipefail
+this=$(realpath "$0"); readonly this
+cd "$(dirname "$this")"
+shellcheck "$this"
+
+if (( $# != 1 && $# != 2  )); then
+    cat >&2 <<'EOF'
+usage:
+    ci-run.sh <tmp_dir> [<cache_dir>]
+
+This script wraps ci/run.sh:
+* If <tmp_dir> is a ramdisk, you can reduce writes to your SSD. If <tmp_dir> is not a ramdisk, keep in mind that total writes will increase by the size of <cache_dir>.
+    (openllama_3b_v2: quantized models are about 30GB)
+* Persistent model and data files are synced to and from <cache_dir>,
+    excluding generated .gguf files.
+    (openllama_3b_v2: persistent files are about 6.6GB)
+* <cache_dir> defaults to  ~/.cache/llama.cpp
+EOF
+    exit 1
+fi
+
+cd .. # => llama.cpp repo root
+
+tmp="$1"
+mkdir -p "$tmp"
+tmp=$(realpath "$tmp")
+echo >&2 "Using tmp=$tmp"
+
+cache="${2-$HOME/.cache/llama.cpp}"
+mkdir -p "$cache"
+cache=$(realpath "$cache")
+echo >&2 "Using cache=$cache"
+
+_sync() {
+    local from="$1"; shift
+    local to="$1"; shift
+
+    echo >&2 "Syncing from $from to $to"
+    mkdir -p "$from" "$to"
+    rsync -a "$from" "$to" --delete-during "$@"
+}
+
+_sync "$(realpath .)/" "$tmp/llama.cpp"
+_sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/"
+
+cd "$tmp/llama.cpp"
+bash ci/run.sh ci-out ci-mnt
+
+_sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P

scripts/run-with-preset.py

@@ -46,7 +46,7 @@ Formatting considerations:
 - 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> ...]"
+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.")
 

scripts/sync-ggml.last

@@ -1 +1 @@
-6c1ce0bd591a430c1d3f6797d905194581c878c1
+f2a9472b23cf27e672ed70a2a6eb078f7b060f18

tests/.gitignore

@@ -0,0 +1,3 @@
+*
+!*.*
+test-c.o

tests/CMakeLists.txt

@@ -1,6 +1,6 @@
 function(llama_build_executable source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_executable(${TEST_TARGET} ${source})
+    add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE common)
 endfunction()
@@ -8,14 +8,20 @@ endfunction()
 function(llama_test_executable name source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_test(NAME ${name} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+    set_property(TEST ${name} PROPERTY LABELS "main")
 endfunction()
 
 function(llama_build_and_test_executable source)
+    llama_build_and_test_executable_with_label(${source} "main")
+endfunction()
+
+function(llama_build_and_test_executable_with_label source label)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_executable(${TEST_TARGET} ${source})
+    add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE common)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
+    set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label})
 endfunction()
 
 # llama_build_and_test_executable(test-double-float.cpp) # SLOW
@@ -49,12 +55,14 @@ llama_build_and_test_executable(test-llama-grammar.cpp)
 llama_build_and_test_executable(test-grad0.cpp)
 # llama_build_and_test_executable(test-opt.cpp) # SLOW
 llama_build_and_test_executable(test-backend-ops.cpp)
-llama_build_and_test_executable(test-autorelease.cpp)
 
 llama_build_and_test_executable(test-rope.cpp)
 
 llama_build_executable(test-flash-attention.cpp)
 
+llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model")
+llama_build_and_test_executable_with_label(test-autorelease.cpp "model")
+
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)

tests/get-model.cpp

@@ -0,0 +1,21 @@
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+#include "get-model.h"
+
+char * get_model_or_exit(int argc, char *argv[]) {
+    char * model_path;
+    if (argc > 1) {
+        model_path = argv[1];
+
+    } else {
+        model_path = getenv("LLAMACPP_TEST_MODELFILE");
+        if (!model_path || strlen(model_path) == 0) {
+            fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n\033[0m");
+            exit(EXIT_SUCCESS);
+        }
+    }
+
+    return model_path;
+}

tests/get-model.h

@@ -0,0 +1,2 @@
+#pragma once
+char * get_model_or_exit(int, char*[]);

tests/test-autorelease.cpp

@@ -5,19 +5,15 @@
 #include <thread>
 
 #include "llama.h"
+#include "get-model.h"
 
 // This creates a new context inside a pthread and then tries to exit cleanly.
 int main(int argc, char ** argv) {
-    if (argc < 2) {
+    auto * model_path = get_model_or_exit(argc, argv);
-        printf("Usage: %s model.gguf\n", argv[0]);
-        return 0; // intentionally return success
-    }
 
-    const std::string fname = argv[1];
+    std::thread([&model_path]() {
-
-    std::thread([&fname]() {
         llama_backend_init(false);
-        auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params());
+        auto * model = llama_load_model_from_file(model_path, llama_model_default_params());
         auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
         llama_free(ctx);
         llama_free_model(model);

tests/test-backend-ops.cpp

@@ -102,7 +102,6 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                     } else if (t->type == GGML_TYPE_I8) {
                         tv.push_back((float)*(int8_t *) &buf[i]);
                     } else if (quantized) {
-                        std::vector<float> vq(ggml_blck_size(t->type));
                         tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {
@@ -240,10 +239,17 @@ static std::string var_to_str(ggml_type type) {
 #define VARS_TO_STR10(a, b, c, d, e, f, g, h, i, j) VAR_TO_STR(a) + "," + VARS_TO_STR9(b, c, d, e, f, g, h, i, j)
 #define VARS_TO_STR11(a, b, c, d, e, f, g, h, i, j, k) VAR_TO_STR(a) + "," + VARS_TO_STR10(b, c, d, e, f, g, h, i, j, k)
 
+#ifdef GGML_USE_SYCL
+static bool inline _isinf(float f) {
+    return (*(uint32_t *)&f & 0x7fffffff) == 0x7f800000;
+}
+#else
+static bool inline _isinf(float f) { return std::isinf(f); }
+#endif
 
 // accept FLT_MAX as infinity
 static bool isinf_or_max(float f) {
-    return std::isinf(f) || f == FLT_MAX || f == -FLT_MAX;
+    return _isinf(f) || f == FLT_MAX || f == -FLT_MAX;
 }
 
 static bool ggml_is_view_op(enum ggml_op op) {
@@ -1396,7 +1402,7 @@ struct test_flash_attn_ext : public test_case {
     }
 
     double max_nmse_err() override {
-        return 5e-5;
+        return 5e-4;
     }
 
     test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
@@ -1412,6 +1418,48 @@ struct test_flash_attn_ext : public test_case {
     }
 };
 
+// Attention
+struct test_attn : public test_case {
+    const int64_t hs; // head size
+    const int64_t nh; // num heads
+    const int64_t kv; // kv size
+    const int64_t nb; // batch size
+
+    std::string op_desc(ggml_tensor * t) override {
+        return "ATTN";
+
+        GGML_UNUSED(t);
+    }
+
+    std::string vars() override {
+        return VARS_TO_STR4(hs, nh, kv, nb);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    test_attn(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
+        : hs(hs), nh(nh), kv(kv), nb(nb) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
+        ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
+        ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, hs, nh, 1); // transposed
+        ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, kv, nb, 1, 1);
+
+        struct ggml_tensor * cur;
+
+        cur = ggml_mul_mat     (ctx, k, q);
+        cur = ggml_soft_max_ext(ctx, cur, mask, 1.0f/sqrtf(hs));
+        cur = ggml_mul_mat     (ctx, v, cur);
+        cur = ggml_permute     (ctx, cur, 0, 2, 1, 3);
+        cur = ggml_cont_2d     (ctx, cur, hs*nh, nb);
+
+        return cur;
+    }
+};
+
 // Mixtral MOE
 struct test_moe : public test_case {
     const int n_experts;
@@ -1678,9 +1726,16 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_leaky_relu());
 
-    test_cases.emplace_back(new test_flash_attn_ext(128, 32, 256, 8));
+    for (int hs : { 64,  80,  96, 112, 128, 256, }) {
-    test_cases.emplace_back(new test_flash_attn_ext(128, 32, 256, 7));
+        for (int nh : { 32, }) {
-    test_cases.emplace_back(new test_flash_attn_ext(128, 32, 256, 1));
+            for (int kv : { 512, 1024, 2048, 4096, }) {
+                for (int nb : { 1, 2, 4, 8, 512, 1024, 2048, }) {
+                    test_cases.emplace_back(new test_attn          (hs, nh, kv, nb));
+                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb));
+                }
+            }
+        }
+    }
 
 #if !defined(__SANITIZE_THREAD__)
     // FIXME: these tests use too much memory with thread sanitizer

tests/test-llama-grammar.cpp

@@ -190,7 +190,6 @@ int main()
         index++;
     }
 
-    std::vector<std::vector<const llama_grammar_element *>> next_stacks;
     std::vector<llama_grammar_candidate> next_candidates;
     next_candidates.resize(24);
 

tests/test-model-load-cancel.cpp

@@ -0,0 +1,27 @@
+#include "llama.h"
+#include "get-model.h"
+
+#include <cstdlib>
+
+int main(int argc, char *argv[] ) {
+    auto * model_path = get_model_or_exit(argc, argv);
+    auto * file = fopen(model_path, "r");
+    if (file == nullptr) {
+        fprintf(stderr, "no model at '%s' found\n", model_path);
+        return EXIT_FAILURE;
+    }
+
+    fprintf(stderr, "using '%s'\n", model_path);
+    fclose(file);
+
+    llama_backend_init(false);
+    auto params = llama_model_params{};
+    params.use_mmap = false;
+    params.progress_callback = [](float progress, void * ctx){
+        (void) ctx;
+        return progress > 0.50;
+    };
+    auto * model = llama_load_model_from_file(model_path, params);
+    llama_backend_free();
+    return model == nullptr ? EXIT_SUCCESS : EXIT_FAILURE;
+}

tests/test-sampling.cpp

@@ -5,11 +5,10 @@
 #undef NDEBUG
 #endif
 
-#include <cmath>
-#include <numeric>
-#include <cassert>
-#include <vector>
 #include <algorithm>
+#include <cmath>
+#include <string>
+#include <vector>
 
 static void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
@@ -20,11 +19,11 @@ static 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)
 
 static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -41,11 +40,11 @@ static void test_top_k(const std::vector<float> & probs, const std::vector<float
 }
 
 static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -62,11 +61,11 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
 }
 
 static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -81,12 +80,33 @@ static void test_tfs(const std::vector<float> & probs, const std::vector<float>
     }
 }
 
-static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
+static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
+        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    }
+
+    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+    DUMP(&candidates_p);
+    llama_sample_min_p(nullptr, &candidates_p, p, 1);
+    DUMP(&candidates_p);
+    llama_sample_softmax(nullptr, &candidates_p);
+
+    GGML_ASSERT(candidates_p.size == expected_probs.size());
+    for (size_t i = 0; i < candidates_p.size; i++) {
+        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
+    }
+}
+
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
+    const size_t n_vocab = probs.size();
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -107,11 +127,11 @@ static void test_repetition_penalties(
 ) {
     GGML_ASSERT(probs.size() == expected_probs.size());
 
-    size_t n_vocab = probs.size();
+    const size_t n_vocab = probs.size();
     std::vector<llama_token_data> candidates;
     candidates.reserve(n_vocab);
     for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        float logit = log(probs[token_id]);
+        const float logit = logf(probs[token_id]);
         candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
     }
 
@@ -128,6 +148,88 @@ static void test_repetition_penalties(
     }
 }
 
+static void test_sampler_queue(
+    const size_t n_vocab, const std::string samplers_sequence, const int top_k, const float top_p, const float min_p
+) {
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+        const float logit = logf(token_id);
+        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    }
+
+    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+          llama_token min_token_id = 0;
+    const llama_token max_token_id = n_vocab-1;
+
+    for (auto s : samplers_sequence) {
+        switch (s){
+            case 'k': llama_sample_top_k    (nullptr, &candidates_p, top_k, 1); break;
+            case 'f': GGML_ASSERT(false && "tail_free test not implemented");   break;
+            case 'y': GGML_ASSERT(false && "typical test not implemented");     break;
+            case 'p': llama_sample_top_p    (nullptr, &candidates_p, top_p, 1); break;
+            case 'm': llama_sample_min_p    (nullptr, &candidates_p, min_p, 1); break;
+            case 't': GGML_ASSERT(false && "temperature test not implemented"); break;
+            default : GGML_ASSERT(false && "Unknown sampler");                  break;
+        }
+
+        llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
+
+        const int size = candidates_p.size;
+
+        if (s == 'k') {
+            const int expected_size = std::min(size, top_k);
+            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k));
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else if (s == 'p') {
+            const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2;
+            const int softmax_numerator_target = ceilf(top_p * softmax_divisor);
+
+                min_token_id  = n_vocab;
+            int expected_size = 0;
+            int cumsum        = 0;
+            do { // do-while because always at least one token is sampled
+                min_token_id--;
+                expected_size++;
+
+                cumsum += min_token_id;
+            } while (cumsum < softmax_numerator_target);
+
+            // token 0 has p == 0, need special consideration for cumsum because top_p immediately returns
+            if (min_token_id == 1) {
+                min_token_id--;
+                expected_size += 1;
+            }
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else if (s == 'm') {
+            int expected_size = ceilf((1.0f-min_p) * n_vocab);
+            expected_size = std::max(expected_size, 1);
+            expected_size = std::min(expected_size, size);
+
+            min_token_id = floorf(min_p * n_vocab);
+            min_token_id = std::max(min_token_id, 1);
+            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size));
+            min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
+
+            GGML_ASSERT(size == expected_size);
+            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
+            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+        } else {
+            GGML_ASSERT(false);
+        }
+    }
+
+    printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n",
+           samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
+}
+
 int main(void) {
     ggml_time_init();
 
@@ -139,6 +241,15 @@ int main(void) {
     test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f);
     test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
 
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.26f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.49f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.51f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.74f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  0.76f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.00f);
+
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
     test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
@@ -154,6 +265,34 @@ int main(void) {
     test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
     test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
 
+    test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
+
+    test_sampler_queue(10000, "k",   100, 1.0000f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f);
+    test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f);
+
+    test_sampler_queue(10000, "kp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "km", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mp", 100, 0.8f, 9997.9f/9999.0f);
+    test_sampler_queue(10000, "mp", 100, 0.8f, 0.1f);
+
+    test_sampler_queue(10000, "kpm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "kmp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pkm", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "pmk", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mkp", 100, 0.8f, 0.1f);
+    test_sampler_queue(10000, "mpk", 100, 0.8f, 0.1f);
+
     printf("OK\n");
 
     return 0;

unicode.h

@@ -2,8 +2,9 @@
 
 #include <cassert>
 #include <stdexcept>
-#include <vector>
+#include <string>
 #include <unordered_map>
+#include <vector>
 
 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},