diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile
index 2b3a20c63..a75bc976f 100644
--- a/.devops/full.Dockerfile
+++ b/.devops/full.Dockerfile
@@ -6,7 +6,8 @@ RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip
 
 RUN pip install --upgrade pip setuptools wheel \
-    && pip install numpy requests sentencepiece torch tqdm
+    && pip install numpy requests sentencepiece tqdm \
+    && pip install torch --index-url https://download.pytorch.org/whl/cpu
 
 WORKDIR /app
 
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index e9826a735..88e70e495 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -8,10 +8,10 @@ on:
         required: true
         type: boolean
   push:
-    paths: ['.github/workflows/**', 'CMakeLists.txt', 'Makefile', '**.h', '*.c', '**.cpp']
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
   pull_request:
     types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
-    paths: ['CMakeLists.txt', 'Makefile', '**.h', '*.c', '**.cpp']
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
 
 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -62,7 +62,43 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --output-on-failure
+          ctest --verbose
+
+  ubuntu-latest-cmake-sanitizer:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
+        accelerate: [ON, OFF]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_ACCELERATE=${{ matrix.accelerate }}
+          cmake --build . --config ${{ matrix.build_type }}
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest --verbose
 
   macOS-latest-make:
     runs-on: macos-latest
@@ -107,11 +143,21 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --output-on-failure
+          ctest --verbose
 
   windows-latest-cmake:
     runs-on: windows-latest
 
+    strategy:
+      matrix:
+        include:
+         - build: 'avx2'
+           defines: ''
+         - build: 'avx'
+           defines: '-DLLAMA_AVX2=OFF'
+         - build: 'avx512'
+           defines: '-DLLAMA_AVX512=ON'
+
     steps:
       - name: Clone
         id: checkout
@@ -122,14 +168,28 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. ${{ matrix.defines }}
           cmake --build . --config Release
 
+      - name: Check AVX512F support
+        id: check_avx512f
+        if: ${{ matrix.build == 'avx512' }}
+        continue-on-error: true
+        run: |
+          cd build
+          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
+          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
+          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
+          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
+          & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
+          .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
+
       - name: Test
         id: cmake_test
+        if: ${{ matrix.build != 'avx512' || env.HAS_AVX512F == '1' }} # Test AVX-512 only when possible
         run: |
           cd build
-          ctest -C Release --output-on-failure
+          ctest -C Release --verbose
 
       - name: Get commit hash
         id: commit
@@ -140,12 +200,39 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-x64.zip .\build\bin\Release\*
+          7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
+
+  release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - ubuntu-latest-make
+      - ubuntu-latest-cmake
+      - macOS-latest-make
+      - macOS-latest-cmake
+      - windows-latest-cmake
+
+    steps:
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v3
+
+      - name: Get commit hash
+        id: commit
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Create release
         id: create_release
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: zendesk/action-create-release@v1
+        uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
@@ -153,15 +240,25 @@ jobs:
 
       - name: Upload release
         id: upload_release
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        uses: actions/github-script@v3
         with:
-          upload_url: ${{ steps.create_release.outputs.upload_url }} 
-          asset_path: .\llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-x64.zip
-          asset_name: llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-x64.zip
-          asset_content_type: application/octet-stream
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/${file}`)
+                });
+              }
+            }
 
 #  ubuntu-latest-gcc:
 #    runs-on: ubuntu-latest
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index d1a43caa6..f70821de2 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -49,6 +49,7 @@ jobs:
         with:
           context: .
           push: true
+          platforms: linux/amd64,linux/arm64
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
@@ -57,5 +58,6 @@ jobs:
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
+          platforms: linux/amd64,linux/arm64
           tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}
\ No newline at end of file
diff --git a/.gitignore b/.gitignore
index ce01fd541..1c75d38d1 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,6 +5,7 @@
 .vscode/
 .DS_Store
 
+.build/
 build/
 build-em/
 build-debug/
@@ -20,9 +21,15 @@ models/*
 /quantize
 /result
 /perplexity
+/embedding
+/Pipfile
 
 arm_neon.h
 compile_commands.json
 
 .envrc
 .direnv/
+
+.venv
+__pycache__
+.swiftpm
diff --git a/CMakeLists.txt b/CMakeLists.txt
index a1ff5a44e..1a434f07b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -54,6 +54,7 @@ option(LLAMA_SANITIZE_UNDEFINED     "llama: enable undefined sanitizer"
 # instruction set specific
 option(LLAMA_AVX                    "llama: enable AVX"                                     ON)
 option(LLAMA_AVX2                   "llama: enable AVX2"                                    ON)
+option(LLAMA_AVX512                 "llama: enable AVX512"                                  OFF)
 option(LLAMA_FMA                    "llama: enable FMA"                                     ON)
 
 # 3rd party libs
@@ -67,7 +68,9 @@ option(LLAMA_BUILD_EXAMPLES         "llama: build examples" ${LLAMA_STANDALONE})
 # Compile flags
 #
 
+set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
+set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
@@ -75,14 +78,17 @@ find_package(Threads REQUIRED)
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
+        link_libraries(-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        link_libraries(-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
+        link_libraries(-fsanitize=undefined)
     endif()
 endif()
 
@@ -120,8 +126,9 @@ if (LLAMA_ALL_WARNINGS)
             -Wall
             -Wextra
             -Wpedantic
-            -Wshadow
             -Wcast-qual
+            -Wdouble-promotion
+            -Wshadow
             -Wstrict-prototypes
             -Wpointer-arith
             -Wno-unused-function
@@ -131,6 +138,7 @@ if (LLAMA_ALL_WARNINGS)
             -Wextra
             -Wpedantic
             -Wcast-qual
+            -Wno-unused-function
         )
     else()
         # todo : msvc
@@ -185,7 +193,9 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     message(STATUS "x86 detected")
     if (MSVC)
-        if (LLAMA_AVX2)
+        if (LLAMA_AVX512)
+            add_compile_options(/arch:AVX512)
+        elseif (LLAMA_AVX2)
             add_compile_options(/arch:AVX2)
         elseif (LLAMA_AVX)
             add_compile_options(/arch:AVX)
@@ -201,6 +211,12 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
         if (LLAMA_AVX2)
             add_compile_options(-mavx2)
         endif()
+        if (LLAMA_AVX512)
+            add_compile_options(-mavx512f)
+            # add_compile_options(-mavx512cd)
+            # add_compile_options(-mavx512dq)
+            # add_compile_options(-mavx512bw)
+        endif()
     endif()
 else()
     # TODO: support PowerPC
@@ -239,7 +255,7 @@ endif()
 #
 
 if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    enable_testing()
+    include(CTest)
     add_subdirectory(tests)
 endif ()
 
diff --git a/Makefile b/Makefile
index 98a2d85f3..2f828bf10 100644
--- a/Makefile
+++ b/Makefile
@@ -35,6 +35,10 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# warnings
+CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -66,92 +70,8 @@ endif
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
-	ifeq ($(UNAME_S),Darwin)
-		CFLAGS += -mf16c
-		AVX1_M := $(shell sysctl machdep.cpu.features)
-		ifneq (,$(findstring FMA,$(AVX1_M)))
-			CFLAGS += -mfma
-		endif
-		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysctl machdep.cpu.leaf7_features)
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-	else ifeq ($(UNAME_S),Linux)
-		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell grep "fma " /proc/cpuinfo)
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-		endif
-		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
-		ifneq (,$(findstring sse3,$(SSE3_M)))
-			CFLAGS += -msse3
-		endif
-		AVX512F_M := $(shell grep "avx512f " /proc/cpuinfo)
-		ifneq (,$(findstring avx512f,$(AVX512F_M)))
-			CFLAGS += -mavx512f
-		endif
-		AVX512BW_M := $(shell grep "avx512bw " /proc/cpuinfo)
-		ifneq (,$(findstring avx512bw,$(AVX512BW_M)))
-			CFLAGS += -mavx512bw
-		endif
-		AVX512DQ_M := $(shell grep "avx512dq " /proc/cpuinfo)
-		ifneq (,$(findstring avx512dq,$(AVX512DQ_M)))
-			CFLAGS += -mavx512dq
-		endif
-		AVX512VL_M := $(shell grep "avx512vl " /proc/cpuinfo)
-		ifneq (,$(findstring avx512vl,$(AVX512VL_M)))
-			CFLAGS += -mavx512vl
-		endif
-		AVX512CD_M := $(shell grep "avx512cd " /proc/cpuinfo)
-		ifneq (,$(findstring avx512cd,$(AVX512CD_M)))
-			CFLAGS += -mavx512cd
-		endif
-		AVX512ER_M := $(shell grep "avx512er " /proc/cpuinfo)
-		ifneq (,$(findstring avx512er,$(AVX512ER_M)))
-			CFLAGS += -mavx512er
-		endif
-		AVX512IFMA_M := $(shell grep "avx512ifma " /proc/cpuinfo)
-		ifneq (,$(findstring avx512ifma,$(AVX512IFMA_M)))
-			CFLAGS += -mavx512ifma
-		endif
-		AVX512PF_M := $(shell grep "avx512pf " /proc/cpuinfo)
-		ifneq (,$(findstring avx512pf,$(AVX512PF_M)))
-			CFLAGS += -mavx512pf
-		endif
-	else ifeq ($(UNAME_S),Haiku)
-		AVX1_M := $(shell sysinfo -cpu | grep -w "AVX")
-		ifneq (,$(findstring AVX,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysinfo -cpu | grep -w "AVX2")
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell sysinfo -cpu | grep -w "FMA")
-		ifneq (,$(findstring FMA,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell sysinfo -cpu | grep -w "F16C")
-		ifneq (,$(findstring F16C,$(F16C_M)))
-			CFLAGS += -mf16c
-		endif
-	else
-		CFLAGS += -mfma -mf16c -mavx -mavx2
-	endif
+	# Use all CPU extensions that are available:
+	CFLAGS += -march=native -mtune=native
 endif
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
@@ -212,7 +132,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main quantize perplexity
+default: main quantize perplexity embedding
 
 #
 # Build library
@@ -228,7 +148,7 @@ common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o
 
 clean:
-	rm -vf *.o main quantize perplexity
+	rm -vf *.o main quantize perplexity embedding
 
 main: examples/main/main.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS)
@@ -242,6 +162,9 @@ quantize: examples/quantize/quantize.cpp ggml.o llama.o
 perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
 	$(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS)
 
+embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
+	$(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS)
+
 #
 # Tests
 #
diff --git a/Package.swift b/Package.swift
new file mode 100644
index 000000000..79d13c82d
--- /dev/null
+++ b/Package.swift
@@ -0,0 +1,20 @@
+// swift-tools-version:5.3
+
+import PackageDescription
+
+let package = Package(
+    name: "llama",
+    products: [
+        .library(name: "llama", targets: ["llama"]),
+    ],
+    targets: [
+        .target(
+            name: "llama",
+            path: ".",
+            sources: ["ggml.c", "llama.cpp"],
+            publicHeadersPath: "spm-headers",
+            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"])]
+        ),
+    ],
+    cxxLanguageStandard: .cxx11
+)
diff --git a/README.md b/README.md
index 9ba6241da..508d315d5 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,7 @@
 # llama.cpp
 
+![llama](https://user-images.githubusercontent.com/1991296/227761327-6d83e30e-2200-41a6-bfbb-f575231c54f4.png)
+
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
@@ -8,9 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 **Hot topics:**
 
 - [Roadmap (short-term)](https://github.com/ggerganov/llama.cpp/discussions/457)
-- New C-style API is now available: https://github.com/ggerganov/llama.cpp/pull/370
-- Cache input prompts for faster initialization: https://github.com/ggerganov/llama.cpp/issues/64
-- Create a `llama.cpp` logo: https://github.com/ggerganov/llama.cpp/issues/105
+- Support for [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 
 ## Description
 
@@ -35,6 +35,14 @@ Supported platforms:
 - [X] Windows (via CMake)
 - [X] Docker
 
+Supported models:
+
+- [X] LLaMA 🦙
+- [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
+- [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
+- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
+- [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+
 ---
 
 Here is a typical run using LLaMA-7B:
@@ -147,8 +155,8 @@ python3 -m pip install torch numpy sentencepiece
 # convert the 7B model to ggml FP16 format
 python3 convert-pth-to-ggml.py models/7B/ 1
 
-# quantize the model to 4-bits
-python3 quantize.py 7B
+# quantize the model to 4-bits (using method 2 = q4_0)
+./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin 2
 
 # run the inference
 ./main -m ./models/7B/ggml-model-q4_0.bin -n 128
@@ -220,6 +228,21 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 > 
 ```
 
+### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
+
+- Obtain the `gpt4all-lora-quantized.bin` model
+- It is distributed in the old `ggml` format which is now obsoleted
+- You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to
+convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
+
+  ```bash
+  python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model 
+  python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin
+  ```
+  
+- You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models
+- The original model is saved in the same folder with a suffix `.orig`
+
 ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
 
 - **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.**
@@ -246,7 +269,7 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
     
 ### Perplexity (Measuring model quality)
 
-You can pass `--perplexity` as a command line option to measure perplexity over the given prompt.  For more background,
+You can use the `perplexity` example to measure perplexity over the given prompt.  For more background,
 see https://huggingface.co/docs/transformers/perplexity.  However, in general, lower perplexity is better for LLMs.
 
 #### Latest measurements
@@ -269,10 +292,10 @@ Perplexity - model options
 #### How to run
 
 1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
-2. Run `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+2. Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
 3. Output:
 ```
-Calculating perplexity over 655 chunks
+perplexity : calculating perplexity over 655 chunks
 24.43 seconds per pass - ETA 4.45 hours
 [1]4.5970,[2]5.1807,[3]6.0382,...
 ```
@@ -280,7 +303,7 @@ And after 4.45 hours, you will have the final perplexity.
 
 ### Android
 
-You can easily run `llama.cpp` on Android device with [termux](https://play.google.com/store/apps/details?id=com.termux).
+You can easily run `llama.cpp` on Android device with [termux](https://termux.dev/).
 First, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
 ```
 $ mkdir build-android
@@ -289,7 +312,7 @@ $ export NDK=<your_ndk_directory>
 $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
 $ make
 ```
-Install [termux](https://play.google.com/store/apps/details?id=com.termux) on your device and run `termux-setup-storage` to get access to your SD card.
+Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
 Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
diff --git a/convert-ggml-to-pth.py b/convert-ggml-to-pth.py
new file mode 100644
index 000000000..25a44237a
--- /dev/null
+++ b/convert-ggml-to-pth.py
@@ -0,0 +1,299 @@
+# Author: github.com/ductai199x
+import argparse
+import os
+import struct
+
+import numpy as np
+import torch
+from numba import njit
+from tqdm.auto import tqdm
+
+
+def read_header(fin):
+    values = struct.unpack("i" * 9, fin.read(4 * 9))
+    _, _, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype = values
+    return {
+        "vocab_size": vocab_size,
+        "dim": dim,
+        "multiple_of": multiple_of,
+        "n_heads": n_heads,
+        "n_layers": n_layers,
+    }, ftype
+
+
+def read_tokens(fin, vocab_size):
+    tokens = []
+    for _ in range(vocab_size):
+        text_len = struct.unpack("i", fin.read(4))[0]
+        text_bytes = fin.read(text_len)
+        try:
+            text = text_bytes.decode()
+        except UnicodeDecodeError:
+            text = text_bytes.decode(errors="replace")
+        score = struct.unpack("f", fin.read(4))[0]
+        tokens.append((text, score))
+    return tokens
+
+
+@njit
+def dequantize_weights_numba(fin_data, n_rows, n_cols):
+    qk = 32
+    nb = n_cols // qk
+    bs = 4 + (qk // 2)
+
+    weights = np.zeros((n_rows, n_cols), dtype=np.float32)
+    data_pos = 0
+
+    for row in range(n_rows):
+        for block in range(nb):
+            d = np.frombuffer(fin_data[data_pos : data_pos + 4], dtype=np.float32)[0]
+            data_pos += 4
+            packed_values = fin_data[data_pos : data_pos + (qk // 2)]
+            data_pos += qk // 2
+
+            for i in range(qk // 2):
+                packed_value = packed_values[i]
+                v0 = np.float32((packed_value & 0b00001111) - 8) * d
+                v1 = np.float32((packed_value >> 4) - 8) * d
+
+                weights[row, block * qk + 2 * i] = v0
+                weights[row, block * qk + 2 * i + 1] = v1
+
+    return weights
+
+
+def dequantize_weights(fin, n_rows, n_cols):
+    qk = 32
+    nb = n_cols // qk
+    data_size = n_rows * n_cols // 2 + n_rows * nb * 4
+    fin_data = fin.read(data_size)
+    return dequantize_weights_numba(fin_data, n_rows, n_cols)
+
+
+def read_variables(fin):
+    model = {}
+    pbar = tqdm(total=os.path.getsize(fin.name), unit="B", unit_scale=True, desc="Reading variables")
+    while True:
+        start_pos = fin.tell()
+        try:
+            n_dims, name_length, ftype_cur = struct.unpack("iii", fin.read(4 * 3))
+        except struct.error:
+            break
+
+        shape = tuple(struct.unpack("i" * n_dims, fin.read(4 * n_dims)))
+        shape = shape[::-1]
+        name = fin.read(name_length).decode()
+
+        # ensure tensor data is aligned
+        tensor_data_offset = fin.tell()
+        tensor_data_offset = (tensor_data_offset + 31) & -32
+        fin.seek(tensor_data_offset)
+
+        if ftype_cur == 2:
+            # 4-bit quantized weights
+            dtype = np.uint8
+            data = dequantize_weights(fin, shape[0], shape[1])
+            data = data.reshape(shape)
+        elif ftype_cur == 0:
+            dtype = np.float32
+            data_size = np.prod(shape)
+            data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
+        elif ftype_cur == 1:
+            dtype = np.float16
+            data_size = np.prod(shape)
+            data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape)
+
+        model[name] = torch.tensor(data, dtype=torch.float32 if dtype == np.float32 else torch.float16)
+
+        pbar.update(fin.tell() - start_pos)
+
+    return model
+
+
+def convert_to_hf_format(model, hparams):
+    # This works for llama 7B, need to test with other models
+    n_layers = hparams["n_layers"]
+    n_heads = hparams["n_heads"]
+    dim = hparams["dim"]
+    dims_per_head = dim // n_heads
+    base = 10000.0
+    inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
+
+    # permute for sliced rotary
+    def permute(w):
+        return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim)
+
+    state_dict = {}
+    for layer_i in range(n_layers):
+        state_dict.update(
+            {
+                f"model.layers.{layer_i}.self_attn.q_proj.weight": permute(
+                    model[f"layers.{layer_i}.attention.wq.weight"]
+                ),
+                f"model.layers.{layer_i}.self_attn.k_proj.weight": permute(
+                    model[f"layers.{layer_i}.attention.wk.weight"]
+                ),
+                f"model.layers.{layer_i}.self_attn.v_proj.weight": model[
+                    f"layers.{layer_i}.attention.wv.weight"
+                ],
+                f"model.layers.{layer_i}.self_attn.o_proj.weight": model[
+                    f"layers.{layer_i}.attention.wo.weight"
+                ],
+                f"model.layers.{layer_i}.mlp.gate_proj.weight": model[
+                    f"layers.{layer_i}.feed_forward.w1.weight"
+                ],
+                f"model.layers.{layer_i}.mlp.down_proj.weight": model[
+                    f"layers.{layer_i}.feed_forward.w2.weight"
+                ],
+                f"model.layers.{layer_i}.mlp.up_proj.weight": model[
+                    f"layers.{layer_i}.feed_forward.w3.weight"
+                ],
+                f"model.layers.{layer_i}.input_layernorm.weight": model[
+                    f"layers.{layer_i}.attention_norm.weight"
+                ],
+                f"model.layers.{layer_i}.post_attention_layernorm.weight": model[
+                    f"layers.{layer_i}.ffn_norm.weight"
+                ],
+            }
+        )
+        state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
+    state_dict.update(
+        {
+            "model.embed_tokens.weight": model["tok_embeddings.weight"],
+            "model.norm.weight": model["norm.weight"],
+            "lm_head.weight": model["output.weight"],
+        }
+    )
+
+    return state_dict
+
+
+def chat(model, hparams, llama_dir):
+    from transformers import (GenerationConfig, LlamaForCausalLM,
+                              LlamaTokenizer, StoppingCriteria,
+                              StoppingCriteriaList)
+    from transformers.models.llama.configuration_llama import LlamaConfig
+
+    class StoppingCriteriaSub(StoppingCriteria):
+        def __init__(self):
+            super().__init__()
+
+        def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops=[]):
+            print(tokenizer.decode(input_ids[0]), end="", flush=True)
+            if input_ids[0][-1] == 13:
+                return True
+
+            return False
+
+    config = LlamaConfig(
+        vocab_size=hparams["vocab_size"],
+        dim=hparams["dim"],
+        num_hidden_layers=hparams["n_layers"],
+        num_attention_heads=hparams["n_heads"],
+    )
+
+    llama = LlamaForCausalLM(config=config)
+    llama.load_state_dict(state_dict=model, strict=True)
+    tokenizer = LlamaTokenizer.from_pretrained(llama_dir)
+
+    device = torch.device("cpu")
+    llama = llama.to(device)
+
+    ctx = """You are AI.
+This is a dialog, where User interacts with AI. AI is helpful, kind, obedient, honest, respectful, direct, concise, should try to protect User's privacy, and knows its own limits. Also, AI must answer User and AI cannot stop the conversation by itself.
+User: Hello, AI.
+AI: Hello! How can I assist you today?
+"""
+    print(ctx.rstrip("\n"))
+    while True:
+        print("-" * 60)
+        prompt = input("User: ")
+        if ctx != "":
+            ctx = f"{ctx}User: {prompt}\n"
+        else:
+            ctx = f"{prompt}\nAI:"
+
+        ctx = (ctx[-1920:]) if len(ctx) >= 2048 else ctx
+
+        print("-" * 60)
+        if len(ctx.strip()) > 0:
+            input_ids = tokenizer(ctx, return_tensors="pt")["input_ids"].to(device)
+            generation_config = GenerationConfig(
+                temperature=0.8,
+                top_p=0.95,
+                top_k=50,
+                repetition_penalty=1.1764,
+            )
+            with torch.no_grad():
+                generation_output = llama.generate(
+                    input_ids=input_ids,
+                    generation_config=generation_config,
+                    return_dict_in_generate=True,
+                    output_scores=True,
+                    max_length=2048,
+                    do_sample=True,
+                    stopping_criteria=StoppingCriteriaList([StoppingCriteriaSub()]),
+                )
+            s = generation_output.sequences[0]
+            decoded = tokenizer.decode(s)
+            ctx = f"{decoded}\n"
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input_dir", "-i", type=str, required=True, help="The input directory containing the ggml files."
+    )
+    parser.add_argument(
+        "--prefix",
+        "-p",
+        type=str,
+        required=True,
+        help="The prefix of the ggml files (ggml-model-f16 or ggml-model-q4_0).",
+    )
+    parser.add_argument(
+        "--hf",
+        action="store_true",
+        help="Whether to save the model in the Hugging Face format. (default: False)",
+    )
+    parser.add_argument(
+        "--chat", "-c", action="store_true", help="Whether to open a chat with the model. (default: False)"
+    )
+    args = parser.parse_args()
+
+    llama_dir = os.path.abspath(f"{args.input_dir}/../")
+
+    ggml_files = sorted(
+        [f"{args.input_dir}/{f}" for f in os.listdir(args.input_dir) if f.startswith(args.prefix)]
+    )
+
+    fin = open(ggml_files[0], "rb")
+    hparams, ftype = read_header(fin)
+    tokens = read_tokens(fin, hparams["vocab_size"])
+    model = read_variables(fin)
+
+    for f in tqdm(ggml_files[1:]):
+        fin = open(f, "rb")
+        read_header(fin)
+        read_tokens(fin, hparams["vocab_size"])
+        model.update(read_variables(fin))
+
+    if args.hf:
+        model = convert_to_hf_format(model, hparams)
+
+    pth_ckpt = {
+        "state_dict": model,
+        "hparams": hparams,
+        "tokens": tokens,
+    }
+
+    torch.save(pth_ckpt, f"{args.input_dir}/{args.prefix}-to-torch.pth")
+
+    if args.chat:
+        if not args.hf:
+            model = convert_to_hf_format(model, hparams)
+        chat(model, hparams, llama_dir)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/convert-gpt4all-to-ggml.py b/convert-gpt4all-to-ggml.py
new file mode 100644
index 000000000..b1a5e0560
--- /dev/null
+++ b/convert-gpt4all-to-ggml.py
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+
+#
+# TODO: deduplicate GPT4All with convert-unversioned-ggml-to-ggml.py
+#
+
+# Original by https://github.com/eiz
+# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
+import argparse
+import glob
+import os
+import struct
+import sys
+from sentencepiece import SentencePieceProcessor
+
+HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Upgrade a GPT4All model to the current format')
+    parser.add_argument('gpt4all_model', help='path to gpt4all-lora-quantized.bin')
+    parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
+    return parser.parse_args()
+
+def read_header(f_in):
+    struct_fmt = "i" * (3 + len(HPARAMS))
+    struct_size = struct.calcsize(struct_fmt)
+    buf = f_in.read(struct_size)
+    return struct.unpack(struct_fmt, buf)
+
+def write_header(f_out, header):
+    (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
+
+    if magic != 0x67676d6c:
+        raise Exception('Invalid file magic. Must be an old style ggml file.')
+
+    values = [
+        0x67676d66, # magic: ggml in hex
+        1,          # file version
+        vocab_size,
+        dim,
+        multiple_of,
+        n_heads,
+        n_layers,
+        rot,
+        ftype
+    ]
+    f_out.write(struct.pack("i" * len(values), *values))
+
+def write_tokens(fout, tokenizer):
+    for i in range(tokenizer.vocab_size()):
+        if tokenizer.is_unknown(i):
+            text = " \u2047 ".encode()
+        elif tokenizer.is_control(i):
+            text = b""
+        elif tokenizer.is_byte(i):
+            piece = tokenizer.id_to_piece(i)
+            if len(piece) != 6:
+                print(f"Invalid token: {piece}")
+                sys.exit(1)
+            byte_value = int(piece[3:-1], 16)
+            text = struct.pack("B", byte_value)
+        else:
+            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
+        fout.write(struct.pack("i", len(text)))
+        fout.write(text)
+        fout.write(struct.pack("f", tokenizer.get_score(i)))
+
+    # TODO: GPT4All - add extra <pad> token
+    text = "<pad>".encode()
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    fout.write(struct.pack("f", 0.0))
+
+def read_tokens(f_in, tokenizer):
+    for i in range(tokenizer.vocab_size()):
+        len_b = f_in.read(4)
+        (length,) = struct.unpack("i", len_b)
+        f_in.read(length)
+
+def copy_all_data(f_out, f_in):
+    while True:
+        buf = f_in.read(1024 * 1024)
+        if not buf:
+            break
+        f_out.write(buf)
+
+def convert_one_file(path_in, tokenizer):
+    path_tmp = f"{path_in}.tmp"
+    path_orig= f"{path_in}.orig"
+    print(f"converting {path_in}")
+    with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
+        write_header(f_out, read_header(f_in))
+        read_tokens(f_in, tokenizer)
+        write_tokens(f_out, tokenizer)
+        copy_all_data(f_out, f_in)
+    os.rename(path_in, path_orig)
+    os.rename(path_tmp, path_in)
+
+def main():
+    args = parse_args()
+
+    tokenizer = SentencePieceProcessor(args.tokenizer_model)
+
+    convert_one_file(args.gpt4all_model, tokenizer)
+
+if __name__ == "__main__":
+    main()
diff --git a/convert-gptq-to-ggml.py b/convert-gptq-to-ggml.py
index 6c77808fc..42e99c2ff 100644
--- a/convert-gptq-to-ggml.py
+++ b/convert-gptq-to-ggml.py
@@ -50,7 +50,7 @@ fout.write(struct.pack("i", 4))
 # This loop unchanged from convert-pth-to-ggml.py:
 for i in range(tokenizer.vocab_size()):
     if tokenizer.is_unknown(i):
-        text = " \u2047 ".encode("utf-8")
+        text = " \u2047 ".encode()
     elif tokenizer.is_control(i):
         text = b""
     elif tokenizer.is_byte(i):
@@ -61,21 +61,26 @@ for i in range(tokenizer.vocab_size()):
         byte_value = int(piece[3:-1], 16)
         text = struct.pack("B", byte_value)
     else:
-        text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
+        text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
     fout.write(struct.pack("i", len(text)))
     fout.write(text)
     fout.write(struct.pack("f", tokenizer.get_score(i)))
 
 def write_header(shape, dst_name, ftype_cur):
-    sname = dst_name.encode('utf-8')
+    sname = dst_name.encode()
     fout.write(struct.pack("iii", len(shape), len(sname), ftype_cur))
     fout.write(struct.pack("i" * len(shape), *shape[::-1]))
     fout.write(sname)
 
+    # ensure tensor data is aligned
+    tensor_data_offset = fout.tell()
+    tensor_data_offset = (tensor_data_offset + 31) & -32
+    fout.seek(tensor_data_offset)
+
 def convert_non_q4(src_name, dst_name):
     v = model[src_name]
     shape = v.shape
-    print("Processing non-Q4 variable: " + src_name + " with shape: ", shape, " and type: ", v.dtype)
+    print(f"Processing non-Q4 variable: {src_name} with shape: {shape} and type: {v.dtype}")
     if len(shape) == 1:
         print("  Converting to float32")
         v = v.to(torch.float32)
@@ -100,7 +105,7 @@ def convert_q4(src_name, dst_name, permute=False):
     # Each int32 item is actually 8 int4 items packed together, and it's transposed.
     shape = (qweight.shape[0], qweight.shape[1] * 8)
 
-    print("Processing Q4 variable: " + src_name + " with shape: ", shape)
+    print(f"Processing Q4 variable: {src_name} with shape: {shape}")
 
     # The output format has the int4 weights in groups of 32 rather than 8.
     # It looks like this:
@@ -163,5 +168,5 @@ for i in range(n_layer):
 
 fout.close()
 
-print("Done. Output file: " + fname_out)
-print("")
+print(f"Done. Output file: {fname_out}")
+print()
diff --git a/convert-pth-to-ggml.py b/convert-pth-to-ggml.py
index f0f6b0ec4..dcef2f6a3 100644
--- a/convert-pth-to-ggml.py
+++ b/convert-pth-to-ggml.py
@@ -1,4 +1,4 @@
-# Convert a LLaMA model checkpoint to a ggml compatible file
+# Convert a LLaMA model checkpoint to a ggjt compatible file
 #
 # Load the model using Torch
 # Iterate over all variables and write them to a binary file.
@@ -24,8 +24,57 @@ import torch
 
 from sentencepiece import SentencePieceProcessor
 
-def parse_args():
+QK = 32
 
+GGML_TYPE_Q4_0  = 0
+GGML_TYPE_Q4_1  = 1
+GGML_TYPE_I8    = 2
+GGML_TYPE_I16   = 3
+GGML_TYPE_I32   = 4
+GGML_TYPE_F16   = 5
+GGML_TYPE_F32   = 6
+
+WTYPES = {
+    0: GGML_TYPE_F32,
+    1: GGML_TYPE_F16,
+    2: GGML_TYPE_Q4_0,
+    3: GGML_TYPE_Q4_1,
+}
+
+GGML_BLCK_SIZE = {
+    GGML_TYPE_Q4_0:  QK,
+    GGML_TYPE_Q4_1:  QK,
+    GGML_TYPE_I8:    1,
+    GGML_TYPE_I16:   1,
+    GGML_TYPE_I32:   1,
+    GGML_TYPE_F16:   1,
+    GGML_TYPE_F32:   1,
+}
+
+GGML_TYPE_SIZE = {
+    GGML_TYPE_Q4_0: 4   + QK//2,
+    GGML_TYPE_Q4_1: 4*2 + QK//2,
+    GGML_TYPE_I8:   1,
+    GGML_TYPE_I16:  2,
+    GGML_TYPE_I32:  4,
+    GGML_TYPE_F16:  2,
+    GGML_TYPE_F32:  4,
+}
+
+def ggml_nelements(shape):
+    r = 1
+    for i in shape:
+        r *= i
+    return r
+
+def ggml_nbytes(shape, ftype):
+    x = ggml_nelements(shape)
+    t = WTYPES[ftype]
+    x *= GGML_TYPE_SIZE[t]
+    x //= GGML_BLCK_SIZE[t]
+    return x
+
+def parse_args():
     parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
     parser.add_argument('dir_model',  help='directory containing the model checkpoint')
     parser.add_argument('ftype',      help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1)
@@ -33,7 +82,6 @@ def parse_args():
     return parser.parse_args()
 
 def get_n_parts(dim):
-
     mappings = {4096: 1, 5120: 2, 6656: 4, 8192: 8}
     n_parts = mappings.get(dim)
     if n_parts is None:
@@ -44,30 +92,24 @@ def get_n_parts(dim):
     return n_parts
 
 def load_hparams_and_tokenizer(dir_model):
-
     # `dir_model` is something like `models/7B` or `models/7B/`.
     # "tokenizer.model" is expected under model's parent dir.
     # When `dir_model` is a symlink, f"{dir_model}/../tokenizer.model" would not be found.
     # Let's use the model's parent dir directly.
     model_parent_dir = os.path.dirname(os.path.normpath(dir_model))
-
     fname_hparams = f"{dir_model}/params.json"
     fname_tokenizer = f"{model_parent_dir}/tokenizer.model"
-
     with open(fname_hparams, "r") as f:
         hparams = json.load(f)
         print(hparams)
-
     tokenizer = SentencePieceProcessor(fname_tokenizer)
     hparams.update({"vocab_size": tokenizer.vocab_size()})
-
     return hparams, tokenizer
 
 def write_header(fout, hparams, ftype):
-
     keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
     values = [
-        0x67676d66,  # magic: ggmf in hex
+        0x67676a74,  # magic: ggjt in hex
         1, # file version
         *[hparams[key] for key in keys],
         hparams["dim"] // hparams["n_heads"],  # rot (obsolete)
@@ -76,10 +118,9 @@ def write_header(fout, hparams, ftype):
     fout.write(struct.pack("i" * len(values), *values))
 
 def write_tokens(fout, tokenizer):
-
     for i in range(tokenizer.vocab_size()):
         if tokenizer.is_unknown(i):
-            text = " \u2047 ".encode("utf-8")
+            text = " \u2047 ".encode()
         elif tokenizer.is_control(i):
             text = b""
         elif tokenizer.is_byte(i):
@@ -90,92 +131,144 @@ def write_tokens(fout, tokenizer):
             byte_value = int(piece[3:-1], 16)
             text = struct.pack("B", byte_value)
         else:
-            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
+            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
         fout.write(struct.pack("i", len(text)))
         fout.write(text)
         fout.write(struct.pack("f", tokenizer.get_score(i)))
 
-def process_and_write_variables(fout, model, ftype):
-
+def process_and_write_variables(fout, model, ftype, part_id, n_parts):
     for name, datao in model.items():
-
         if name.endswith("freqs"):
             continue
 
-        shape = datao.shape
-
-        print(f"Processing variable: {name} with shape: {shape} and type: {datao.dtype}")
-
+        # remove dimensions with a single element
         data = datao.numpy().squeeze()
-        n_dims = len(shape)
+        partshape = data.shape
+        n_dims = len(data.shape)
+        assert n_dims in (1, 2)
 
-        # default type is fp16
+        print(f"Processing variable: {name} with shape: {partshape} and type: {datao.dtype}")
+
+        # coerce single-dimensional tensors from float16 to float32
         ftype_cur = 1
         if ftype == 0 or n_dims == 1:
             print("  Converting to float32")
             data = data.astype(np.float32)
             ftype_cur = 0
+        blck_size = GGML_BLCK_SIZE[WTYPES[ftype_cur]]
+        type_size = GGML_TYPE_SIZE[WTYPES[ftype_cur]]
 
-        # header
-        sname = name.encode('utf-8')
-        fout.write(struct.pack("iii", len(data.shape), len(sname), ftype_cur))
-        for dim in reversed(data.shape):
+        # determine dimension along which multipart tensor is sharded
+        #
+        # split_dim 0 regex:
+        #   - output.*
+        #   - layers.*.attention.wq.weight
+        #   - layers.*.attention.wk.weight
+        #   - layers.*.attention.wv.weight
+        #   - layers.*.feed_forward.w1.weight
+        #   - layers.*.feed_forward.w3.weight
+        #
+        # split_dim 1 regex:
+        #   - tok_embeddings.*
+        #   - layers.*.attention.wo.weight
+        #   - layers.*.feed_forward.w2.weight
+        #
+        if n_dims > 1:
+            split_dim = 1
+            if "tok_embeddings" in name:
+                split_dim = 1
+            elif "layers" in name:
+                if "attention.wo.weight" in name:
+                    split_dim = 1
+                elif "feed_forward.w2.weight" in name:
+                    split_dim = 1
+                else:
+                    split_dim = 0
+            elif "output" in name:
+                split_dim = 0
+
+        # output tensor header
+        fullshape = list(partshape)
+        if n_dims > 1:
+            fullshape[split_dim] *= n_parts
+        sname = name.encode()
+        fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
+        for dim in reversed(fullshape):
             fout.write(struct.pack("i", dim))
         fout.write(sname)
 
-        # data output to file
-        data.tofile(fout)
+        # ensure tensor data is aligned
+        tensor_data_offset = fout.tell()
+        while tensor_data_offset % QK != 0:
+            fout.write(struct.pack("B", 0))
+            tensor_data_offset += 1
+
+        # output unified mappable tensor data
+        if n_dims == 1 or n_parts == 1:
+            # copy tensor which we thankfully received in one piece
+            if part_id == 0:
+                data.tofile(fout)
+        elif split_dim == 0:
+            # reassemble multifile tensor containing some of the rows
+            rows_per_chunk = partshape[0]
+            current_row = part_id * rows_per_chunk
+            bytes_per_row = fullshape[1] // blck_size * type_size
+            offset = current_row * bytes_per_row
+            fout.seek(tensor_data_offset + offset)
+            data.tofile(fout)
+        elif split_dim == 1:
+            # reassemble multifile tensor containing some of the cols
+            cols_per_chunk = partshape[1]
+            current_col = part_id * cols_per_chunk
+            bytes_per_row = fullshape[1] // blck_size * type_size
+            offset_current_col = current_col // blck_size * type_size
+            for row in range(partshape[0]):
+                offset_row = row * bytes_per_row
+                offset = offset_row + offset_current_col
+                fout.seek(tensor_data_offset + offset)
+                data[row].tofile(fout)
+
+        # advance file position to next tensor
+        fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype_cur))
 
 def main():
-
     args = parse_args()
     dir_model = args.dir_model
     ftype = args.ftype
     ftype_str = ["f32", "f16"]
-
     hparams, tokenizer = load_hparams_and_tokenizer(dir_model)
 
     print(args)
 
     # if only writing vocab to file
     if args.vocab_only:
-
         fname_model = f"{dir_model}/consolidated.00.pth"
         fname_out = f"{dir_model}/ggml-vocab.bin"
-
         print(f"Extracting only the vocab from '{fname_model}'\n")
-
-        model = torch.load(fname_model, map_location="cpu")
-
         with open(fname_out, "wb") as fout:
             write_header(fout, hparams, ftype)
             write_tokens(fout, tokenizer)
-
-        del model
-
         print(f"Done. Output file: {fname_out}\n")
-
         return
 
     n_parts = get_n_parts(hparams["dim"])
+    fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin"
 
-    for p in range(n_parts):
+    # we output a single file for ggml
+    with open(fname_out, "wb") as fout:
+        write_header(fout, hparams, ftype)
+        write_tokens(fout, tokenizer)
+        offset_of_tensors = fout.tell()
+        # the tensors we load could be split across multiple files
+        for part_id in range(n_parts):
+            fout.seek(offset_of_tensors)
+            print(f"Processing part {part_id+1} of {n_parts}\n")
+            fname_model = f"{dir_model}/consolidated.0{part_id}.pth"
+            model = torch.load(fname_model, map_location="cpu")
+            process_and_write_variables(fout, model, ftype, part_id, n_parts)
+            del model
 
-        print(f"Processing part {p}\n")
-
-        fname_model = f"{dir_model}/consolidated.0{p}.pth"
-        fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin{'' if p == 0 else '.' + str(p)}"
-
-        model = torch.load(fname_model, map_location="cpu")
-
-        with open(fname_out, "wb") as fout:
-            write_header(fout, hparams, ftype)
-            write_tokens(fout, tokenizer)
-            process_and_write_variables(fout, model, ftype)
-
-        del model
-
-        print(f"Done. Output file: {fname_out}, (part {p})\n")
+    print(f"Done. Output file: {fname_out}\n")
 
 if __name__ == "__main__":
     main()
diff --git a/convert-unversioned-ggml-to-ggml.py b/convert-unversioned-ggml-to-ggml.py
new file mode 100644
index 000000000..5151d9081
--- /dev/null
+++ b/convert-unversioned-ggml-to-ggml.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+# Original by https://github.com/eiz
+# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818
+import argparse
+import glob
+import os
+import struct
+import sys
+from sentencepiece import SentencePieceProcessor
+
+HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Upgrade old ggml model files to the current format')
+    parser.add_argument('dir_model', help='directory containing ggml .bin files')
+    parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file')
+    return parser.parse_args()
+
+def read_header(f_in):
+    struct_fmt = "i" * (3 + len(HPARAMS))
+    struct_size = struct.calcsize(struct_fmt)
+    buf = f_in.read(struct_size)
+    return struct.unpack(struct_fmt, buf)
+
+def write_header(f_out, header):
+    (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header
+
+    if magic != 0x67676d6c:
+        raise Exception('Invalid file magic. Must be an old style ggml file.')
+
+    values = [
+        0x67676d66,  # magic: ggml in hex
+        1, # file version
+        vocab_size,
+        dim,
+        multiple_of,
+        n_heads,
+        n_layers,
+        rot,
+        ftype
+    ]
+    f_out.write(struct.pack("i" * len(values), *values))
+
+def write_tokens(fout, tokenizer):
+    for i in range(tokenizer.vocab_size()):
+        if tokenizer.is_unknown(i):
+            text = " \u2047 ".encode()
+        elif tokenizer.is_control(i):
+            text = b""
+        elif tokenizer.is_byte(i):
+            piece = tokenizer.id_to_piece(i)
+            if len(piece) != 6:
+                print(f"Invalid token: {piece}")
+                sys.exit(1)
+            byte_value = int(piece[3:-1], 16)
+            text = struct.pack("B", byte_value)
+        else:
+            text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode()
+        fout.write(struct.pack("i", len(text)))
+        fout.write(text)
+        fout.write(struct.pack("f", tokenizer.get_score(i)))
+
+def read_tokens(f_in, tokenizer):
+    for i in range(tokenizer.vocab_size()):
+        len_b = f_in.read(4)
+        (length,) = struct.unpack("i", len_b)
+        f_in.read(length)
+
+def copy_all_data(f_out, f_in):
+    while True:
+        buf = f_in.read(1024 * 1024)
+        if not buf:
+            break
+        f_out.write(buf)
+
+def convert_one_file(path_in, tokenizer):
+    path_tmp = f"{path_in}.tmp"
+    path_orig= f"{path_in}.orig"
+    print(f"converting {path_in}")
+    with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out:
+        write_header(f_out, read_header(f_in))
+        read_tokens(f_in, tokenizer)
+        write_tokens(f_out, tokenizer)
+        copy_all_data(f_out, f_in)
+    os.rename(path_in, path_orig)
+    os.rename(path_tmp, path_in)
+
+def main():
+    args = parse_args()
+    files = []
+    files.extend(glob.glob(f"{args.dir_model}/*.bin"))
+    files.extend(glob.glob(f"{args.dir_model}/*.bin.*"))
+
+    tokenizer = SentencePieceProcessor(args.tokenizer_model)
+
+    for file in files:
+        convert_one_file(file, tokenizer)
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/chat-13B.bat b/examples/chat-13B.bat
new file mode 100644
index 000000000..c5c8ac6ef
--- /dev/null
+++ b/examples/chat-13B.bat
@@ -0,0 +1,57 @@
+@setlocal disabledelayedexpansion enableextensions
+@echo off
+
+cd /d "%~dp0.."
+if not "%errorlevel%"=="0" (
+    echo Unable to change directory.
+    pause
+    exit /b 1
+)
+
+if not defined MODEL set "MODEL=models\13B\ggml-model-q4_0.bin"
+if not defined USER_NAME set "USER_NAME=User"
+if not defined AI_NAME set "AI_NAME=ChatLLaMa"
+rem Adjust to the number of CPU cores you want to use.
+rem if not defined N_THREAD set "N_THREAD=8"
+rem Number of tokens to predict (made it larger than default because we want a long interaction)
+if not defined N_PREDICTS set "N_PREDICTS=2048"
+if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
+
+rem Default main script paths
+set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"
+
+rem Get main script path from command line arguments
+set "MAIN_SCRIPT_PATH=%~1"
+
+rem If the main script path was not specified, try the default paths
+if not defined MAIN_SCRIPT_PATH (
+    for %%i in (%DEFAULT_MAIN_SCRIPT_PATHS%) do (
+        if exist "%%i" set "MAIN_SCRIPT_PATH=%%i"
+    )
+)
+
+rem If the main script path was not found, tell the user how to specify it
+if not defined MAIN_SCRIPT_PATH (
+    echo The main script could not be found. Please provide the path to the main script as 1st argument to this script, or place the main script in one of the default locations:
+    echo %DEFAULT_MAIN_SCRIPT_PATHS%
+    pause
+    exit /b 1
+)
+
+rem Default context, feel free to edit it
+set "PROMPT_TEXT=Text transcript of a never ending dialog, where %USER_NAME% interacts with an AI assistant named %AI_NAME%. %AI_NAME% is helpful, kind, honest, friendly, good at writing and never fails to answer %USER_NAME%'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what %USER_NAME% and %AI_NAME% say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown."
+
+rem Set a temporary variable if N_THREAD is set
+if defined N_THREAD (
+    set "_N_THREAD=--threads %N_THREAD%"
+) else (
+    set "_N_THREAD="
+)
+
+rem Run the script
+echo "%MAIN_SCRIPT_PATH%" %GEN_OPTIONS% %_N_THREAD% ^
+  --model "%MODEL%" ^
+  --n_predict %N_PREDICTS% ^
+  --color --interactive ^
+  --reverse-prompt "%USER_NAME%:" ^
+  --prompt "%PROMPT_TEXT%"
diff --git a/examples/common.cpp b/examples/common.cpp
index 2ab000f4f..5400f6b01 100644
--- a/examples/common.cpp
+++ b/examples/common.cpp
@@ -9,11 +9,20 @@
 #include <iterator>
 #include <algorithm>
 
- #if defined(_MSC_VER) || defined(__MINGW32__)
- #include <malloc.h> // using malloc.h with MSC/MINGW
- #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
- #include <alloca.h>
- #endif
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <malloc.h> // using malloc.h with MSC/MINGW
+#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
+#include <alloca.h>
+#endif
+
+#if defined (_WIN32)
+#pragma comment(lib,"kernel32.lib")
+extern "C" __declspec(dllimport) void* __stdcall GetStdHandle(unsigned long nStdHandle);
+extern "C" __declspec(dllimport) int __stdcall GetConsoleMode(void* hConsoleHandle, unsigned long* lpMode);
+extern "C" __declspec(dllimport) int __stdcall SetConsoleMode(void* hConsoleHandle, unsigned long dwMode);
+extern "C" __declspec(dllimport) int __stdcall SetConsoleCP(unsigned int wCodePageID);
+extern "C" __declspec(dllimport) int __stdcall SetConsoleOutputCP(unsigned int wCodePageID);
+#endif
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     // determine sensible default number of threads.
@@ -30,6 +39,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 
     bool invalid_param = false;
     std::string arg;
+    gpt_params default_params;
+
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
 
@@ -57,6 +68,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (params.prompt.back() == '\n') {
                 params.prompt.pop_back();
@@ -159,7 +175,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
             params.n_parts = std::stoi(argv[i]);
         } else if (arg == "-h" || arg == "--help") {
-            gpt_print_usage(argc, argv, params);
+            gpt_print_usage(argc, argv, default_params);
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -171,13 +187,13 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.input_prefix = argv[i];
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            gpt_print_usage(argc, argv, params);
+            gpt_print_usage(argc, argv, default_params);
             exit(1);
         }
     }
     if (invalid_param) {
         fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        gpt_print_usage(argc, argv, params);
+        gpt_print_usage(argc, argv, default_params);
         exit(1);
     }
 
@@ -204,19 +220,19 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
     fprintf(stderr, "                        prompt file to start generation.\n");
-    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d, -1 - infinity)\n", params.n_predict);
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
     fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
-    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", (double)params.top_p);
     fprintf(stderr, "  --repeat_last_n N     last n tokens to consider for penalize (default: %d)\n", params.repeat_last_n);
-    fprintf(stderr, "  --repeat_penalty N    penalize repeat sequence of tokens (default: %.1f)\n", params.repeat_penalty);
+    fprintf(stderr, "  --repeat_penalty N    penalize repeat sequence of tokens (default: %.1f)\n", (double)params.repeat_penalty);
     fprintf(stderr, "  -c N, --ctx_size N    size of the prompt context (default: %d)\n", params.n_ctx);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating\n");
     fprintf(stderr, "  --memory_f32          use f32 instead of f16 for memory key+value\n");
-    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
     fprintf(stderr, "  --n_parts N           number of model parts (default: -1 = determine from dimensions)\n");
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
-    fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt\n");
+    fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     if (ggml_mlock_supported()) {
         fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
@@ -256,3 +272,47 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 
     return res;
 }
+
+/* Keep track of current color of output, and emit ANSI code if it changes. */
+void set_console_color(console_state & con_st, console_color_t color) {
+    if (con_st.use_color && con_st.color != color) {
+        switch(color) {
+            case CONSOLE_COLOR_DEFAULT:
+                printf(ANSI_COLOR_RESET);
+                break;
+            case CONSOLE_COLOR_PROMPT:
+                printf(ANSI_COLOR_YELLOW);
+                break;
+            case CONSOLE_COLOR_USER_INPUT:
+                printf(ANSI_BOLD ANSI_COLOR_GREEN);
+                break;
+        }
+        con_st.color = color;
+    }
+}
+
+#if defined (_WIN32)
+void win32_console_init(bool enable_color) {
+    unsigned long dwMode = 0;
+    void* hConOut = GetStdHandle((unsigned long)-11); // STD_OUTPUT_HANDLE (-11)
+    if (!hConOut || hConOut == (void*)-1 || !GetConsoleMode(hConOut, &dwMode)) {
+        hConOut = GetStdHandle((unsigned long)-12); // STD_ERROR_HANDLE (-12)
+        if (hConOut && (hConOut == (void*)-1 || !GetConsoleMode(hConOut, &dwMode))) {
+            hConOut = 0;
+        }
+    }
+    if (hConOut) {
+        // Enable ANSI colors on Windows 10+
+        if (enable_color && !(dwMode & 0x4)) {
+            SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4)
+        }
+        // Set console output codepage to UTF8
+        SetConsoleOutputCP(65001); // CP_UTF8
+    }
+    void* hConIn = GetStdHandle((unsigned long)-10); // STD_INPUT_HANDLE (-10)
+    if (hConIn && hConIn != (void*)-1 && GetConsoleMode(hConIn, &dwMode)) {
+        // Set console input codepage to UTF8
+        SetConsoleCP(65001); // CP_UTF8
+    }
+}
+#endif
diff --git a/examples/common.h b/examples/common.h
index 8caefd859..1505aa927 100644
--- a/examples/common.h
+++ b/examples/common.h
@@ -63,3 +63,33 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 //
 
 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos);
+
+//
+// Console utils
+//
+
+#define ANSI_COLOR_RED     "\x1b[31m"
+#define ANSI_COLOR_GREEN   "\x1b[32m"
+#define ANSI_COLOR_YELLOW  "\x1b[33m"
+#define ANSI_COLOR_BLUE    "\x1b[34m"
+#define ANSI_COLOR_MAGENTA "\x1b[35m"
+#define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_RESET   "\x1b[0m"
+#define ANSI_BOLD          "\x1b[1m"
+
+enum console_color_t {
+    CONSOLE_COLOR_DEFAULT=0,
+    CONSOLE_COLOR_PROMPT,
+    CONSOLE_COLOR_USER_INPUT
+};
+
+struct console_state {
+    bool use_color = false;
+    console_color_t color = CONSOLE_COLOR_DEFAULT;
+};
+
+void set_console_color(console_state & con_st, console_color_t color);
+
+#if defined (_WIN32)
+void win32_console_init(bool enable_color);
+#endif
diff --git a/examples/gpt4all.sh b/examples/gpt4all.sh
new file mode 100755
index 000000000..d974f95a9
--- /dev/null
+++ b/examples/gpt4all.sh
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main --color --instruct --threads 4 \
+       --model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
+       --file ./prompts/alpaca.txt \
+       --batch_size 8 --ctx_size 2048 \
+       --repeat_last_n 64 --repeat_penalty 1.3 \
+       --n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 7bb2b6bc4..3130aef0c 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -18,67 +18,13 @@
 #include <signal.h>
 #endif
 
-#if defined (_WIN32)
-#pragma comment(lib,"kernel32.lib")
-extern "C" __declspec(dllimport) void* __stdcall GetStdHandle(unsigned long nStdHandle);
-extern "C" __declspec(dllimport) int __stdcall GetConsoleMode(void* hConsoleHandle, unsigned long* lpMode);
-extern "C" __declspec(dllimport) int __stdcall SetConsoleMode(void* hConsoleHandle, unsigned long dwMode);
-#endif
-
-#define ANSI_COLOR_RED     "\x1b[31m"
-#define ANSI_COLOR_GREEN   "\x1b[32m"
-#define ANSI_COLOR_YELLOW  "\x1b[33m"
-#define ANSI_COLOR_BLUE    "\x1b[34m"
-#define ANSI_COLOR_MAGENTA "\x1b[35m"
-#define ANSI_COLOR_CYAN    "\x1b[36m"
-#define ANSI_COLOR_RESET   "\x1b[0m"
-#define ANSI_BOLD          "\x1b[1m"
-
-/* Keep track of current color of output, and emit ANSI code if it changes. */
-enum console_state {
-    CONSOLE_STATE_DEFAULT=0,
-    CONSOLE_STATE_PROMPT,
-    CONSOLE_STATE_USER_INPUT
-};
-
-static console_state con_st = CONSOLE_STATE_DEFAULT;
-static bool con_use_color = false;
-
-void enable_console_colors() {
-#if defined (_WIN32)
-    // Enable ANSI colors on Windows 10+
-    unsigned long dwMode = 0;
-    void* hConOut = GetStdHandle((unsigned long)-11); // STD_OUTPUT_HANDLE (-11)
-    if (hConOut && hConOut != (void*)-1 && GetConsoleMode(hConOut, &dwMode) && !(dwMode & 0x4)) {
-        SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4)
-    }
-#endif
-}
-
-void set_console_state(console_state new_st) {
-    if (!con_use_color) return;
-    // only emit color code if state changed
-    if (new_st != con_st) {
-        con_st = new_st;
-        switch(con_st) {
-        case CONSOLE_STATE_DEFAULT:
-            printf(ANSI_COLOR_RESET);
-            return;
-        case CONSOLE_STATE_PROMPT:
-            printf(ANSI_COLOR_YELLOW);
-            return;
-        case CONSOLE_STATE_USER_INPUT:
-            printf(ANSI_BOLD ANSI_COLOR_GREEN);
-            return;
-        }
-    }
-}
+static console_state con_st;
 
 static bool is_interacting = false;
 
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 void sigint_handler(int signo) {
-    set_console_state(CONSOLE_STATE_DEFAULT);
+    set_console_color(con_st, CONSOLE_COLOR_DEFAULT);
     printf("\n"); // this also force flush stdout.
     if (signo == SIGINT) {
         if (!is_interacting) {
@@ -98,6 +44,14 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    // save choice to use color for later
+    // (note for later: this is a slightly awkward choice)
+    con_st.use_color = params.use_color;
+
+#if defined (_WIN32)
+    win32_console_init(params.use_color);
+#endif
+
     if (params.perplexity) {
         printf("\n************\n");
         printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
@@ -130,10 +84,6 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    // save choice to use color for later
-    // (note for later: this is a slightly awkward choice)
-    con_use_color = params.use_color;
-
 //    params.prompt = R"(// this function checks if the number n is prime
 //bool is_prime(int n) {)";
 
@@ -196,7 +146,10 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    params.n_keep    = std::min(params.n_keep,    (int) embd_inp.size());
+    // number of tokens to keep when resetting context
+    if (params.n_keep < 0 || params.n_keep > (int)embd_inp.size() || params.instruct) {
+        params.n_keep = (int)embd_inp.size();
+    }
 
     // prefix & suffix for instruct mode
     const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true);
@@ -204,16 +157,12 @@ int main(int argc, char ** argv) {
 
     // in instruct mode, we inject a prefix and a suffix to each input by the user
     if (params.instruct) {
-        params.interactive = true;
+        params.interactive_start = true;
         params.antiprompt.push_back("### Instruction:\n\n");
     }
 
-    // enable interactive mode if reverse prompt is specified
-    if (params.antiprompt.size() != 0) {
-        params.interactive = true;
-    }
-
-    if (params.interactive_start) {
+    // enable interactive mode if reverse prompt or interactive start is specified
+    if (params.antiprompt.size() != 0 || params.interactive_start) { 
         params.interactive = true;
     }
 
@@ -260,7 +209,8 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str());
         }
     }
-    fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n", params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
+    fprintf(stderr, "sampling: temp = %f, top_k = %d, top_p = %f, repeat_last_n = %i, repeat_penalty = %f\n",
+        params.temp, params.top_k, params.top_p, params.repeat_last_n, params.repeat_penalty);
     fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     fprintf(stderr, "\n\n");
 
@@ -275,20 +225,18 @@ int main(int argc, char ** argv) {
 #endif
                " - Press Return to return control to LLaMa.\n"
                " - If you want to submit another line, end your input in '\\'.\n\n");
-        is_interacting = params.interactive_start || params.instruct;
+        is_interacting = params.interactive_start;
     }
 
-    bool input_noecho = false;
+    bool is_antiprompt = false;
+    bool input_noecho  = false;
 
     int n_past     = 0;
     int n_remain   = params.n_predict;
     int n_consumed = 0;
 
     // the first thing we will do is to output the prompt, so set color accordingly
-    if (params.use_color) {
-        enable_console_colors();
-    }
-    set_console_state(CONSOLE_STATE_PROMPT);
+    set_console_color(con_st, CONSOLE_COLOR_PROMPT);
 
     std::vector<llama_token> embd;
 
@@ -327,10 +275,10 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // out of user input, sample next token
-            const float top_k          = params.top_k;
-            const float top_p          = params.top_p;
-            const float temp           = params.temp;
-            const float repeat_penalty = params.repeat_penalty;
+            const int32_t top_k          = params.top_k;
+            const float   top_p          = params.top_p;
+            const float   temp           = params.temp;
+            const float   repeat_penalty = params.repeat_penalty;
 
             llama_token id = 0;
 
@@ -389,36 +337,38 @@ int main(int argc, char ** argv) {
         }
         // reset color to default if we there is no pending user input
         if (!input_noecho && (int)embd_inp.size() == n_consumed) {
-            set_console_state(CONSOLE_STATE_DEFAULT);
+            set_console_color(con_st, CONSOLE_COLOR_DEFAULT);
         }
 
         // in interactive mode, and not currently processing queued inputs;
         // check if we should prompt the user for more
         if (params.interactive && (int) embd_inp.size() <= n_consumed) {
-            // check for reverse prompt
-            std::string last_output;
-            for (auto id : last_n_tokens) {
-                last_output += llama_token_to_str(ctx, id);
-            }
 
-            // Check if each of the reverse prompts appears at the end of the output.
-            for (std::string & antiprompt : params.antiprompt) {
-                if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
-                    is_interacting = true;
-                    set_console_state(CONSOLE_STATE_USER_INPUT);
-                    fflush(stdout);
-                    break;
+            // check for reverse prompt
+            if (params.antiprompt.size()) {
+                std::string last_output;
+                for (auto id : last_n_tokens) {
+                    last_output += llama_token_to_str(ctx, id);
+                }
+
+                is_antiprompt = false;
+                // Check if each of the reverse prompts appears at the end of the output.
+                for (std::string & antiprompt : params.antiprompt) {
+                    if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                        is_interacting = true;
+                        is_antiprompt = true;
+                        set_console_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                        fflush(stdout);
+                        break;
+                    }
                 }
             }
 
             if (n_past > 0 && is_interacting) {
                 // potentially set color to indicate we are taking user input
-                set_console_state(CONSOLE_STATE_USER_INPUT);
+                set_console_color(con_st, CONSOLE_COLOR_USER_INPUT);
 
                 if (params.instruct) {
-                    n_consumed = embd_inp.size();
-                    embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
-
                     printf("\n> ");
                 }
 
@@ -431,7 +381,10 @@ int main(int argc, char ** argv) {
                 std::string line;
                 bool another_line = true;
                 do {
-                    std::getline(std::cin, line);
+                    if (!std::getline(std::cin, line)) {
+                        // input stream is bad or EOF received
+                        return 0;
+                    }
                     if (line.empty() || line.back() != '\\') {
                         another_line = false;
                     } else {
@@ -441,17 +394,29 @@ int main(int argc, char ** argv) {
                 } while (another_line);
 
                 // done taking input, reset color
-                set_console_state(CONSOLE_STATE_DEFAULT);
+                set_console_color(con_st, CONSOLE_COLOR_DEFAULT);
 
-                auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+                // Add tokens to embd only if the input buffer is non-empty
+                // Entering a empty line lets the user pass control back
+                if (buffer.length() > 1) {
 
-                if (params.instruct) {
-                    embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                    // instruct mode: insert instruction prefix
+                    if (params.instruct && !is_antiprompt) {
+                        n_consumed = embd_inp.size();
+                        embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
+                    }
+
+                    auto line_inp = ::llama_tokenize(ctx, buffer, false);
+                    embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
+
+                    // instruct mode: insert response suffix
+                    if (params.instruct) {
+                        embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
+                    }
+
+                    n_remain -= line_inp.size();
                 }
 
-                n_remain -= line_inp.size();
-
                 input_noecho = true; // do not echo this again
             }
 
@@ -471,7 +436,7 @@ int main(int argc, char ** argv) {
         }
 
         // In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
-        if (params.interactive && n_remain <= 0) {
+        if (params.interactive && n_remain <= 0 && params.n_predict != -1) {
             n_remain = params.n_predict;
             is_interacting = true;
         }
@@ -484,7 +449,7 @@ int main(int argc, char ** argv) {
     llama_print_timings(ctx);
     llama_free(ctx);
 
-    set_console_state(CONSOLE_STATE_DEFAULT);
+    set_console_color(con_st, CONSOLE_COLOR_DEFAULT);
 
     return 0;
 }
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index 91f0bf6b9..d8195a1f8 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -1,15 +1,17 @@
 #include "common.h"
 #include "llama.h"
 
-std::vector<double> softmax(const std::vector<float>& logits) {
-    std::vector<double> probs(logits.size());
+#include <cmath>
+
+std::vector<float> softmax(const std::vector<float>& logits) {
+    std::vector<float> probs(logits.size());
     float max_logit = logits[0];
     for (float v : logits) max_logit = std::max(max_logit, v);
     double sum_exp = 0.0;
     for (size_t i = 0; i < logits.size(); i++) {
         // Subtract the maximum logit value from the current logit value for numerical stability
-        float logit = logits[i] - max_logit;
-        double exp_logit = std::exp(logit);
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
         sum_exp += exp_logit;
         probs[i] = exp_logit;
     }
@@ -19,15 +21,15 @@ std::vector<double> softmax(const std::vector<float>& logits) {
 
 void perplexity(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 `./main --perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
+    // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     auto tokens = ::llama_tokenize(ctx, params.prompt, true);
 
     int count = 0;
-    double nll = 0.0;
     int seq_count = tokens.size() / params.n_ctx;
     int n_vocab = llama_n_vocab(ctx);
 
+    double nll = 0.0;
     fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
 
     for (int i = 0; i < seq_count; ++i) {
@@ -49,7 +51,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         }
         auto end_t = std::chrono::high_resolution_clock::now();
         if (i == 0) {
-            double seconds = std::chrono::duration<double>(end_t - start_t).count();
+            const float seconds = std::chrono::duration<float>(end_t - start_t).count();
             printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
         }
         // We get the logits for all the tokens in the context window (params.n_ctx)
@@ -70,7 +72,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             std::vector<float> tok_logits(
                 logits.begin() + j * n_vocab,
                 logits.begin() + (j + 1) * n_vocab);
-            double prob = softmax(tok_logits)[tokens[start + j + 1]];
+            float prob = softmax(tok_logits)[tokens[start + j + 1]];
             nll += -std::log(prob);
             ++count;
         }
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp
index f0230f5dc..680757c6b 100644
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -4,8 +4,6 @@
 #include <cstdio>
 #include <string>
 
-const int QK = 32;
-
 // usage:
 //  ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
 //
@@ -21,7 +19,7 @@ int main(int argc, char ** argv) {
 
     // needed to initialize f16 tables
     {
-        struct ggml_init_params params = { 0, NULL };
+        struct ggml_init_params params = { 0, NULL, false };
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
@@ -39,7 +37,7 @@ int main(int argc, char ** argv) {
     {
         const int64_t t_start_us = ggml_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype, QK)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }
@@ -52,8 +50,8 @@ int main(int argc, char ** argv) {
         const int64_t t_main_end_us = ggml_time_us();
 
         printf("\n");
-        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
-        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0);
     }
 
     return 0;
diff --git a/examples/reason-act.sh b/examples/reason-act.sh
new file mode 100755
index 000000000..e7fe655db
--- /dev/null
+++ b/examples/reason-act.sh
@@ -0,0 +1,17 @@
+
+#!/bin/bash
+
+cd `dirname $0`
+cd ..
+
+# get -m model parameter otherwise defer to default
+if [ "$1" == "-m" ]; then
+  MODEL="-m $2 "
+fi
+
+./main $MODEL --color \
+    -f ./prompts/reason-act.txt \
+    -i --interactive-first \
+    --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7 -c 2048 \
+    -r "Question:" -r "Observation:" --in-prefix " " \
+    -n -1
diff --git a/ggml.c b/ggml.c
index c9a4e8675..63aa5eb6e 100644
--- a/ggml.c
+++ b/ggml.c
@@ -16,6 +16,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <stdint.h>
+#include <inttypes.h>
 #include <stdio.h>
 #include <float.h>
 
@@ -79,6 +80,19 @@ static int sched_yield (void) {
 typedef void* thread_ret_t;
 #endif
 
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
 #ifdef __HAIKU__
 #define static_assert(cond, msg) _Static_assert(cond, msg)
 #endif
@@ -150,10 +164,10 @@ typedef double ggml_float;
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) (x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
 #define GGML_COMPUTE_FP32_TO_FP16(x) (x)
 
-#define GGML_FP16_TO_FP32(x) (x)
+#define GGML_FP16_TO_FP32(x) ((float) (x))
 #define GGML_FP32_TO_FP16(x) (x)
 
 #else
@@ -172,8 +186,13 @@ typedef double ggml_float;
 
 #ifdef __F16C__
 
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
 #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
 
 #elif defined(__POWER9_VECTOR__)
 
@@ -322,7 +341,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 // note: do not use these inside ggml.c
 // these are meant to be used via the ggml.h API
 float ggml_fp16_to_fp32(ggml_fp16_t x) {
-    return GGML_FP16_TO_FP32(x);
+    return (float) GGML_FP16_TO_FP32(x);
 }
 
 ggml_fp16_t ggml_fp32_to_fp16(float x) {
@@ -443,22 +462,65 @@ static inline __m128i packNibbles( __m256i bytes )
     __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
     return _mm_packus_epi16( r0, r1 );
 }
+#elif __AVX__
+static inline __m128i bytesFromNibbles( const uint8_t* rsi )
+{
+    // Load 8 bytes from memory
+    __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi );
+
+    // Expand bytes into uint16_t values
+    __m128i bytes = _mm_cvtepu8_epi16( tmp );
+
+    // Unpack values into individual bytes
+    const __m128i lowMask = _mm_set1_epi8( 0xF );
+    __m128i high = _mm_andnot_si128( lowMask, bytes );
+    __m128i low = _mm_and_si128( lowMask, bytes );
+    high = _mm_slli_epi16( high, 4 );
+    bytes = _mm_or_si128( low, high );
+    return bytes;
+}
+
+static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m128i lowByte = _mm_set1_epi16( 0xFF );
+    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
+    __m128i low = _mm_and_si128( lowByte, bytes1 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes1 = _mm_or_si128( low, high );
+    high = _mm_andnot_si128( lowByte, bytes2 );
+    low = _mm_and_si128( lowByte, bytes2 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes2 = _mm_or_si128( low, high );
+
+    return _mm_packus_epi16( bytes1, bytes2);
+}
 #endif
 
 // method 5
 // blocks of QK elements
 // represented with a single float (delta) and QK/2 8-bit ints (i.e QK 4-bit signed integer factors)
+typedef struct {
+    float   d; // delta
+    uint8_t qs[QK / 2]; // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK / 2, "wrong q4_0 block size/padding");
+
+// method 4
+// blocks of QK elements
+// represented with 2 floats (delta + min) and QK/2 8-bit ints (i.e QK 4-bit unsigned integer factors)
+typedef struct {
+    float   d;
+    float   m;
+    uint8_t qs[QK / 2]; // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK / 2, "wrong q4_1 block size/padding");
 
 // reference implementation for deterministic creation of model files
-static void quantize_row_q4_0_reference(const float * restrict x, void * restrict y, int k) {
+static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) {
     assert(k % QK == 0);
     const int nb = k / QK;
 
-    const size_t bs = sizeof(float) + QK/2;
-
-    uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
-    uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
-
     uint8_t pp[QK/2];
 
     for (int i = 0; i < nb; i++) {
@@ -472,39 +534,30 @@ static void quantize_row_q4_0_reference(const float * restrict x, void * restric
         const float d = amax / ((1 << 3) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        *(float *)pd = d;
-        pd += bs;
+        y[i].d = d;
 
         for (int l = 0; l < QK; l += 2) {
             const float v0 = x[i*QK + l + 0]*id;
             const float v1 = x[i*QK + l + 1]*id;
 
-            const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
-            const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
+            const uint8_t vi0 = (int8_t)roundf(v0) + 8;
+            const uint8_t vi1 = (int8_t)roundf(v1) + 8;
 
-            assert(vi0 >= 0 && vi0 < 16);
-            assert(vi1 >= 0 && vi1 < 16);
+            assert(vi0 < 16);
+            assert(vi1 < 16);
 
             pp[l/2] = vi0 | (vi1 << 4);
         }
 
-        memcpy(pb, pp, sizeof(pp));
-        pb += bs;
+        memcpy(y[i].qs, pp, sizeof(pp));
     }
 }
 
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
+static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK == 0);
-
-#if defined(__ARM_NEON) || defined(__AVX2__) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__)
     const int nb = k / QK;
-    const size_t bs = sizeof(float) + QK/2;
 
-    uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
-    uint8_t * restrict pb = ((uint8_t *)y + 0*bs + sizeof(float));
-
-    uint8_t pp[QK/2];
-#endif
+    block_q4_0 * restrict y = vy;
 
 #if defined(__POWER9_VECTOR__)
     const vector float v85 = vec_splats(8.5f);
@@ -532,10 +585,10 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
         const float d = amax / ((1 << 3) - 1);
         const float id = d ? 1.0/d : 0.0;
 
-        *(float *)pd = d;
-        pd += bs;
+        y[i].d = d;
 
         const vector float vid = vec_splats(id);
+        uint8_t * restrict pb = y[i].qs;
         for (int l = 0; l < 8; l++) {
             const vector float vf  = vec_madd(srcv[l], vid, v85);
             const vector signed int vi = vec_signed(vf);
@@ -543,14 +596,9 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
             pb[2*l + 0] = vec_extract(vi, 0) | (vec_extract(vi, 1) << 4);
             pb[2*l + 1] = vec_extract(vi, 2) | (vec_extract(vi, 3) << 4);
         }
-
-        //memcpy(pb, pp, sizeof(pp));
-        pb += bs;
     }
 #elif __ARM_NEON
     for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-
         float32x4_t srcv [8];
         float32x4_t asrcv[8];
         float32x4_t amaxv[8];
@@ -562,27 +610,24 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
         for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]);
         for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]);
 
-        amax = MAX(
+        // absolute max
+        const float amax = MAX(
                 MAX(vgetq_lane_f32(amaxv[0], 0), vgetq_lane_f32(amaxv[0], 1)),
                 MAX(vgetq_lane_f32(amaxv[0], 2), vgetq_lane_f32(amaxv[0], 3)));
 
         const float d = amax / ((1 << 3) - 1);
-        const float id = d ? 1.0/d : 0.0;
+        const float id = d ? 1.0f/d : 0.0f;
 
-        *(float *)pd = d;
-        pd += bs;
+        y[i].d = d;
 
         for (int l = 0; l < 8; l++) {
             const float32x4_t v  = vmulq_n_f32(srcv[l], id);
             const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f));
             const int32x4_t   vi = vcvtq_s32_f32(vf);
 
-            pp[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
-            pp[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
+            y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
+            y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
         }
-
-        memcpy(pb, pp, sizeof(pp));
-        pb += bs;
     }
 #elif defined(__AVX2__)
     for (int i = 0; i < nb; i++) {
@@ -607,8 +652,7 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
 
         // Quantize these floats
         const float d = maxScalar / 7.0f;
-        *(float *)pd = d;
-        pd += bs;
+        y[i].d = d;
         const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -648,8 +692,81 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
 
         // Compress the vector into 4 bit/value, and store
         __m128i res = packNibbles( i0 );
-        _mm_storeu_si128( ( __m128i* )pb, res );
-        pb += bs;
+        _mm_storeu_si128( ( __m128i* )y[i].qs, res );
+    }
+#elif defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float maxScalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = maxScalar / 7.0f;
+        y[i].d = d;
+        const float id = ( maxScalar != 0.0f ) ? 7.0f / maxScalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        // Apply offset to translate the range from [ -7 .. +7 ] into [ +1 .. +15 ]
+        const __m128i off = _mm_set1_epi8( 8);
+        ni0 = _mm_add_epi8( ni0, off );
+        ni4 = _mm_add_epi8( ni4, off );
+
+        // Compress the vector into 4 bit/value, and store
+        __m128i res = packNibbles( ni0, ni4 );
+        _mm_storeu_si128( ( __m128i* )y[i].qs, res );
     }
 #elif defined(__wasm_simd128__)
     for (int i = 0; i < nb; i++) {
@@ -673,20 +790,16 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
         const float d = amax / ((1 << 3) - 1);
         const float id = d ? 1.0/d : 0.0;
 
-        *(float *)pd = d;
-        pd += bs;
+        y[i].d = d;
 
         for (int l = 0; l < 8; l++) {
             const v128_t v  = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id));
             const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f));
             const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf);
 
-            pp[2*l + 0] = wasm_i32x4_extract_lane(vi, 0) | (wasm_i32x4_extract_lane(vi, 1) << 4);
-            pp[2*l + 1] = wasm_i32x4_extract_lane(vi, 2) | (wasm_i32x4_extract_lane(vi, 3) << 4);
+            y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vi, 0) | (wasm_i32x4_extract_lane(vi, 1) << 4);
+            y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vi, 2) | (wasm_i32x4_extract_lane(vi, 3) << 4);
         }
-
-        memcpy(pb, pp, sizeof(pp));
-        pb += bs;
     }
 #else
     // scalar
@@ -694,18 +807,11 @@ void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
 #endif
 }
 
-// method 4
-// blocks of QK elements
-// represented with 2 floats (min + delta) and QK/2 8-bit ints (i.e QK 4-bit unsigned integer factors)
-void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) {
+static void quantize_row_q4_1_reference(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK == 0);
-
     const int nb = k / QK;
-    const size_t bs = 2*sizeof(float) + QK/2;
 
-    uint8_t * restrict pd = ((uint8_t *)y + 0*bs);
-    uint8_t * restrict pm = ((uint8_t *)y + 0*bs +   sizeof(float));
-    uint8_t * restrict pb = ((uint8_t *)y + 0*bs + 2*sizeof(float));
+    block_q4_1 * restrict y = vy;
 
     uint8_t pp[QK/2];
 
@@ -722,45 +828,160 @@ void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) {
         const float d = (max - min) / ((1 << 4) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        *(float *)pm = min;
-        *(float *)pd = d;
-        pm += bs;
-        pd += bs;
+        y[i].d = d;
+        y[i].m = min;
 
         for (int l = 0; l < QK; l += 2) {
             const float v0 = (x[i*QK + l + 0] - min)*id;
             const float v1 = (x[i*QK + l + 1] - min)*id;
 
-            const uint8_t vi0 = round(v0);
-            const uint8_t vi1 = round(v1);
+            const uint8_t vi0 = roundf(v0);
+            const uint8_t vi1 = roundf(v1);
 
-            assert(vi0 >= 0 && vi0 < 16);
-            assert(vi1 >= 0 && vi1 < 16);
+            assert(vi0 < 16);
+            assert(vi1 < 16);
 
             pp[l/2] = vi0 | (vi1 << 4);
         }
 
-        memcpy(pb, pp, sizeof(pp));
-        pb += bs;
+        memcpy(y[i].qs, pp, sizeof(pp));
     }
 }
 
-// TODO: vectorize
-void dequantize_row_q4_0(const void * restrict x, float * restrict y, int k) {
+static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK == 0);
 
     const int nb = k / QK;
-    const size_t bs = sizeof(float) + QK/2;
 
-    const uint8_t * restrict pd = ((const uint8_t *)x + 0*bs);
-    const uint8_t * restrict pb = ((const uint8_t *)x + 0*bs + sizeof(float));
+    block_q4_1 * restrict y = vy;
+
+#if defined(__AVX2__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max for the block
+        __m256 vmax;
+        vmax = _mm256_max_ps( v0, v1 );
+        vmax = _mm256_max_ps( vmax, v2 );
+        vmax = _mm256_max_ps( vmax, v3 );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( vmax, 1 ), _mm256_castps256_ps128( vmax ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float maxScalar = _mm_cvtss_f32( max4 );
+
+        // Compute min for the block
+        __m256 vmin;
+        vmin = _mm256_min_ps( v0, v1 );
+        vmin = _mm256_min_ps( vmin, v2 );
+        vmin = _mm256_min_ps( vmin, v3 );
+
+        __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( vmin, 1 ), _mm256_castps256_ps128( vmin ) );
+        min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) );
+        min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) );
+        const float minScalar = _mm_cvtss_f32( min4 );
+
+        // Quantize these floats
+        const float d = (maxScalar - minScalar) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].m = minScalar;
+        y[i].d = d;
+
+        // x = (x-min)*id
+        const __m256 mul = _mm256_set1_ps( id );
+        const __m256 off = _mm256_set1_ps( minScalar );
+        v0 = _mm256_mul_ps( _mm256_sub_ps( v0, off ), mul );
+        v1 = _mm256_mul_ps( _mm256_sub_ps( v1, off ), mul );
+        v2 = _mm256_mul_ps( _mm256_sub_ps( v2, off ), mul );
+        v3 = _mm256_mul_ps( _mm256_sub_ps( v3, off ), mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        // Compress the vector into 4 bit/value, and store
+        __m128i res = packNibbles( i0 );
+        _mm_storeu_si128( ( __m128i* )y[i].qs, res );
+    }
+#elif __ARM_NEON
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv[8];
+        float32x4_t minv[8];
+        float32x4_t maxv[8];
+
+        for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l);
+
+        for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]);
+        for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]);
+        for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l + 4]);
+
+        for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l + 1]);
+        for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l + 2]);
+        for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l + 4]);
+
+        const float min = vminvq_f32(minv[0]);
+        const float max = vmaxvq_f32(maxv[0]);
+
+        const float d = (max - min) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = d;
+        y[i].m = min;
+
+        const float32x4_t minv0 = vdupq_n_f32(min);
+
+        for (int l = 0; l < 8; l++) {
+            const float32x4_t v  = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id);
+            const int32x4_t   vi = vcvtq_s32_f32(v);
+
+            y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4);
+            y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4);
+        }
+    }
+#else
+    // scalar
+    quantize_row_q4_1_reference(x, vy, k);
+#endif
+}
+
+static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) {
+    assert(k % QK == 0);
+    const int nb = k / QK;
+
+    const block_q4_0 * restrict x = vx;
 
 #if defined(__AVX2__)
     for (int i = 0; i < nb; i++) {
         // scale factor
-        const __m256 d_v = _mm256_broadcast_ss((const float *) (pd + i*bs));
+        const __m256 d_v = _mm256_broadcast_ss(&x[i].d);
 
-        const uint8_t * restrict pp = pb + i*bs;
+        const uint8_t * restrict pp = x[i].qs;
 
         for (int l = 0; l < QK; l += 32) {
             // Load 32x4-bit integers into 32x8-bit integers
@@ -790,17 +1011,15 @@ void dequantize_row_q4_0(const void * restrict x, float * restrict y, int k) {
     }
 #elif defined(__ARM_NEON)
     for (int i = 0; i < nb; i++) {
-        const float d = *(const float *) (pd + i*bs);
+        const float32x4_t vd = vdupq_n_f32(x[i].d);
 
-        const uint8_t * restrict pp = pb + i*bs;
-
-        const float32x4_t vd = vdupq_n_f32(d);
+        const uint8_t * restrict pp = x[i].qs;
 
         for (int l = 0; l < QK; l += 16) {
             // Load 16x4-bit integers into 8x8-bit integers
             const uint8x8_t v8 = vld1_u8(pp + l/2);
 
-            // Expand 4-bit nibbles to 8-bit bytes
+            // Expand 4-bit qs to 8-bit bytes
             const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0f));
             const uint8x8_t v1 = vshr_n_u8(v8, 4);
 
@@ -844,9 +1063,9 @@ void dequantize_row_q4_0(const void * restrict x, float * restrict y, int k) {
 #else
     // scalar
     for (int i = 0; i < nb; i++) {
-        const float d = *(const float *) (pd + i*bs);
+        const float d = x[i].d;
 
-        const uint8_t * restrict pp = pb + i*bs;
+        const uint8_t * restrict pp = x[i].qs;
 
         for (int l = 0; l < QK; l += 2) {
             const uint8_t vi = pp[l/2];
@@ -869,22 +1088,18 @@ void dequantize_row_q4_0(const void * restrict x, float * restrict y, int k) {
 #endif
 }
 
-void dequantize_row_q4_1(const void * restrict x, float * restrict y, int k) {
+static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, int k) {
     assert(k % QK == 0);
-
     const int nb = k / QK;
-    const size_t bs = 2*sizeof(float) + QK/2;
 
-    const uint8_t * restrict pd = ((const uint8_t *)x + 0*bs);
-    const uint8_t * restrict pm = ((const uint8_t *)x + 0*bs + sizeof(float));
-    const uint8_t * restrict pb = ((const uint8_t *)x + 0*bs + 2*sizeof(float));
+    const block_q4_1 * restrict x = vx;
 
 #if defined(__AVX2__)
     for (int i = 0; i < nb; i++) {
-        const __m256 d_v = _mm256_broadcast_ss((const float *) (pd + i*bs));
-        const __m256 d_m = _mm256_broadcast_ss((const float *) (pm + i*bs));
+        const __m256 d_v = _mm256_broadcast_ss(&x[i].d);
+        const __m256 d_m = _mm256_broadcast_ss(&x[i].m);
 
-        const uint8_t * restrict pp = pb + i*bs;
+        const uint8_t * restrict pp = x[i].qs;
 
         for (int l = 0; l < QK; l += 32) {
             // Load 32x4-bit integers into 32x8-bit integers
@@ -909,12 +1124,56 @@ void dequantize_row_q4_1(const void * restrict x, float * restrict y, int k) {
             }
         }
     }
+#elif defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        const float32x4_t vd = vdupq_n_f32(x[i].d);
+        const float32x4_t vm = vdupq_n_f32(x[i].m);
+
+        const uint8_t * restrict pp = x[i].qs;
+
+        for (int l = 0; l < QK; l += 16) {
+            // Load 16x4-bit integers into 8x8-bit integers
+            const uint8x8_t v8 = vld1_u8(pp + l/2);
+
+            // Expand 4-bit qs to 8-bit bytes
+            const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0f));
+            const uint8x8_t v1 = vshr_n_u8(v8, 4);
+
+            // Interleave and combine
+            const uint8x8_t vx_0 = vzip1_u8(v0, v1);
+            const uint8x8_t vx_1 = vzip2_u8(v0, v1);
+
+            const uint8x16_t vq = vcombine_u8(vx_0, vx_1);
+
+            // convert to 2x uint16x8_t
+            const uint16x8_t vi_0 = vmovl_u8(vget_low_u8 (vq));
+            const uint16x8_t vi_1 = vmovl_u8(vget_high_u8(vq));
+
+            // convert to 4x float32x4_t
+            const float32x4_t vf_0 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_0)));
+            const float32x4_t vf_1 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_0)));
+            const float32x4_t vf_2 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_1)));
+            const float32x4_t vf_3 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_1)));
+
+            // multiply by d and add m
+            const float32x4_t r0 = vmlaq_f32(vm, vf_0, vd);
+            const float32x4_t r1 = vmlaq_f32(vm, vf_1, vd);
+            const float32x4_t r2 = vmlaq_f32(vm, vf_2, vd);
+            const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd);
+
+            // Store
+            vst1q_f32(y + i*QK + l +  0, r0);
+            vst1q_f32(y + i*QK + l +  4, r1);
+            vst1q_f32(y + i*QK + l +  8, r2);
+            vst1q_f32(y + i*QK + l + 12, r3);
+        }
+    }
 #else
     for (int i = 0; i < nb; i++) {
-        const float d = *(const float *) (pd + i*bs);
-        const float m = *(const float *) (pm + i*bs);
+        const float d = x[i].d;
+        const float m = x[i].m;
 
-        const uint8_t * restrict pp = pb + i*bs;
+        const uint8_t * restrict pp = x[i].qs;
 
         for (int l = 0; l < QK; l += 2) {
             const uint8_t vi = pp[l/2];
@@ -1027,7 +1286,7 @@ void dequantize_row_q4_1(const void * restrict x, float * restrict y, int k) {
         }                                                         \
         const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
         const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
-        res = vaddvq_f32(vaddq_f32(t0, t1));                      \
+        res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
     }
 
     #define GGML_F16_VEC                GGML_F16x8
@@ -1122,13 +1381,36 @@ void dequantize_row_q4_1(const void * restrict x, float * restrict y, int k) {
 #define GGML_F16_EPR  8
 
 // F16 arithmetic is not supported by AVX, so we use F32 instead
-// we take advantage of the _mm256_cvt intrinsics to convert F16 <-> F32
 
 #define GGML_F32Cx8             __m256
 #define GGML_F32Cx8_ZERO        _mm256_setzero_ps()
 #define GGML_F32Cx8_SET1(x)     _mm256_set1_ps(x)
+
+#if defined(__F16C__)
+// the  _mm256_cvt intrinsics require F16C
 #define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x)))
 #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
+#else
+static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+    float tmp[8];
+
+    for (int i = 0; i < 8; i++)
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+
+    return _mm256_loadu_ps(tmp);
+}
+static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
+    float arr[8];
+
+    _mm256_storeu_ps(arr, y);
+
+    for (int i = 0; i < 8; i++)
+        x[i] = GGML_FP32_TO_FP16(arr[i]);
+}
+#define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
+#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
+#endif
+
 #define GGML_F32Cx8_FMA         GGML_F32x8_FMA
 #define GGML_F32Cx8_ADD         _mm256_add_ps
 #define GGML_F32Cx8_MUL         _mm256_mul_ps
@@ -1440,9 +1722,8 @@ inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, co
 inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
 
 inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y) {
-    ggml_float sumf = 0.0;
-
 #ifdef GGML_SIMD
+    float sumf = 0.0f;
     const int np = (n & ~(GGML_F32_STEP - 1));
 
     GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
@@ -1468,8 +1749,9 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float
     }
 #else
     // scalar
+    ggml_float sumf = 0.0;
     for (int i = 0; i < n; ++i) {
-        sumf += x[i]*y[i];
+        sumf += (ggml_float)(x[i]*y[i]);
     }
 #endif
 
@@ -1479,25 +1761,15 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float
 #if __AVX512F__ && QK == 32
 static inline __m512 dot_q4_0_oneblock_avx512(
     __m512 acc,
-    const uint8_t * pd0,
-    const uint8_t * pd1,
-    const uint8_t * pb0,
-    const uint8_t * pb1,
-    size_t bs,
+    const block_q4_0 * restrict x,
+    const block_q4_0 * restrict y,
     int i
 ) {
-    const float * d0_0 = (const float *) (pd0 + i*bs);
-    const float * d1_0 = (const float *) (pd1 + i*bs);
-
-    const uint8_t * restrict p0 = pb0 + (i+0)*bs;
-    const uint8_t * restrict p1 = pb1 + (i+0)*bs;
-
     // Compute combined scale for the block
-    float scaleScalar = d0_0[0] * d1_0[0];
-    __m512 scale = _mm512_set1_ps( scaleScalar );
+    __m512 d = _mm512_set1_ps( x[i].d * y[i].d );
 
-    __m256i bx = bytesFromNibbles( p0 );
-    __m256i by = bytesFromNibbles( p1 );
+    __m256i bx = bytesFromNibbles( x[i].qs );
+    __m256i by = bytesFromNibbles( y[i].qs );
 
     // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
     const __m256i off = _mm256_set1_epi8( 8 );
@@ -1513,7 +1785,7 @@ static inline __m512 dot_q4_0_oneblock_avx512(
     // Convert int32_t to float
     __m512 p = _mm512_cvtepi32_ps( i64 );
     // Apply the scale, and accumulate
-    return _mm512_fmadd_ps( scale, p, acc );
+    return _mm512_fmadd_ps( d, p, acc );
 }
 #endif
 
@@ -1542,30 +1814,25 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        sumf += GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]);
+        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
     }
 #else
     for (int i = 0; i < n; ++i) {
-        sumf += GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]);
+        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
     }
 #endif
 
     *s = sumf;
 }
 
-inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * restrict x, const void * restrict y) {
+static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     const int nb = n / QK;
 
     assert(n % QK == 0);
     assert(nb % 2 == 0);
 
-    const size_t bs = sizeof(float) + QK/2;
-
-    const uint8_t * restrict pd0 = ((const uint8_t *)x + 0*bs);
-    const uint8_t * restrict pd1 = ((const uint8_t *)y + 0*bs);
-
-    const uint8_t * restrict pb0 = ((const uint8_t *)x + 0*bs + sizeof(float));
-    const uint8_t * restrict pb1 = ((const uint8_t *)y + 0*bs + sizeof(float));
+    const block_q4_0 * restrict x = vx;
+    const block_q4_0 * restrict y = vy;
 
     float sumf = 0.0;
 
@@ -1574,23 +1841,18 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
     float sum1 = 0.0f;
 
     for (int i = 0; i < nb; i += 2) {
-        const float d0_0 = *(const float *) (pd0 + i*bs);
-        const float d1_0 = *(const float *) (pd1 + i*bs);
-        const float d0_1 = *(const float *) (pd0 + (i + 1)*bs);
-        const float d1_1 = *(const float *) (pd1 + (i + 1)*bs);
-
-        //printf("d0_0: %f, d1_0: %f, d0_1: %f, d1_1: %f\n", d0_0, d1_0, d0_1, d1_1);
-
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
+        const block_q4_0 * restrict x0 = &x[i + 0];
+        const block_q4_0 * restrict y0 = &y[i + 0];
+        const block_q4_0 * restrict x1 = &x[i + 1];
+        const block_q4_0 * restrict y1 = &y[i + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0xf);
         const int8x16_t  s8b = vdupq_n_s8(0x8);
 
-        const uint8x16_t v0_0 = vld1q_u8(p0);
-        const uint8x16_t v1_0 = vld1q_u8(p1);
-        const uint8x16_t v0_1 = vld1q_u8(p0 + bs);
-        const uint8x16_t v1_1 = vld1q_u8(p1 + bs);
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v1_0 = vld1q_u8(y0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+        const uint8x16_t v1_1 = vld1q_u8(y1->qs);
 
         // 4-bit -> 8-bit
         const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8(v0_0, m4b));
@@ -1628,11 +1890,11 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
 
         // scalar
 #if defined(__ARM_FEATURE_QRDMX)
-        sum0 += d0_0*d1_0*vaddvq_s32(p_0);
-        sum1 += d0_1*d1_1*vaddvq_s32(p_1);
+        sum0 += x0->d * y0->d * vaddvq_s32(p_0);
+        sum1 += x1->d * y1->d * vaddvq_s32(p_1);
 #else
-        sum0 += d0_0*d1_0*(vgetq_lane_s32(p_0, 0) + vgetq_lane_s32(p_0, 1) + vgetq_lane_s32(p_0, 2) + vgetq_lane_s32(p_0, 3));
-        sum1 += d0_1*d1_1*(vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3));
+        sum0 += x0->d * y0->d * (vgetq_lane_s32(p_0, 0) + vgetq_lane_s32(p_0, 1) + vgetq_lane_s32(p_0, 2) + vgetq_lane_s32(p_0, 3));
+        sum1 += x1->d * y1->d * (vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3));
 #endif
 #else
 	    const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
@@ -1658,11 +1920,11 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
 
         // scalar
 #if defined(__ARM_FEATURE_QRDMX)
-        sum0 += d0_0*d1_0*vaddvq_s16(p_0);
-        sum1 += d0_1*d1_1*vaddvq_s16(p_1);
+        sum0 += x0->d * y0->d * vaddvq_s16(p_0);
+        sum1 += x1->d * y1->d * vaddvq_s16(p_1);
 #else
-        sum0 += d0_0*d1_0*(vgetq_lane_s16(p_0, 0) + vgetq_lane_s16(p_0, 1) + vgetq_lane_s16(p_0, 2) + vgetq_lane_s16(p_0, 3) + vgetq_lane_s16(p_0, 4) + vgetq_lane_s16(p_0, 5) + vgetq_lane_s16(p_0, 6) + vgetq_lane_s16(p_0, 7));
-        sum1 += d0_1*d1_1*(vgetq_lane_s16(p_1, 0) + vgetq_lane_s16(p_1, 1) + vgetq_lane_s16(p_1, 2) + vgetq_lane_s16(p_1, 3) + vgetq_lane_s16(p_1, 4) + vgetq_lane_s16(p_1, 5) + vgetq_lane_s16(p_1, 6) + vgetq_lane_s16(p_1, 7));
+        sum0 += x0->d * y0->d * (vgetq_lane_s16(p_0, 0) + vgetq_lane_s16(p_0, 1) + vgetq_lane_s16(p_0, 2) + vgetq_lane_s16(p_0, 3) + vgetq_lane_s16(p_0, 4) + vgetq_lane_s16(p_0, 5) + vgetq_lane_s16(p_0, 6) + vgetq_lane_s16(p_0, 7));
+        sum1 += x1->d * y1->d * (vgetq_lane_s16(p_1, 0) + vgetq_lane_s16(p_1, 1) + vgetq_lane_s16(p_1, 2) + vgetq_lane_s16(p_1, 3) + vgetq_lane_s16(p_1, 4) + vgetq_lane_s16(p_1, 5) + vgetq_lane_s16(p_1, 6) + vgetq_lane_s16(p_1, 7));
 #endif
 #endif
     }
@@ -1675,70 +1937,112 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
 
     const int superblock_size = 8;
     const int superblock_count = nb / superblock_size;
-    const int remainder = nb % superblock_size;
 
     for (int superblock_ix = 0; superblock_ix < superblock_count; superblock_ix += 1) {
         int i = superblock_ix * superblock_size;
 
-        acc0 = dot_q4_0_oneblock_avx512( acc0, pd0, pd1, pb0, pb1, bs, i+0 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, pd0, pd1, pb0, pb1, bs, i+1 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, pd0, pd1, pb0, pb1, bs, i+2 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, pd0, pd1, pb0, pb1, bs, i+3 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, pd0, pd1, pb0, pb1, bs, i+4 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, pd0, pd1, pb0, pb1, bs, i+5 );
-        acc0 = dot_q4_0_oneblock_avx512( acc0, pd0, pd1, pb0, pb1, bs, i+6 );
-        acc1 = dot_q4_0_oneblock_avx512( acc1, pd0, pd1, pb0, pb1, bs, i+7 );
+        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+0 );
+        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+1 );
+        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+2 );
+        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+3 );
+        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+4 );
+        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+5 );
+        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+6 );
+        acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+7 );
     }
 
     // Remainders
     for (int i = superblock_count * superblock_size; i < nb; ++i) {
-        acc0 = dot_q4_0_oneblock_avx512( acc0, pd0, pd1, pb0, pb1, bs, i );
+        acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i );
     }
 
     // Horizontal sum of all lanes of the accumulator
     sumf = _mm512_reduce_add_ps( acc0 ) + _mm512_reduce_add_ps( acc1 );
 #elif defined(__AVX2__)
-    const size_t countBlocks = nb;
-
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
     // Main loop
+    // TODO: figure a way to do this in a portable way
+    #ifdef __GNUC__
+    #pragma GCC unroll 16
+    #endif
     for (int i = 0; i < nb; ++i) {
-        const float * d0_0 = (const float *) (pd0 + i*bs);
-        const float * d1_0 = (const float *) (pd1 + i*bs);
-
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
-
         // Compute combined scale for the block
-        const __m256 scale = _mm256_mul_ps( _mm256_broadcast_ss( d0_0 ), _mm256_broadcast_ss( d1_0 ) );
+        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        __m256i bx = bytesFromNibbles( p0 );
-        __m256i by = bytesFromNibbles( p1 );
+        __m256i bx = bytesFromNibbles( x[i].qs );
+        __m256i by = bytesFromNibbles( y[i].qs );
 
         // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
         const __m256i off = _mm256_set1_epi8( 8 );
         bx = _mm256_sub_epi8( bx, off );
         by = _mm256_sub_epi8( by, off );
 
-        // Sign-extend first 16 signed bytes into int16_t
-        __m256i x16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( bx ) );
-        __m256i y16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) );
-        // Compute products of int16_t integers, add pairwise
-        __m256i i32 = _mm256_madd_epi16( x16, y16 );
+        // Get absolute values of x vectors
+        const __m256i ax = _mm256_sign_epi8(bx, bx);
 
-        // Sign-extend last 16 signed bytes into int16_t vectors
-        x16 = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( bx, 1 ) );
-        y16 = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) );
-        // Accumulate products of int16_t integers
-        i32 = _mm256_add_epi32( i32, _mm256_madd_epi16( x16, y16 ) );
+        // Sign the values of the y vectors
+        const __m256i sy = _mm256_sign_epi8(by, bx);
+
+        // Perform multiplication and create 16-bit values
+        const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+
+        const __m256i ones = _mm256_set1_epi16(1);
+        const __m256i i32 = _mm256_madd_epi16(ones, dot);
 
         // Convert int32_t to float
-        __m256 p = _mm256_cvtepi32_ps( i32 );
+        const __m256 p = _mm256_cvtepi32_ps( i32 );
+
         // Apply the scale, and accumulate
-        acc = _mm256_fmadd_ps( scale, p, acc );
+        acc = _mm256_fmadd_ps( d, p, acc );
+    }
+
+    // Return horizontal sum of the acc vector
+    __m128 res = _mm256_extractf128_ps( acc, 1 );
+    res = _mm_add_ps( res, _mm256_castps256_ps128( acc ) );
+    res = _mm_add_ps( res, _mm_movehl_ps( res, res ) );
+    res = _mm_add_ss( res, _mm_movehdup_ps( res ) );
+
+    sumf = _mm_cvtss_f32( res );
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (int i = 0; i < nb; ++i) {
+        // Compute combined scale for the block
+        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
+
+        __m128i i32[2];
+        for (int j = 0; j < 2; ++j) {
+            // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes
+            __m128i bx = bytesFromNibbles( x[i].qs + 8*j );
+            __m128i by = bytesFromNibbles( y[i].qs + 8*j );
+
+            // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+            const __m128i off = _mm_set1_epi8( 8 );
+            bx = _mm_sub_epi8( bx, off );
+            by = _mm_sub_epi8( by, off );
+
+	    // Get absolute values of x vectors
+            const __m128i ax = _mm_sign_epi8(bx, bx);
+
+            // Sign the values of the y vectors
+            const __m128i sy = _mm_sign_epi8(by, bx);
+
+            // Perform multiplication and create 16-bit values
+            const __m128i dot = _mm_maddubs_epi16(ax, sy);
+
+            const __m128i ones = _mm_set1_epi16(1);
+            i32[j] = _mm_madd_epi16(ones, dot);
+        }
+
+        // Convert int32_t to float
+        __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] ));
+        // Apply the scale, and accumulate
+        acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
     }
 
     // Return horizontal sum of the acc vector
@@ -1754,21 +2058,18 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
     float sum1 = 0.0f;
 
     for (int i = 0; i < nb; i += 2) {
-        const float d0_0 = *(const float *) (pd0 + i*bs);
-        const float d1_0 = *(const float *) (pd1 + i*bs);
-        const float d0_1 = *(const float *) (pd0 + (i + 1)*bs);
-        const float d1_1 = *(const float *) (pd1 + (i + 1)*bs);
-
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
+        const block_q4_0 * restrict x0 = &px[i + 0];
+        const block_q4_0 * restrict y0 = &py[i + 0];
+        const block_q4_0 * restrict x1 = &px[i + 1];
+        const block_q4_0 * restrict y1 = &py[i + 1];
 
         const v128_t m4b = wasm_u8x16_splat(0xf);
         const v128_t s8b = wasm_i8x16_splat(0x8);
 
-        const v128_t v0_0 = wasm_v128_load(p0);
-        const v128_t v0_1 = wasm_v128_load(p0 + bs);
-        const v128_t v1_0 = wasm_v128_load(p1);
-        const v128_t v1_1 = wasm_v128_load(p1 + bs);
+        const v128_t v0_0 = wasm_v128_load(x0.qs);
+        const v128_t v0_1 = wasm_v128_load(y0.qs);
+        const v128_t v1_0 = wasm_v128_load(x1.qs);
+        const v128_t v1_1 = wasm_v128_load(y1.qs);
 
         // 4-bit -> 8-bit
         const v128_t v0_0l = wasm_v128_and(v0_0, m4b);
@@ -1818,12 +2119,12 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
         const v128_t p_0 = wasm_i16x8_add(pl_0, ph_0);
         const v128_t p_1 = wasm_i16x8_add(pl_1, ph_1);
 
-        sum0 += d0_0*d1_0*(
+        sum0 += x0->d * y0->d * (
                 wasm_i16x8_extract_lane(p_0, 0) + wasm_i16x8_extract_lane(p_0, 1) +
                 wasm_i16x8_extract_lane(p_0, 2) + wasm_i16x8_extract_lane(p_0, 3) +
                 wasm_i16x8_extract_lane(p_0, 4) + wasm_i16x8_extract_lane(p_0, 5) +
                 wasm_i16x8_extract_lane(p_0, 6) + wasm_i16x8_extract_lane(p_0, 7));
-        sum1 += d0_1*d1_1*(
+        sum1 += x1->d * y1->d * (
                 wasm_i16x8_extract_lane(p_1, 0) + wasm_i16x8_extract_lane(p_1, 1) +
                 wasm_i16x8_extract_lane(p_1, 2) + wasm_i16x8_extract_lane(p_1, 3) +
                 wasm_i16x8_extract_lane(p_1, 4) + wasm_i16x8_extract_lane(p_1, 5) +
@@ -1834,11 +2135,11 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
 #else
     // scalar
     for (int i = 0; i < nb; i++) {
-        const float d0 = *(const float *) (pd0 + i*bs);
-        const float d1 = *(const float *) (pd1 + i*bs);
+        const float d0 = x[i].d;
+        const float d1 = y[i].d;
 
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
+        const uint8_t * restrict p0 = x[i].qs;
+        const uint8_t * restrict p1 = y[i].qs;
 
         for (int j = 0; j < QK/2; j++) {
             const uint8_t v0 = p0[j];
@@ -1858,19 +2159,11 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
     *s = sumf;
 }
 
-inline static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict x, const void * restrict y) {
+static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
     const int nb = n / QK;
 
-    const size_t bs = 2*sizeof(float) + QK/2;
-
-    const uint8_t * restrict pd0 = ((const uint8_t *)x + 0*bs);
-    const uint8_t * restrict pd1 = ((const uint8_t *)y + 0*bs);
-
-    const uint8_t * restrict pm0 = ((const uint8_t *)x + 0*bs + sizeof(float));
-    const uint8_t * restrict pm1 = ((const uint8_t *)y + 0*bs + sizeof(float));
-
-    const uint8_t * restrict pb0 = ((const uint8_t *)x + 0*bs + 2*sizeof(float));
-    const uint8_t * restrict pb1 = ((const uint8_t *)y + 0*bs + 2*sizeof(float));
+    const block_q4_1 * restrict x = vx;
+    const block_q4_1 * restrict y = vy;
 
     float sumf = 0.0;
 
@@ -1882,32 +2175,28 @@ inline static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void
 
     // Main loop
     for (int i = 0; i < nb; ++i) {
-        const float * m0 = (const float *) (pm0 + i*bs);
-        const float * m1 = (const float *) (pm1 + i*bs);
+        const float * d0 = &x[i].d;
+        const float * d1 = &y[i].d;
 
-        const float * d0 = (const float *) (pd0 + i*bs);
-        const float * d1 = (const float *) (pd1 + i*bs);
-
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
+        const float * m0 = &x[i].m;
+        const float * m1 = &y[i].m;
 
         const __m256 d0v = _mm256_broadcast_ss( d0 );
         const __m256 d1v = _mm256_broadcast_ss( d1 );
         const __m256 m0v = _mm256_broadcast_ss( m0 );
         const __m256 m1v = _mm256_broadcast_ss( m1 );
 
-
         // Compute combined scale for the block
         const __m256 scale_01 = _mm256_mul_ps( d0v, d1v );
 
         // Compute cross scales for the block
         const __m256 scale_0 = _mm256_mul_ps( d0v, m1v );
         const __m256 scale_1 = _mm256_mul_ps( m0v, d1v );
-        const __m256 cross_scales = _mm256_blend_ps( scale_0, scale_1, 0b10101010 );
+        const __m256 cross_scales = _mm256_blend_ps( scale_0, scale_1, 0xAA /* 0b10101010 */ );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        __m256i bx = bytesFromNibbles( p0 );
-        __m256i by = bytesFromNibbles( p1 );
+        __m256i bx = bytesFromNibbles( x[i].qs );
+        __m256i by = bytesFromNibbles( y[i].qs );
 
         // Now we have a vector with bytes in [ 0 .. 15 ] interval.
 
@@ -1949,17 +2238,56 @@ inline static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void
     res = _mm_add_ss( res, _mm_movehdup_ps( res ) );
 
     sumf = _mm_cvtss_f32( res ) + acc_offset * QK;
+#elif defined(__ARM_NEON)
+    float sum00 = 0.0f;
+    float sum01 = 0.0f;
+    float sum10 = 0.0f;
+    float sum11 = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const block_q4_1 * restrict x0 = &x[i + 0];
+        const block_q4_1 * restrict y0 = &y[i + 0];
+
+        const uint8x16_t m4b = vdupq_n_u8(0xf);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v1_0 = vld1q_u8(y0->qs);
+
+        // and with 0xf
+        const uint8x16_t v0_0l = vandq_u8(v0_0, m4b);
+        const uint8x16_t v1_0l = vandq_u8(v1_0, m4b);
+
+        const uint8x16_t v0_0h = vshrq_n_u8(v0_0, 4);
+        const uint8x16_t v1_0h = vshrq_n_u8(v1_0, 4);
+
+        // dot product into uint16x8_t
+        const uint16x8_t pl0l = vmull_u8(vget_low_u8 (v0_0l), vget_low_u8 (v1_0l));
+        const uint16x8_t pl0h = vmull_u8(vget_high_u8(v0_0l), vget_high_u8(v1_0l));
+
+        const uint16x8_t ph0l = vmull_u8(vget_low_u8 (v0_0h), vget_low_u8 (v1_0h));
+        const uint16x8_t ph0h = vmull_u8(vget_high_u8(v0_0h), vget_high_u8(v1_0h));
+
+        const uint16x8_t pl0 = vaddq_u16(pl0l, pl0h);
+        const uint16x8_t ph0 = vaddq_u16(ph0l, ph0h);
+
+        sum00 += x0->m*y0->m;
+        sum01 += y0->m*x0->d*(vaddvq_u8(v0_0l) + vaddvq_u8(v0_0h));
+        sum10 += x0->m*y0->d*(vaddvq_u8(v1_0l) + vaddvq_u8(v1_0h));
+        sum11 += x0->d*y0->d*vaddvq_u16(vaddq_u16(pl0, ph0));
+    }
+
+    sumf = QK*sum00 + sum01 + sum10 + sum11;
 #else
     // scalar
     for (int i = 0; i < nb; i++) {
-        const float m0 = *(const float *) (pm0 + i*bs);
-        const float m1 = *(const float *) (pm1 + i*bs);
+        const float d0 = x[i].d;
+        const float d1 = y[i].d;
 
-        const float d0 = *(const float *) (pd0 + i*bs);
-        const float d1 = *(const float *) (pd1 + i*bs);
+        const float m0 = x[i].m;
+        const float m1 = y[i].m;
 
-        const uint8_t * restrict p0 = pb0 + i*bs;
-        const uint8_t * restrict p1 = pb1 + i*bs;
+        const uint8_t * restrict p0 = x[i].qs;
+        const uint8_t * restrict p1 = y[i].qs;
 
         for (int j = 0; j < QK/2; j++) {
             const uint8_t v0 = p0[j];
@@ -2018,13 +2346,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * re
     // leftovers
     for (int i = np; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]);
+            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
         }
     }
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]);
+            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
         }
     }
 #endif
@@ -2095,19 +2423,19 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #endif
 }
 
-inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrt(*s);   }
+inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrtf(*s);   }
 inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
-inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrt(x[i]); }
+inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
 inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 
-static const ggml_float GELU_COEF_A    = 0.044715;
-static const ggml_float SQRT_2_OVER_PI = 0.79788456080286535587989211986876;
+static const float GELU_COEF_A    = 0.044715f;
+static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
 
 inline static float ggml_gelu_f32(float x) {
-    return 0.5*x*(1.0 + tanh(SQRT_2_OVER_PI*x*(1.0 + GELU_COEF_A*x*x)));
+    return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
 inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
@@ -2136,7 +2464,7 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
 
 // Sigmoid Linear Unit (SiLU) function
 inline static float ggml_silu_f32(float x) {
-    return x/(1.0 + exp(-x));
+    return x/(1.0f + expf(-x));
 }
 
 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
@@ -2167,7 +2495,7 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     ggml_float sum = 0.0;
     for (int i = 0; i < n; ++i) {
-        sum += x[i];
+        sum += (ggml_float)x[i];
     }
     *s = sum;
 #else
@@ -2177,7 +2505,7 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 
 inline static void ggml_vec_max_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
-    ggml_float max = -INFINITY;
+    float max = -INFINITY;
     for (int i = 0; i < n; ++i) {
         max = MAX(max, x[i]);
     }
@@ -2187,7 +2515,10 @@ inline static void ggml_vec_max_f32(const int n, float * s, const float * x) {
 #endif
 }
 
-inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) { ggml_vec_norm_f32(n, s, x); *s = 1./(*s); }
+inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) {
+    ggml_vec_norm_f32(n, s, x);
+    *s = 1.f/(*s);
+}
 
 //
 // logging
@@ -2230,8 +2561,8 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = {
 static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_COUNT != 5");
 
 static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
-    sizeof(float  )   + QK/2,
-    sizeof(float  )*2 + QK/2,
+    sizeof(block_q4_0),
+    sizeof(block_q4_1),
     sizeof(int8_t ),
     sizeof(int16_t),
     sizeof(int32_t),
@@ -2357,8 +2688,9 @@ struct ggml_context {
     void * mem_buffer;
     bool   mem_buffer_owned;
     bool   mem_buffer_mlocked;
+    bool   no_alloc;
 
-    int n_objects;
+    int    n_objects;
 
     struct ggml_object * objects_begin;
     struct ggml_object * objects_end;
@@ -2443,7 +2775,7 @@ void ggml_print_objects(const struct ggml_context * ctx) {
     GGML_PRINT("%s: --- end ---\n", __func__);
 }
 
-int ggml_nelements(const struct ggml_tensor * tensor) {
+int64_t ggml_nelements(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return tensor->ne[0]*tensor->ne[1]*tensor->ne[2]*tensor->ne[3];
@@ -2575,6 +2907,9 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
     static bool is_first_call = true;
 
     if (is_first_call) {
+        // initialize time system (required on Windows)
+        ggml_time_init();
+
         // initialize GELU, SILU and EXP F32 tables
         {
             const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
@@ -2586,7 +2921,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
                 const float f = table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii);
                 table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
                 table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f));
-                table_exp_f16[i]  = GGML_FP32_TO_FP16(exp(f));
+                table_exp_f16[i]  = GGML_FP32_TO_FP16(expf(f));
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
@@ -2640,6 +2975,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         /*.mem_buffer         =*/ params.mem_buffer ? params.mem_buffer : malloc(params.mem_size),
         /*.mem_buffer_owned   =*/ params.mem_buffer ? false : true,
         /*.mem_buffer_mlocked =*/ false,
+        /*.no_alloc           =*/ params.no_alloc,
         /*.n_objects          =*/ 0,
         /*.objects_begin      =*/ NULL,
         /*.objects_end        =*/ NULL,
@@ -2707,36 +3043,47 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch)
     return result;
 }
 
+#ifdef __APPLE__
+#define MLOCK_SUGGESTION \
+    "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+    "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+#else
+#define MLOCK_SUGGESTION \
+    "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+#endif
+
 bool ggml_mlock_supported(void) {
     return GGML_MLOCK_SUPPORT;
 }
 
+bool ggml_mlock(
+        struct ggml_context * ctx,
+        const void *opt_extra_addr,
+        size_t opt_extra_len,
+        char **err_p) {
+    // TODO: Use SetProcessWorkingSetSize() + VirtualLock() on WIN32
 #if GGML_MLOCK_SUPPORT
-#ifdef __APPLE__
-    #define MLOCK_SUGGESTION "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or\n" \
-                             "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l)."
-#else
-    #define MLOCK_SUGGESTION "Try increasing RLIMIT_MLOCK (ulimit -l)."
-#endif
-bool ggml_mlock(struct ggml_context * ctx, char ** err_p) {
     if (ctx->mem_buffer_mlocked) {
         return true;
     }
-    if (mlock(ctx->mem_buffer, ctx->mem_size)) {
-        int ret = asprintf(err_p, "failed to mlock %zu-byte buffer: %s\n" MLOCK_SUGGESTION,
-                           ctx->mem_size, strerror(errno));
-        GGML_ASSERT(ret >= 0);
+    if (mlock(ctx->mem_buffer, ctx->mem_size) ||
+        (opt_extra_len &&
+         mlock(opt_extra_addr, opt_extra_len))) {
+        if ((*err_p = malloc(1024))) {
+            snprintf(*err_p, 1024,
+                     "failed to mlock %zu-byte buffer: %s\n" MLOCK_SUGGESTION,
+                     ctx->mem_size + opt_extra_len,
+                     strerror(errno));
+        }
         return false;
     }
     ctx->mem_buffer_mlocked = true;
     return true;
-}
 #else // GGML_MLOCK_SUPPORT
-bool ggml_mlock(struct ggml_context * ctx, char ** err_p) {
     *err_p = strdup("can't mlock because it's not supported on this system");
     return false;
-}
 #endif // GGML_MLOCK_SUPPORT
+}
 
 ////////////////////////////////////////////////////////////////////////////////
 
@@ -2744,7 +3091,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int* ne,
+        const int64_t* ne,
         void*  data) {
     // always insert objects at the end of the context's memory pool
     struct ggml_object * obj_cur = ctx->objects_end;
@@ -2755,7 +3102,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
 
     size_t size_needed = 0;
 
-    if (data == NULL) {
+    if (data == NULL && !ctx->no_alloc) {
         size_needed += GGML_TYPE_SIZE[type]*(ne[0]/GGML_BLCK_SIZE[type]);
         for (int i = 1; i < n_dims; i++) {
             size_needed *= ne[i];
@@ -2839,7 +3186,7 @@ struct ggml_tensor * ggml_new_tensor_impl(
         /*.perf_runs    =*/ 0,
         /*.perf_cycles  =*/ 0,
         /*.perf_time_us =*/ 0,
-        /*.data         =*/ data == NULL ? (void *)(result + 1) : data,
+        /*.data         =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data,
         /*.pad          =*/ { 0 },
     };
 
@@ -2864,44 +3211,44 @@ struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int * ne) {
+        const int64_t * ne) {
     return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
 }
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0) {
+        int64_t ne0) {
     return ggml_new_tensor(ctx, type, 1, &ne0);
 }
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1) {
-    const int ne[2] = { ne0, ne1 };
+        int64_t ne0,
+        int64_t ne1) {
+    const int64_t ne[2] = { ne0, ne1 };
     return ggml_new_tensor(ctx, type, 2, ne);
 }
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2) {
-    const int ne[3] = { ne0, ne1, ne2 };
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2) {
+    const int64_t ne[3] = { ne0, ne1, ne2 };
     return ggml_new_tensor(ctx, type, 3, ne);
 }
 
 struct ggml_tensor * ggml_new_tensor_4d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2,
-        int    ne3) {
-    const int ne[4] = { ne0, ne1, ne2, ne3 };
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2,
+        int64_t ne3) {
+    const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
     return ggml_new_tensor(ctx, type, 4, ne);
 }
 
@@ -3548,7 +3895,7 @@ struct ggml_tensor * ggml_mean(
         is_node = true;
     }
 
-    int ne[GGML_MAX_DIMS] = { 1, a->ne[1], a->ne[2], a->ne[3] };
+    int64_t ne[GGML_MAX_DIMS] = { 1, a->ne[1], a->ne[2], a->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, ne);
 
     result->op   = GGML_OP_MEAN;
@@ -3909,7 +4256,7 @@ struct ggml_tensor * ggml_mul_mat(
         is_node = true;
     }
 
-    const int ne[4] = { a->ne[1], b->ne[1], a->ne[2], b->ne[3] };
+    const int64_t ne[4] = { a->ne[1], b->ne[1], a->ne[2], b->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MIN(a->n_dims, b->n_dims), ne);
 
     result->op   = GGML_OP_MUL_MAT;
@@ -4034,8 +4381,8 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1) {
+        int64_t               ne0,
+        int64_t               ne1) {
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0*ne1);
 
@@ -4046,7 +4393,7 @@ struct ggml_tensor * ggml_reshape_2d(
         is_node = true;
     }
 
-    const int ne[2] = { ne0, ne1 };
+    const int64_t ne[2] = { ne0, ne1 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a->data);
 
     result->op   = GGML_OP_RESHAPE;
@@ -4060,9 +4407,9 @@ struct ggml_tensor * ggml_reshape_2d(
 struct ggml_tensor * ggml_reshape_3d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2) {
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2) {
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2);
 
@@ -4073,7 +4420,7 @@ struct ggml_tensor * ggml_reshape_3d(
         is_node = true;
     }
 
-    const int ne[3] = { ne0, ne1, ne2 };
+    const int64_t ne[3] = { ne0, ne1, ne2 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a->data);
 
     result->op   = GGML_OP_RESHAPE;
@@ -4089,7 +4436,7 @@ struct ggml_tensor * ggml_reshape_3d(
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
+        int64_t               ne0,
         size_t                offset) {
     if (a->grad) {
         GGML_ASSERT(false); // gradient propagation is not supported
@@ -4110,15 +4457,15 @@ struct ggml_tensor * ggml_view_1d(
 struct ggml_tensor * ggml_view_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
+        int64_t               ne0,
+        int64_t               ne1,
         size_t                nb1,
         size_t                offset) {
     if (a->grad) {
         GGML_ASSERT(false); // gradient propagation is not supported
     }
 
-    const int ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
+    const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 };
 
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset);
 
@@ -4349,7 +4696,7 @@ struct ggml_tensor * ggml_conv_1d_1s(
         is_node = true;
     }
 
-    const int ne[4] = { b->ne[0], a->ne[2], 1, 1, };
+    const int64_t ne[4] = { b->ne[0], a->ne[2], 1, 1, };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     result->op   = GGML_OP_CONV_1D_1S;
@@ -4376,7 +4723,7 @@ struct ggml_tensor * ggml_conv_1d_2s(
         is_node = true;
     }
 
-    const int ne[4] = { b->ne[0]/2, a->ne[2], 1, 1, };
+    const int64_t ne[4] = { b->ne[0]/2, a->ne[2], 1, 1, };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
 
     result->op   = GGML_OP_CONV_1D_2S;
@@ -4476,10 +4823,10 @@ static void ggml_compute_forward_dup_f16(
         return;
     }
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb00 = src0->nb[0];
     const size_t nb01 = src0->nb[1];
@@ -4496,9 +4843,9 @@ static void ggml_compute_forward_dup_f16(
             size_t id = 0;
             const size_t rs = ne00*nb00;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
                         const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
                         char * dst_ptr = (char *) dst->data + id*rs;
 
@@ -4512,10 +4859,10 @@ static void ggml_compute_forward_dup_f16(
             size_t id = 0;
             float * dst_ptr = (float *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr);
@@ -4534,10 +4881,10 @@ static void ggml_compute_forward_dup_f16(
             size_t id = 0;
             float * dst_ptr = (float *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr);
@@ -4550,10 +4897,10 @@ static void ggml_compute_forward_dup_f16(
             size_t id = 0;
             ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = *src0_ptr;
@@ -4580,10 +4927,10 @@ static void ggml_compute_forward_dup_f32(
         return;
     }
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb00 = src0->nb[0];
     const size_t nb01 = src0->nb[1];
@@ -4600,9 +4947,9 @@ static void ggml_compute_forward_dup_f32(
             size_t id = 0;
             const size_t rs = ne00*nb00;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
                         const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
                         char * dst_ptr = (char *) dst->data + id*rs;
 
@@ -4616,10 +4963,10 @@ static void ggml_compute_forward_dup_f32(
             size_t id = 0;
             ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
@@ -4638,10 +4985,10 @@ static void ggml_compute_forward_dup_f32(
             size_t id = 0;
             float * dst_ptr = (float *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = *src0_ptr;
@@ -4654,10 +5001,10 @@ static void ggml_compute_forward_dup_f32(
             size_t id = 0;
             ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
 
-            for (int i03 = 0; i03 < ne03; i03++) {
-                for (int i02 = 0; i02 < ne02; i02++) {
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    for (int64_t i01 = 0; i01 < ne01; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
                             const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                             dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
@@ -5043,18 +5390,18 @@ static void ggml_compute_forward_sum_f32(
     assert(ggml_is_scalar(dst));
     assert(src0->nb[0] == sizeof(float));
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb01 = src0->nb[1];
     const size_t nb02 = src0->nb[2];
     const size_t nb03 = src0->nb[3];
 
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = 0; i02 < ne02; i02++) {
-            for (int i01 = 0; i01 < ne01; i01++) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = 0; i01 < ne01; i01++) {
                 ggml_vec_sum_f32(ne00,
                         (float *) (dst->data),
                         (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03));
@@ -5099,19 +5446,19 @@ static void ggml_compute_forward_mean_f32(
 
     assert(src0->nb[0] == sizeof(float));
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb01 = src0->nb[1];
     const size_t nb02 = src0->nb[2];
     const size_t nb03 = src0->nb[3];
 
-    const int ne0 = dst->ne[0];
-    const int ne1 = dst->ne[1];
-    const int ne2 = dst->ne[2];
-    const int ne3 = dst->ne[3];
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const int64_t ne3 = dst->ne[3];
 
     assert(ne0 == 1);
     assert(ne1 == ne01);
@@ -5127,9 +5474,9 @@ static void ggml_compute_forward_mean_f32(
     const size_t nb2 = dst->nb[2];
     const size_t nb3 = dst->nb[3];
 
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = 0; i02 < ne02; i02++) {
-            for (int i01 = 0; i01 < ne01; i01++) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = 0; i01 < ne01; i01++) {
                 ggml_vec_sum_f32(ne00,
                         (float *) ((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3),
                         (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03));
@@ -5616,10 +5963,10 @@ static void ggml_compute_forward_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb01 = src0->nb[1];
     const size_t nb02 = src0->nb[2];
@@ -5629,31 +5976,32 @@ static void ggml_compute_forward_norm_f32(
     const size_t nb2 = dst->nb[2];
     const size_t nb3 = dst->nb[3];
 
-    const ggml_float eps = 1e-5f; // TODO: make this a parameter
+    const float eps = 1e-5f; // TODO: make this a parameter
 
     // TODO: optimize
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = 0; i02 < ne02; i02++) {
-            for (int i01 = ith; i01 < ne01; i01 += nth) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
                 const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
-                ggml_float mean = 0.0;
-                for (int i00 = 0; i00 < ne00; i00++) {
-                    mean += x[i00];
+                ggml_float sum = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    sum += (ggml_float)x[i00];
                 }
 
-                mean /= ne00;
+                float mean = sum/ne00;
 
                 float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
 
                 ggml_float sum2 = 0.0;
-                for (int i00 = 0; i00 < ne00; i00++) {
-                    ggml_float v = x[i00] - mean;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    float v = x[i00] - mean;
                     y[i00] = v;
-                    sum2 += v*v;
+                    sum2 += (ggml_float)(v*v);
                 }
 
-                const float scale = 1.0/sqrt(sum2/ne00 + eps);
+                float variance = sum2/ne00;
+                const float scale = 1.0f/sqrtf(variance + eps);
 
                 ggml_vec_scale_f32(ne00, y, scale);
             }
@@ -5698,10 +6046,10 @@ static void ggml_compute_forward_rms_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 size_t nb01 = src0->nb[1];
     const size_t nb02 = src0->nb[2];
@@ -5711,20 +6059,20 @@ static void ggml_compute_forward_rms_norm_f32(
     const size_t nb2 = dst->nb[2];
     const size_t nb3 = dst->nb[3];
 
-    const ggml_float eps = 1e-6f; // TODO: make this a parameter
+    const float eps = 1e-6f; // TODO: make this a parameter
 
     // TODO: optimize
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = 0; i02 < ne02; i02++) {
-            for (int i01 = ith; i01 < ne01; i01 += nth) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
                 const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
-                ggml_float mean = 0.0;
-                for (int i00 = 0; i00 < ne00; i00++) {
-                    mean += x[i00] * x[i00];
+                ggml_float sum = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    sum += (ggml_float)(x[i00] * x[i00]);
                 }
 
-                mean /= ne00;
+                float mean = sum/ne00;
 
                 float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
 
@@ -5733,7 +6081,7 @@ static void ggml_compute_forward_rms_norm_f32(
                 //     y[i00] = x[i00];
                 // }
 
-                const float scale = 1.0/sqrt(mean + eps);
+                const float scale = 1.0f/sqrtf(mean + eps);
 
                 ggml_vec_scale_f32(ne00, y, scale);
             }
@@ -5773,13 +6121,13 @@ static bool ggml_compute_forward_mul_mat_use_blas(
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    //const int ne00 = src0->ne[0];
-    //const int ne01 = src0->ne[1];
+    //const int64_t ne00 = src0->ne[0];
+    //const int64_t ne01 = src0->ne[1];
 
-    const int ne10 = src1->ne[0];
+    const int64_t ne10 = src1->ne[0];
 
-    const int ne0 = dst->ne[0];
-    const int ne1 = dst->ne[1];
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
 
     // TODO: find the optimal values for these
     if (ggml_is_contiguous(src0) &&
@@ -5801,28 +6149,33 @@ static void ggml_compute_forward_mul_mat_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    //const int ne12 = src1->ne[2];
-    //const int ne13 = src1->ne[3];
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+    const int64_t ne10 = src1->ne[0];
+#endif
+    const int64_t ne11 = src1->ne[1];
+#ifndef NDEBUG
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
 
-    //const int ne0  = dst->ne[0];
-    //const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    const int64_t ne2  = dst->ne[2];
+    const int64_t ne3  = dst->ne[3];
 
-    //const int nb00 = src0->nb[0];
+    const int nb00 = src0->nb[0];
+#endif
     const int nb01 = src0->nb[1];
     const int nb02 = src0->nb[2];
     const int nb03 = src0->nb[3];
 
+#ifndef NDEBUG
     const int nb10 = src1->nb[0];
+#endif
     const int nb11 = src1->nb[1];
     const int nb12 = src1->nb[2];
     const int nb13 = src1->nb[3];
@@ -5840,8 +6193,9 @@ static void ggml_compute_forward_mul_mat_f32(
     assert(ne2  == ne12);
     assert(ne3  == ne13);
 
-    // TODO: we don't support permuted src0
+    // we don't support permuted src0 or src1
     assert(nb00 == sizeof(float));
+    assert(nb10 == sizeof(float));
 
     // dst cannot be transposed or permuted
     assert(nb0 == sizeof(float));
@@ -5859,8 +6213,6 @@ static void ggml_compute_forward_mul_mat_f32(
 
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
-        GGML_ASSERT(nb10 == sizeof(float));
-
         if (params->ith != 0) {
             return;
         }
@@ -5873,8 +6225,8 @@ static void ggml_compute_forward_mul_mat_f32(
             return;
         }
 
-        for (int i03 = 0; i03 < ne03; i03++) {
-            for (int i02 = 0; i02 < ne02; i02++) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
                 const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
                 const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
 
@@ -5903,9 +6255,6 @@ static void ggml_compute_forward_mul_mat_f32(
         return;
     }
 
-    // TODO: do not support transposed src1
-    assert(nb10 == sizeof(float));
-
     // parallelize by src0 rows using ggml_vec_dot_f32
 
     // total rows in src0
@@ -5924,7 +6273,7 @@ static void ggml_compute_forward_mul_mat_f32(
         const int i02 = (ir - i03*ne02*ne01)/ne01;
         const int i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-        for (int ic = 0; ic < ne11; ++ic) {
+        for (int64_t ic = 0; ic < ne11; ++ic) {
             // src1 indices
             const int i13 = i03;
             const int i12 = i02;
@@ -5965,21 +6314,21 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    const int ne12 = src1->ne[2];
-    const int ne13 = src1->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 ne0  = dst->ne[0];
-    const int ne1  = dst->ne[1];
-    const int ne2  = dst->ne[2];
-    const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    const int64_t ne2  = dst->ne[2];
+    const int64_t ne3  = dst->ne[3];
+    //const int64_t ne   = ne0*ne1*ne2*ne3;
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -6039,12 +6388,12 @@ static void ggml_compute_forward_mul_mat_f16_f32(
 
         float * const wdata = params->wdata;
 
-        for (int i03 = 0; i03 < ne03; i03++) {
-            for (int i02 = 0; i02 < ne02; i02++) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
                 {
                     size_t id = 0;
-                    for (int i01 = 0; i01 < ne01; ++i01) {
-                        for (int i00 = 0; i00 < ne00; ++i00) {
+                    for (int64_t i01 = 0; i01 < ne01; ++i01) {
+                        for (int64_t i00 = 0; i00 < ne00; ++i00) {
                             wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00));
                         }
                     }
@@ -6074,10 +6423,10 @@ static void ggml_compute_forward_mul_mat_f16_f32(
         ggml_fp16_t * const wdata = params->wdata;
 
         size_t id = 0;
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                for (int i11 = 0; i11 < ne11; ++i11) {
-                    for (int i10 = 0; i10 < ne10; ++i10) {
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                    for (int64_t i10 = 0; i10 < ne10; ++i10) {
                         wdata[id++] = GGML_FP32_TO_FP16(*(float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10));
                     }
                 }
@@ -6129,7 +6478,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
 
         float * dst_col = (float *) ((char *) dst->data + (i0*nb0 + 0*nb1 + i2*nb2 + i3*nb3));
 
-        for (int ic = 0; ic < ne11; ++ic) {
+        for (int64_t ic = 0; ic < ne11; ++ic) {
             ggml_vec_dot_f16(ne00, &dst_col[ic*ne0], src0_row, src1_col + ic*ne00);
         }
     }
@@ -6147,7 +6496,30 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     //}
 }
 
-static void ggml_compute_forward_mul_mat_q4_0_f32(
+typedef void (*dequantize_row_q_t)(const void * restrict x, float * restrict y, int k);
+typedef void (*quantize_row_q_t)(const float * restrict x, void * restrict y, int k);
+typedef void (*vec_dot_q_t)(const int n, float * restrict s, const void * restrict x, const void * restrict y);
+
+typedef struct {
+    dequantize_row_q_t dequantize_row_q;
+    quantize_row_q_t   quantize_row_q;
+    vec_dot_q_t        vec_dot_q;
+} quantize_fns_t;
+
+static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
+    [GGML_TYPE_Q4_0] = {
+        .dequantize_row_q = dequantize_row_q4_0,
+        .quantize_row_q   = quantize_row_q4_0,
+        .vec_dot_q        = ggml_vec_dot_q4_0,
+    },
+    [GGML_TYPE_Q4_1] = {
+        .dequantize_row_q = dequantize_row_q4_1,
+        .quantize_row_q   = quantize_row_q4_1,
+        .vec_dot_q        = ggml_vec_dot_q4_1,
+    },
+};
+
+static void ggml_compute_forward_mul_mat_q_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
@@ -6155,21 +6527,20 @@ static void ggml_compute_forward_mul_mat_q4_0_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    const int ne12 = src1->ne[2];
-    const int ne13 = src1->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 ne0  = dst->ne[0];
-    const int ne1  = dst->ne[1];
-    const int ne2  = dst->ne[2];
-    const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    const int64_t ne2  = dst->ne[2];
+    const int64_t ne3  = dst->ne[3];
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -6194,8 +6565,13 @@ static void ggml_compute_forward_mul_mat_q4_0_f32(
     GGML_ASSERT(ne2  == ne12);
     GGML_ASSERT(ne3  == ne13);
 
-    // TODO: we don't support permuted src0
-    GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[GGML_TYPE_Q4_0]);
+    const enum ggml_type type = src0->type;
+    quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q;
+    vec_dot_q_t      const vec_dot_q      = quantize_fns[type].vec_dot_q;
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]);
+    GGML_ASSERT(nb10 == sizeof(float));
 
     // dst cannot be transposed or permuted
     GGML_ASSERT(nb0 == sizeof(float));
@@ -6213,8 +6589,6 @@ static void ggml_compute_forward_mul_mat_q4_0_f32(
 
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
-        GGML_ASSERT(nb10 == sizeof(float));
-
         if (params->ith != 0) {
             return;
         }
@@ -6228,198 +6602,14 @@ static void ggml_compute_forward_mul_mat_q4_0_f32(
         }
 
         float * const wdata = params->wdata;
+        dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
 
-        for (int i03 = 0; i03 < ne03; i03++) {
-            for (int i02 = 0; i02 < ne02; i02++) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
                 {
                     size_t id = 0;
-                    for (int i01 = 0; i01 < ne01; ++i01) {
-                        dequantize_row_q4_0((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
-                        id += ne00;
-                    }
-                }
-
-                const float * x = wdata;
-                const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
-
-                float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-
-                // zT = y * xT
-                cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne10,
-                        0.0f,    d, ne01);
-            }
-        }
-
-        /*printf("CBLAS Q4_0 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);*/
-
-        return;
-    }
-#endif
-
-    if (params->type == GGML_TASK_INIT) {
-        char * wdata = params->wdata;
-
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                for (int i11 = 0; i11 < ne11; ++i11) {
-                    quantize_row_q4_0((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
-                    wdata += (ne10*GGML_TYPE_SIZE[GGML_TYPE_Q4_0])/GGML_BLCK_SIZE[GGML_TYPE_Q4_0];
-                }
-            }
-        }
-
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // TODO: do not support transposed src1
-
-    // parallelize by src0 rows using ggml_vec_dot_q4_0
-
-    // total rows in src0
-    const int nr = ne01*ne02*ne03;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    void * wdata = params->wdata;
-
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // src0 indices
-        const int i03 = ir/(ne02*ne01);
-        const int i02 = (ir - i03*ne02*ne01)/ne01;
-        const int i01 = (ir - i03*ne02*ne01 - i02*ne01);
-
-        const int i13 = i03;
-        const int i12 = i02;
-
-        const int i0 = i01;
-        const int i2 = i02;
-        const int i3 = i03;
-
-        void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03));
-        char * src1_col =          ((char *)      wdata + (      (0 + i12*ne11 + i13*ne12*ne11)*ne00*GGML_TYPE_SIZE[GGML_TYPE_Q4_0])/GGML_BLCK_SIZE[GGML_TYPE_Q4_0]);
-
-        float * dst_col = (float *) ((char *) dst->data + (i0*nb0 + 0*nb1 + i2*nb2 + i3*nb3));
-
-        assert(ne00 % 32 == 0);
-
-        for (int ic = 0; ic < ne11; ++ic) {
-            ggml_vec_dot_q4_0(ne00, &dst_col[ic*ne0], src0_row, ((void *) (src1_col + (ic*ne00*GGML_TYPE_SIZE[GGML_TYPE_Q4_0])/GGML_BLCK_SIZE[GGML_TYPE_Q4_0])));
-        }
-    }
-
-    //int64_t t1 = ggml_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    printf("\n");
-    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
-
-    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
-}
-
-static void ggml_compute_forward_mul_mat_q4_1_f32(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    const int ne03 = src0->ne[3];
-
-    const int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    const int ne12 = src1->ne[2];
-    const int ne13 = src1->ne[3];
-
-    const int ne0  = dst->ne[0];
-    const int ne1  = dst->ne[1];
-    const int ne2  = dst->ne[2];
-    const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
-
-    const int nb00 = src0->nb[0];
-    const int nb01 = src0->nb[1];
-    const int nb02 = src0->nb[2];
-    const int nb03 = src0->nb[3];
-
-    const int nb10 = src1->nb[0];
-    const int nb11 = src1->nb[1];
-    const int nb12 = src1->nb[2];
-    const int nb13 = src1->nb[3];
-
-    const int nb0  = dst->nb[0];
-    const int nb1  = dst->nb[1];
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_ASSERT(ne02 == ne12);
-    GGML_ASSERT(ne03 == ne13);
-    GGML_ASSERT(ne2  == ne12);
-    GGML_ASSERT(ne3  == ne13);
-
-    // TODO: we don't support permuted src0
-    GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[GGML_TYPE_Q4_1]);
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    GGML_ASSERT(ne0 == ne01);
-    GGML_ASSERT(ne1 == ne11);
-    GGML_ASSERT(ne2 == ne02);
-    GGML_ASSERT(ne3 == ne03);
-
-    // nb01 >= nb00 - src0 is not transposed
-    //   compute by src0 rows
-
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-    if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
-        GGML_ASSERT(nb10 == sizeof(float));
-
-        if (params->ith != 0) {
-            return;
-        }
-
-        if (params->type == GGML_TASK_INIT) {
-            return;
-        }
-
-        if (params->type == GGML_TASK_FINALIZE) {
-            return;
-        }
-
-        float * const wdata = params->wdata;
-
-        for (int i03 = 0; i03 < ne03; i03++) {
-            for (int i02 = 0; i02 < ne02; i02++) {
-                {
-                    size_t id = 0;
-                    for (int i01 = 0; i01 < ne01; ++i01) {
-                        dequantize_row_q4_1((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
+                    for (int64_t i01 = 0; i01 < ne01; ++i01) {
+                        dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
                         id += ne00;
                     }
                 }
@@ -6446,15 +6636,13 @@ static void ggml_compute_forward_mul_mat_q4_1_f32(
 
     if (params->type == GGML_TASK_INIT) {
         char * wdata = params->wdata;
+        const size_t row_size = ne10*GGML_TYPE_SIZE[type]/GGML_BLCK_SIZE[type];
 
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                for (int i11 = 0; i11 < ne11; ++i11) {
-                    //for (int i10 = 0; i10 < ne10; ++i10) {
-                    //    wdata[id++] = GGML_FP32_TO_FP16(*(float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10));
-                    //}
-                    quantize_row_q4_1((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
-                    wdata += (ne10*GGML_TYPE_SIZE[GGML_TYPE_Q4_1])/GGML_BLCK_SIZE[GGML_TYPE_Q4_1];
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                    quantize_row_q((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
+                    wdata += row_size;
                 }
             }
         }
@@ -6466,9 +6654,7 @@ static void ggml_compute_forward_mul_mat_q4_1_f32(
         return;
     }
 
-    // TODO: do not support transposed src1
-
-    // parallelize by src0 rows using ggml_vec_dot_q4_1
+    // parallelize by src0 rows using ggml_vec_dot_q
 
     // total rows in src0
     const int nr = ne01*ne02*ne03;
@@ -6481,6 +6667,7 @@ static void ggml_compute_forward_mul_mat_q4_1_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     void * wdata = params->wdata;
+    const size_t row_size = ne00*GGML_TYPE_SIZE[type]/GGML_BLCK_SIZE[type];
 
     for (int ir = ir0; ir < ir1; ++ir) {
         // src0 indices
@@ -6496,14 +6683,14 @@ static void ggml_compute_forward_mul_mat_q4_1_f32(
         const int i3 = i03;
 
         void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03));
-        char * src1_col =          ((char *)      wdata + (      (0 + i12*ne11 + i13*ne12*ne11)*ne00*GGML_TYPE_SIZE[GGML_TYPE_Q4_1])/GGML_BLCK_SIZE[GGML_TYPE_Q4_1]);
+        char * src1_col =          ((char *)      wdata + (      (0 + i12*ne11 + i13*ne12*ne11)*row_size));
 
         float * dst_col = (float *) ((char *) dst->data + (i0*nb0 + 0*nb1 + i2*nb2 + i3*nb3));
 
         assert(ne00 % 32 == 0);
 
-        for (int ic = 0; ic < ne11; ++ic) {
-            ggml_vec_dot_q4_1(ne00, &dst_col[ic*ne0], src0_row, ((void *) (src1_col + (ic*ne00*GGML_TYPE_SIZE[GGML_TYPE_Q4_1])/GGML_BLCK_SIZE[GGML_TYPE_Q4_1])));
+        for (int64_t ic = 0; ic < ne11; ++ic) {
+            vec_dot_q(ne00, &dst_col[ic*ne0], src0_row, (void *) (src1_col + ic*row_size));
         }
     }
 
@@ -6527,12 +6714,9 @@ static void ggml_compute_forward_mul_mat(
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-            {
-                ggml_compute_forward_mul_mat_q4_0_f32(params, src0, src1, dst);
-            } break;
         case GGML_TYPE_Q4_1:
             {
-                ggml_compute_forward_mul_mat_q4_1_f32(params, src0, src1, dst);
+                ggml_compute_forward_mul_mat_q_f32(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F16:
             {
@@ -6693,7 +6877,7 @@ static void ggml_compute_forward_transpose(
 
 // ggml_compute_forward_get_rows
 
-static void ggml_compute_forward_get_rows_q4_0(
+static void ggml_compute_forward_get_rows_q(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
@@ -6706,42 +6890,17 @@ static void ggml_compute_forward_get_rows_q4_0(
 
     const int nc = src0->ne[0];
     const int nr = ggml_nelements(src1);
+    const enum ggml_type type = src0->type;
+    dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q;
 
     assert( dst->ne[0] == nc);
     assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == GGML_TYPE_SIZE[GGML_TYPE_Q4_0]);
+    assert(src0->nb[0] == GGML_TYPE_SIZE[type]);
 
     for (int i = 0; i < nr; ++i) {
         const int r = ((int32_t *) src1->data)[i];
 
-        dequantize_row_q4_0(
-                (const void *) ((char *) src0->data + r*src0->nb[1]),
-                     (float *) ((char *)  dst->data + i*dst->nb[1]), nc);
-    }
-}
-
-static void ggml_compute_forward_get_rows_q4_1(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    const int nc = src0->ne[0];
-    const int nr = ggml_nelements(src1);
-
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == GGML_TYPE_SIZE[GGML_TYPE_Q4_1]);
-
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
-
-        dequantize_row_q4_1(
+        dequantize_row_q(
                 (const void *) ((char *) src0->data + r*src0->nb[1]),
                      (float *) ((char *)  dst->data + i*dst->nb[1]), nc);
     }
@@ -6809,12 +6968,9 @@ static void ggml_compute_forward_get_rows(
         struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
-            {
-                ggml_compute_forward_get_rows_q4_0(params, src0, src1, dst);
-            } break;
         case GGML_TYPE_Q4_1:
             {
-                ggml_compute_forward_get_rows_q4_1(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_q(params, src0, src1, dst);
             } break;
         case GGML_TYPE_F16:
             {
@@ -6966,12 +7122,12 @@ static void ggml_compute_forward_soft_max_f32(
                 ggml_fp16_t s = GGML_FP32_TO_FP16(p[i] - max);
                 memcpy(&scvt, &s, sizeof(scvt));
                 const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
-                sum += val;
+                sum += (ggml_float)val;
                 p[i] = val;
             }
         }
 
-        assert(sum > 0.0f);
+        assert(sum > 0.0);
 
         sum = 1.0/sum;
         ggml_vec_scale_f32(nc, p, sum);
@@ -7026,10 +7182,10 @@ static void ggml_compute_forward_rope_f32(
     const int n_dims = ((int32_t *) src1->data)[1];
     const int mode   = ((int32_t *) src1->data)[2];
 
-    //const int ne0 = src0->ne[0];
-    const int ne1 = src0->ne[1];
-    const int ne2 = src0->ne[2];
-    const int ne3 = src0->ne[3];
+    //const int64_t ne0 = src0->ne[0];
+    const int64_t ne1 = src0->ne[1];
+    const int64_t ne2 = src0->ne[2];
+    const int64_t ne3 = src0->ne[3];
 
     const int nb0 = src0->nb[0];
     const int nb1 = src0->nb[1];
@@ -7042,21 +7198,21 @@ static void ggml_compute_forward_rope_f32(
     assert(nb0 == sizeof(float));
 
     // TODO: optimize
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
             const int p = (mode == 0 ? n_past + i2 : i2);
-            for (int i1 = 0; i1 < ne1; i1++) {
+            for (int64_t i1 = 0; i1 < ne1; i1++) {
                 for (int i0 = 0; i0 < n_dims; i0 += 2) {
-                    const double theta = pow(10000.0, ((double)-i0)/n_dims);
+                    const float theta = powf(10000.0, ((float)-i0)/n_dims);
 
-                    const double cos_theta = cos(p*theta);
-                    const double sin_theta = sin(p*theta);
+                    const float cos_theta = cosf(p*theta);
+                    const float sin_theta = sinf(p*theta);
 
                     const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
                           float * dst_data  = (float *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
-                    double x0 = src[0];
-                    double x1 = src[1];
+                    const float x0 = src[0];
+                    const float x1 = src[1];
 
                     dst_data[0] = x0*cos_theta - x1*sin_theta;
                     dst_data[1] = x0*sin_theta + x1*cos_theta;
@@ -7083,10 +7239,10 @@ static void ggml_compute_forward_rope_f16(
     const int n_dims = ((int32_t *) src1->data)[1];
     const int mode   = ((int32_t *) src1->data)[2];
 
-    //const int ne0 = src0->ne[0];
-    const int ne1 = src0->ne[1];
-    const int ne2 = src0->ne[2];
-    const int ne3 = src0->ne[3];
+    //const int64_t ne0 = src0->ne[0];
+    const int64_t ne1 = src0->ne[1];
+    const int64_t ne2 = src0->ne[2];
+    const int64_t ne3 = src0->ne[3];
 
     const int nb0 = src0->nb[0];
     const int nb1 = src0->nb[1];
@@ -7098,21 +7254,21 @@ static void ggml_compute_forward_rope_f16(
 
     assert(nb0 == sizeof(ggml_fp16_t));
 
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
             const int p = (mode == 0 ? n_past + i2 : i2);
-            for (int i1 = 0; i1 < ne1; i1++) {
+            for (int64_t i1 = 0; i1 < ne1; i1++) {
                 for (int i0 = 0; i0 < n_dims; i0 += 2) {
-                    const double theta = pow(10000.0, ((double)-i0)/n_dims);
+                    const float theta = powf(10000.0, ((float)-i0)/n_dims);
 
-                    const double cos_theta = cos(p*theta);
-                    const double sin_theta = sin(p*theta);
+                    const float cos_theta = cosf(p*theta);
+                    const float sin_theta = sinf(p*theta);
 
                     const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
                           ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
-                    double x0 = ggml_fp16_to_fp32(src[0]);
-                    double x1 = ggml_fp16_to_fp32(src[1]);
+                    const float x0 = ggml_fp16_to_fp32(src[0]);
+                    const float x1 = ggml_fp16_to_fp32(src[1]);
 
                     dst_data[0] = ggml_fp32_to_fp16(x0*cos_theta - x1*sin_theta);
                     dst_data[1] = ggml_fp32_to_fp16(x0*sin_theta + x1*cos_theta);
@@ -7162,21 +7318,21 @@ static void ggml_compute_forward_conv_1d_1s_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    //const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    //const int ne12 = src1->ne[2];
-    //const int ne13 = src1->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 ne0  = dst->ne[0];
-    //const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    //const int64_t ne0  = dst->ne[0];
+    //const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
+    //const int64_t ne   = ne0*ne1*ne2*ne3;
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -7213,11 +7369,11 @@ static void ggml_compute_forward_conv_1d_1s_f16_f32(
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int i02 = 0; i02 < ne02; i02++) {
-                for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
                     ggml_fp16_t * dst_data = wdata + i02*ew0*ne00;
-                    for (int i00 = 0; i00 < ne00; i00++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
                         dst_data[i00*ew0 + i01] = src[i00];
                     }
                 }
@@ -7228,10 +7384,10 @@ static void ggml_compute_forward_conv_1d_1s_f16_f32(
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + ne02*ew0*ne00;
 
-            for (int i11 = 0; i11 < ne11; i11++) {
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
                 const float * const src = (float *)((char *) src1->data + i11*nb11);
                 ggml_fp16_t * dst_data = wdata;
-                for (int i10 = 0; i10 < ne10; i10++) {
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
                     dst_data[(i10 + nh)*ew0 + i11] = GGML_FP32_TO_FP16(src[i10]);
                 }
             }
@@ -7256,7 +7412,7 @@ static void ggml_compute_forward_conv_1d_1s_f16_f32(
 
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int i0 = 0; i0 < ne10; ++i0) {
+        for (int64_t i0 = 0; i0 < ne10; ++i0) {
             dst_data[i0] = 0;
             for (int k = -nh; k <= nh; k++) {
                 float v = 0.0f;
@@ -7282,21 +7438,21 @@ static void ggml_compute_forward_conv_1d_1s_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    //const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    //const int ne12 = src1->ne[2];
-    //const int ne13 = src1->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 ne0  = dst->ne[0];
-    //const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    //const int64_t ne0  = dst->ne[0];
+    //const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
+    //const int64_t ne   = ne0*ne1*ne2*ne3;
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -7333,11 +7489,11 @@ static void ggml_compute_forward_conv_1d_1s_f32(
         {
             float * const wdata = (float *) params->wdata + 0;
 
-            for (int i02 = 0; i02 < ne02; i02++) {
-                for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
                     float * dst_data = wdata + i02*ew0*ne00;
-                    for (int i00 = 0; i00 < ne00; i00++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
                         dst_data[i00*ew0 + i01] = src[i00];
                     }
                 }
@@ -7348,10 +7504,10 @@ static void ggml_compute_forward_conv_1d_1s_f32(
         {
             float * const wdata = (float *) params->wdata + ne02*ew0*ne00;
 
-            for (int i11 = 0; i11 < ne11; i11++) {
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
                 const float * const src = (float *)((char *) src1->data + i11*nb11);
                 float * dst_data = wdata;
-                for (int i10 = 0; i10 < ne10; i10++) {
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
                     dst_data[(i10 + nh)*ew0 + i11] = src[i10];
                 }
             }
@@ -7376,7 +7532,7 @@ static void ggml_compute_forward_conv_1d_1s_f32(
 
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int i0 = 0; i0 < ne10; ++i0) {
+        for (int64_t i0 = 0; i0 < ne10; ++i0) {
             dst_data[i0] = 0;
             for (int k = -nh; k <= nh; k++) {
                 float v = 0.0f;
@@ -7430,21 +7586,21 @@ static void ggml_compute_forward_conv_1d_2s_f16_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    //const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    //const int ne12 = src1->ne[2];
-    //const int ne13 = src1->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 ne0  = dst->ne[0];
-    //const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    //const int64_t ne0  = dst->ne[0];
+    //const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
+    //const int64_t ne   = ne0*ne1*ne2*ne3;
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -7481,11 +7637,11 @@ static void ggml_compute_forward_conv_1d_2s_f16_f32(
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int i02 = 0; i02 < ne02; i02++) {
-                for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
                     ggml_fp16_t * dst_data = wdata + i02*ew0*ne00;
-                    for (int i00 = 0; i00 < ne00; i00++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
                         dst_data[i00*ew0 + i01] = src[i00];
                     }
                 }
@@ -7496,10 +7652,10 @@ static void ggml_compute_forward_conv_1d_2s_f16_f32(
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + ne02*ew0*ne00;
 
-            for (int i11 = 0; i11 < ne11; i11++) {
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
                 const float * const src = (float *)((char *) src1->data + i11*nb11);
                 ggml_fp16_t * dst_data = wdata;
-                for (int i10 = 0; i10 < ne10; i10++) {
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
                     dst_data[(i10 + nh)*ew0 + i11] = GGML_FP32_TO_FP16(src[i10]);
                 }
             }
@@ -7524,7 +7680,7 @@ static void ggml_compute_forward_conv_1d_2s_f16_f32(
 
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int i0 = 0; i0 < ne10; i0 += 2) {
+        for (int64_t i0 = 0; i0 < ne10; i0 += 2) {
             dst_data[i0/2] = 0;
             for (int k = -nh; k <= nh; k++) {
                 float v = 0.0f;
@@ -7550,21 +7706,21 @@ static void ggml_compute_forward_conv_1d_2s_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int ne00 = src0->ne[0];
-    const int ne01 = src0->ne[1];
-    const int ne02 = src0->ne[2];
-    //const int ne03 = src0->ne[3];
+    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 int ne10 = src1->ne[0];
-    const int ne11 = src1->ne[1];
-    //const int ne12 = src1->ne[2];
-    //const int ne13 = src1->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 ne0  = dst->ne[0];
-    //const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
-    //const int ne   = ne0*ne1*ne2*ne3;
+    //const int64_t ne0  = dst->ne[0];
+    //const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
+    //const int64_t ne   = ne0*ne1*ne2*ne3;
 
     const int nb00 = src0->nb[0];
     const int nb01 = src0->nb[1];
@@ -7601,11 +7757,11 @@ static void ggml_compute_forward_conv_1d_2s_f32(
         {
             float * const wdata = (float *) params->wdata + 0;
 
-            for (int i02 = 0; i02 < ne02; i02++) {
-                for (int i01 = 0; i01 < ne01; i01++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01);
                     float * dst_data = wdata + i02*ew0*ne00;
-                    for (int i00 = 0; i00 < ne00; i00++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
                         dst_data[i00*ew0 + i01] = src[i00];
                     }
                 }
@@ -7616,10 +7772,10 @@ static void ggml_compute_forward_conv_1d_2s_f32(
         {
             float * const wdata = (float *) params->wdata + ne02*ew0*ne00;
 
-            for (int i11 = 0; i11 < ne11; i11++) {
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
                 const float * const src = (float *)((char *) src1->data + i11*nb11);
                 float * dst_data = wdata;
-                for (int i10 = 0; i10 < ne10; i10++) {
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
                     dst_data[(i10 + nh)*ew0 + i11] = src[i10];
                 }
             }
@@ -7644,7 +7800,7 @@ static void ggml_compute_forward_conv_1d_2s_f32(
 
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * dst_data = (float *)((char *) dst->data + i1*nb1);
-        for (int i0 = 0; i0 < ne10; i0 += 2) {
+        for (int64_t i0 = 0; i0 < ne10; i0 += 2) {
             dst_data[i0/2] = 0;
             for (int k = -nh; k <= nh; k++) {
                 float v = 0.0f;
@@ -7696,25 +7852,25 @@ static void ggml_compute_forward_flash_attn_f32(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int neq0 = q->ne[0];
-    const int neq1 = q->ne[1];
-    const int neq2 = q->ne[2];
-    const int neq3 = q->ne[3];
+    const int64_t neq0 = q->ne[0];
+    const int64_t neq1 = q->ne[1];
+    const int64_t neq2 = q->ne[2];
+    const int64_t neq3 = q->ne[3];
 
-    const int nek0 = k->ne[0];
-    const int nek1 = k->ne[1];
-    //const int nek2 = k->ne[2];
-    //const int nek3 = k->ne[3];
+    const int64_t nek0 = k->ne[0];
+    const int64_t nek1 = k->ne[1];
+    //const int64_t nek2 = k->ne[2];
+    //const int64_t nek3 = k->ne[3];
 
-    //const int nev0 = v->ne[0];
-    const int nev1 = v->ne[1];
-    //const int nev2 = v->ne[2];
-    //const int nev3 = v->ne[3];
+    //const int64_t nev0 = v->ne[0];
+    const int64_t nev1 = v->ne[1];
+    //const int64_t nev2 = v->ne[2];
+    //const int64_t nev3 = v->ne[3];
 
-    const int ne0  = dst->ne[0];
-    const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
 
     const int nbk0 = k->nb[0];
     const int nbk1 = k->nb[1];
@@ -7739,10 +7895,10 @@ static void ggml_compute_forward_flash_attn_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int D = neq0;
-    const int N = neq1;
-    const int P = nek1 - N;
-    const int M = P + N;
+    const int64_t D = neq0;
+    const int64_t N = neq1;
+    const int64_t P = nek1 - N;
+    const int64_t M = P + N;
 
     const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
 
@@ -7788,7 +7944,7 @@ static void ggml_compute_forward_flash_attn_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    const float scale = 1.0/sqrt((double) D);
+    const float scale = 1.0f/sqrtf(D);
 
     //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
 
@@ -7804,7 +7960,7 @@ static void ggml_compute_forward_flash_attn_f32(
             S[i] = -INFINITY;
         }
 
-        for (int ic = 0; ic < nek1; ++ic) {
+        for (int64_t ic = 0; ic < nek1; ++ic) {
             // k indices
             const int ik3 = iq3;
             const int ik2 = iq2;
@@ -7823,7 +7979,7 @@ static void ggml_compute_forward_flash_attn_f32(
         ggml_vec_scale_f32(nek1, S, scale);
 
         if (masked) {
-            for (int i = P; i < M; i++) {
+            for (int64_t i = P; i < M; i++) {
                 if (i > P + iq1) {
                     S[i] = -INFINITY;
                 }
@@ -7835,7 +7991,7 @@ static void ggml_compute_forward_flash_attn_f32(
             float max = -INFINITY;
             ggml_vec_max_f32(M, &max, S);
 
-            float sum = 0.0f;
+            ggml_float sum = 0.0;
             {
 #ifdef GGML_SOFT_MAX_ACCELERATE
                 max = -max;
@@ -7856,7 +8012,7 @@ static void ggml_compute_forward_flash_attn_f32(
                             ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
                             memcpy(&scvt[j], &s, sizeof(uint16_t));
                             const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
-                            sump[j] += val;
+                            sump[j] += (ggml_float)val;
                             SS[j] = val;
                         }
                     }
@@ -7868,7 +8024,7 @@ static void ggml_compute_forward_flash_attn_f32(
 #endif
             }
 
-            assert(sum > 0.0f);
+            assert(sum > 0.0);
 
             sum = 1.0/sum;
             ggml_vec_scale_f32(M, S, sum);
@@ -7881,7 +8037,7 @@ static void ggml_compute_forward_flash_attn_f32(
 #endif
         }
 
-        for (int ic = 0; ic < nev1; ++ic) {
+        for (int64_t ic = 0; ic < nev1; ++ic) {
             // dst indices
             const int i1 = iq1;
             const int i2 = iq2;
@@ -7905,25 +8061,25 @@ static void ggml_compute_forward_flash_attn_f16(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int neq0 = q->ne[0];
-    const int neq1 = q->ne[1];
-    const int neq2 = q->ne[2];
-    const int neq3 = q->ne[3];
+    const int64_t neq0 = q->ne[0];
+    const int64_t neq1 = q->ne[1];
+    const int64_t neq2 = q->ne[2];
+    const int64_t neq3 = q->ne[3];
 
-    const int nek0 = k->ne[0];
-    const int nek1 = k->ne[1];
-    //const int nek2 = k->ne[2];
-    //const int nek3 = k->ne[3];
+    const int64_t nek0 = k->ne[0];
+    const int64_t nek1 = k->ne[1];
+    //const int64_t nek2 = k->ne[2];
+    //const int64_t nek3 = k->ne[3];
 
-    //const int nev0 = v->ne[0];
-    const int nev1 = v->ne[1];
-    //const int nev2 = v->ne[2];
-    //const int nev3 = v->ne[3];
+    //const int64_t nev0 = v->ne[0];
+    const int64_t nev1 = v->ne[1];
+    //const int64_t nev2 = v->ne[2];
+    //const int64_t nev3 = v->ne[3];
 
-    const int ne0  = dst->ne[0];
-    const int ne1  = dst->ne[1];
-    //const int ne2  = dst->ne[2];
-    //const int ne3  = dst->ne[3];
+    const int64_t ne0  = dst->ne[0];
+    const int64_t ne1  = dst->ne[1];
+    //const int64_t ne2  = dst->ne[2];
+    //const int64_t ne3  = dst->ne[3];
 
     const int nbk0 = k->nb[0];
     const int nbk1 = k->nb[1];
@@ -7948,10 +8104,10 @@ static void ggml_compute_forward_flash_attn_f16(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int D = neq0;
-    const int N = neq1;
-    const int P = nek1 - N;
-    const int M = P + N;
+    const int64_t D = neq0;
+    const int64_t N = neq1;
+    const int64_t P = nek1 - N;
+    const int64_t M = P + N;
 
     const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
 
@@ -7997,7 +8153,7 @@ static void ggml_compute_forward_flash_attn_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    const float scale = 1.0/sqrt((double) D);
+    const float scale = 1.0f/sqrtf(D);
 
     //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
 
@@ -8014,7 +8170,7 @@ static void ggml_compute_forward_flash_attn_f16(
         }
 
         if (GGML_VEC_DOT_UNROLL > 2 || nek1 % GGML_VEC_DOT_UNROLL != 0) {
-            for (int ic = 0; ic < nek1; ++ic) {
+            for (int64_t ic = 0; ic < nek1; ++ic) {
                 // k indices
                 const int ik3 = iq3;
                 const int ik2 = iq2;
@@ -8029,7 +8185,7 @@ static void ggml_compute_forward_flash_attn_f16(
                         (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
             }
         } else {
-            for (int ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
+            for (int64_t ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
                 // k indices
                 const int ik3 = iq3;
                 const int ik2 = iq2;
@@ -8049,7 +8205,7 @@ static void ggml_compute_forward_flash_attn_f16(
         ggml_vec_scale_f32(nek1, S, scale);
 
         if (masked) {
-            for (int i = P; i < M; i++) {
+            for (int64_t i = P; i < M; i++) {
                 if (i > P + iq1) {
                     S[i] = -INFINITY;
                 }
@@ -8061,7 +8217,7 @@ static void ggml_compute_forward_flash_attn_f16(
             float max = -INFINITY;
             ggml_vec_max_f32(M, &max, S);
 
-            float sum = 0.0f;
+            ggml_float sum = 0.0;
             {
 #ifdef GGML_SOFT_MAX_ACCELERATE
                 max = -max;
@@ -8082,7 +8238,7 @@ static void ggml_compute_forward_flash_attn_f16(
                             ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
                             memcpy(&scvt[j], &s, sizeof(uint16_t));
                             const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
-                            sump[j] += val;
+                            sump[j] += (ggml_float)val;
                             SS[j] = val;
                         }
                     }
@@ -8094,7 +8250,7 @@ static void ggml_compute_forward_flash_attn_f16(
 #endif
             }
 
-            assert(sum > 0.0f);
+            assert(sum > 0.0);
 
             sum = 1.0/sum;
             ggml_vec_scale_f32(M, S, sum);
@@ -8109,12 +8265,12 @@ static void ggml_compute_forward_flash_attn_f16(
 
         ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*Mup + CACHE_LINE_SIZE_F32) + Mup);
 
-        for (int i = 0; i < M; i++) {
+        for (int64_t i = 0; i < M; i++) {
             S16[i] = GGML_FP32_TO_FP16(S[i]);
         }
 
         if (GGML_VEC_DOT_UNROLL == 1 || (nev1 % GGML_VEC_DOT_UNROLL != 0)) {
-            for (int ic = 0; ic < nev1; ++ic) {
+            for (int64_t ic = 0; ic < nev1; ++ic) {
                 // dst indices
                 const int i1 = iq1;
                 const int i2 = iq2;
@@ -8126,7 +8282,7 @@ static void ggml_compute_forward_flash_attn_f16(
                         S16);
             }
         } else {
-            for (int ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
+            for (int64_t ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
                 // dst indices
                 const int i1 = iq1;
                 const int i2 = iq2;
@@ -8182,35 +8338,35 @@ static void ggml_compute_forward_flash_ff_f16(
     int64_t t0 = ggml_perf_time_us();
     UNUSED(t0);
 
-    const int nea0 = a->ne[0];
-    const int nea1 = a->ne[1];
-    const int nea2 = a->ne[2];
-    const int nea3 = a->ne[3];
+    const int64_t nea0 = a->ne[0];
+    const int64_t nea1 = a->ne[1];
+    const int64_t nea2 = a->ne[2];
+    const int64_t nea3 = a->ne[3];
 
-    const int neb00 = b0->ne[0];
-    const int neb01 = b0->ne[1];
-    //const int neb02 = b0->ne[2];
-    //const int neb03 = b0->ne[3];
+    const int64_t neb00 = b0->ne[0];
+    const int64_t neb01 = b0->ne[1];
+    //const int64_t neb02 = b0->ne[2];
+    //const int64_t neb03 = b0->ne[3];
 
-    const int neb10 = b1->ne[0];
-    const int neb11 = b1->ne[1];
-    //const int neb12 = b1->ne[2];
-    //const int neb13 = b1->ne[3];
+    const int64_t neb10 = b1->ne[0];
+    const int64_t neb11 = b1->ne[1];
+    //const int64_t neb12 = b1->ne[2];
+    //const int64_t neb13 = b1->ne[3];
 
-    const int nec00 = c0->ne[0];
-    const int nec01 = c0->ne[1];
-    //const int nec02 = c0->ne[2];
-    //const int nec03 = c0->ne[3];
+    const int64_t nec00 = c0->ne[0];
+    const int64_t nec01 = c0->ne[1];
+    //const int64_t nec02 = c0->ne[2];
+    //const int64_t nec03 = c0->ne[3];
 
-    const int nec10 = c1->ne[0];
-    const int nec11 = c1->ne[1];
-    //const int nec12 = c1->ne[2];
-    //const int nec13 = c1->ne[3];
+    const int64_t nec10 = c1->ne[0];
+    const int64_t nec11 = c1->ne[1];
+    //const int64_t nec12 = c1->ne[2];
+    //const int64_t nec13 = c1->ne[3];
 
-    const int ne0 = dst->ne[0];
-    const int ne1 = dst->ne[1];
-    const int ne2 = dst->ne[2];
-    //const int ne3 = dst->ne[3];
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    //const int64_t ne3 = dst->ne[3];
 
     const int nba0 = a->nb[0];
     const int nba1 = a->nb[1];
@@ -8245,9 +8401,9 @@ static void ggml_compute_forward_flash_ff_f16(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int D = nea0;
-    //const int N = nea1;
-    const int M = neb01;
+    const int64_t D = nea0;
+    //const int64_t N = nea1;
+    const int64_t M = neb01;
 
     GGML_ASSERT(ne0 == nea0);
     GGML_ASSERT(ne1 == nea1);
@@ -8303,7 +8459,7 @@ static void ggml_compute_forward_flash_ff_f16(
 
         float * S = (float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32);
 
-        for (int ic = 0; ic < neb01; ++ic) {
+        for (int64_t ic = 0; ic < neb01; ++ic) {
             // b0 indices
             const int ib03 = ia3;
             const int ib02 = ia2;
@@ -8323,7 +8479,7 @@ static void ggml_compute_forward_flash_ff_f16(
 
         ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32) + M);
 
-        for (int i = 0; i < M; i++) {
+        for (int64_t i = 0; i < M; i++) {
             S16[i] = GGML_FP32_TO_FP16(S[i]);
         }
 
@@ -8335,7 +8491,7 @@ static void ggml_compute_forward_flash_ff_f16(
             const int i2 = ia2;
             const int i3 = ia3;
 
-            for (int ic = 0; ic < nec01; ++ic) {
+            for (int64_t ic = 0; ic < nec01; ++ic) {
 
                 ggml_vec_dot_f16(neb01,
                         (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1   + i2*nb2   + i3*nb3)),
@@ -9147,8 +9303,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 
                         size_t cur = 0;
 
-                        if (node->src0->type == GGML_TYPE_F16 &&
-                                node->src1->type == GGML_TYPE_F32) {
+                        if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) {
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1; // TODO: this actually is doing nothing
@@ -9163,33 +9318,18 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 #else
                             cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
 #endif
-                        } else if (node->src0->type == GGML_TYPE_F32 &&
-                                node->src1->type == GGML_TYPE_F32) {
+                        } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
-                        } else if (node->src0->type == GGML_TYPE_Q4_0 &&
-                                node->src1->type == GGML_TYPE_F32) {
+                        } else if (quantize_fns[node->src0->type].vec_dot_q && node->src1->type == GGML_TYPE_F32) {
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
-                            } else {
-                                cur = (GGML_TYPE_SIZE[GGML_TYPE_Q4_0]*ggml_nelements(node->src1))/GGML_BLCK_SIZE[GGML_TYPE_Q4_0];
-                            }
-#else
-                            cur = (GGML_TYPE_SIZE[GGML_TYPE_Q4_0]*ggml_nelements(node->src1))/GGML_BLCK_SIZE[GGML_TYPE_Q4_0];
+                            } else
 #endif
-                        } else if (node->src0->type == GGML_TYPE_Q4_1 &&
-                                node->src1->type == GGML_TYPE_F32) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                            if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
-                                node->n_tasks = 1;
-                                cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
-                            } else {
-                                cur = (GGML_TYPE_SIZE[GGML_TYPE_Q4_1]*ggml_nelements(node->src1))/GGML_BLCK_SIZE[GGML_TYPE_Q4_1];
+                            {
+                                cur = GGML_TYPE_SIZE[node->src0->type]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[node->src0->type];
                             }
-#else
-                            cur = (GGML_TYPE_SIZE[GGML_TYPE_Q4_1]*ggml_nelements(node->src1))/GGML_BLCK_SIZE[GGML_TYPE_Q4_1];
-#endif
                         } else {
                             GGML_ASSERT(false);
                         }
@@ -9254,7 +9394,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 
                         size_t cur = 0;
 
-                        const int ne11 = ggml_up(node->src1->ne[1], GGML_SOFT_MAX_UNROLL);
+                        const int64_t ne11 = ggml_up(node->src1->ne[1], GGML_SOFT_MAX_UNROLL);
 
                         if (node->src1->type == GGML_TYPE_F32) {
                             cur  = sizeof(float)*ne11*node->n_tasks; // TODO: this can become (n_tasks-1)
@@ -9513,7 +9653,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
 
         perf_total_per_op_us[node->op] += node->perf_time_us;
 
-        GGML_PRINT(" - %3d: [ %6d, %6d, %6d] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
+        GGML_PRINT(" - %3d: [ %" PRId64 ", %" PRId64 ", %" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
                 GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
@@ -9527,7 +9667,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * node = cgraph->leafs[i];
 
-        GGML_PRINT(" - %3d: [ %6d, %6d] %8s\n",
+        GGML_PRINT(" - %3d: [ %" PRId64 ", %" PRId64 "] %8s\n",
                 i,
                 node->ne[0], node->ne[1],
                 GGML_OP_LABEL[node->op]);
@@ -9598,7 +9738,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
         fprintf(fp, "  \"%p\" [ \
 style = filled; fillcolor = %s; shape = record; \
-label=\"%d [%d, %d] | <x>%s",
+label=\"%d [%" PRId64 ", %" PRId64 "] | <x>%s",
                 (void *) node, color,
                 i, node->ne[0], node->ne[1],
                 GGML_OP_SYMBOL[node->op]);
@@ -9619,11 +9759,11 @@ label=\"%d [%d, %d] | <x>%s",
             fprintf(fp, "  \"%p\" [ \
 style = filled; fillcolor = %s; shape = record; \
 label=\"<x>%.1e\"; ]\n",
-                    (void *) node, color, ggml_get_f32_1d(node, 0));
+                    (void *) node, color, (double)ggml_get_f32_1d(node, 0));
         } else {
             fprintf(fp, "  \"%p\" [ \
 style = filled; fillcolor = %s; shape = record; \
-label=\"<x>CONST %d [%d, %d]\"; ]\n",
+label=\"<x>CONST %d [%" PRId64 ", %" PRId64 "]\"; ]\n",
                     (void *) node, color,
                     i, node->ne[0], node->ne[1]);
         }
@@ -9687,9 +9827,9 @@ label=\"<x>CONST %d [%d, %d]\"; ]\n",
 static void ggml_opt_set_params(int np, struct ggml_tensor * const ps[], const float * x) {
     int i = 0;
     for (int p = 0; p < np; ++p) {
-        const int ne = ggml_nelements(ps[p]) ;
+        const int64_t ne = ggml_nelements(ps[p]) ;
         // TODO: add function to set tensor from array
-        for (int j = 0; j < ne; ++j) {
+        for (int64_t j = 0; j < ne; ++j) {
             ggml_set_f32_1d(ps[p], j, x[i++]);
         }
     }
@@ -9698,9 +9838,9 @@ static void ggml_opt_set_params(int np, struct ggml_tensor * const ps[], const f
 static void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float * x) {
     int i = 0;
     for (int p = 0; p < np; ++p) {
-        const int ne = ggml_nelements(ps[p]) ;
+        const int64_t ne = ggml_nelements(ps[p]) ;
         // TODO: add function to get all elements at once
-        for (int j = 0; j < ne; ++j) {
+        for (int64_t j = 0; j < ne; ++j) {
             x[i++] = ggml_get_f32_1d(ps[p], j);
         }
     }
@@ -9709,9 +9849,9 @@ static void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float *
 static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g) {
     int i = 0;
     for (int p = 0; p < np; ++p) {
-        const int ne = ggml_nelements(ps[p]) ;
+        const int64_t ne = ggml_nelements(ps[p]) ;
         // TODO: add function to get all elements at once
-        for (int j = 0; j < ne; ++j) {
+        for (int64_t j = 0; j < ne; ++j) {
             g[i++] = ggml_get_f32_1d(ps[p]->grad, j);
         }
     }
@@ -9857,7 +9997,7 @@ static enum ggml_opt_result ggml_opt_adam(
             if (params.past <= t) {
                 const float rate = (pf[t%params.past] - fx)/fx;
 
-                if (fabs(rate) < params.delta) {
+                if (fabsf(rate) < params.delta) {
                     return GGML_OPT_OK;
                 }
             }
@@ -9936,7 +10076,7 @@ static enum ggml_opt_result linesearch_backtracking(
     const float dec = 0.5f;
     const float inc = 2.1f;
 
-    if (*step <= 0.) {
+    if (*step <= 0.f) {
         return GGML_LINESEARCH_INVALID_PARAMETERS;
     }
 
@@ -10024,7 +10164,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         struct ggml_cgraph * gb) {
     if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE ||
         params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
-        if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1. <= params.lbfgs.wolfe) {
+        if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) {
             return GGML_OPT_INVALID_WOLFE;
         }
     }
@@ -10145,8 +10285,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
         GGML_PRINT_DEBUG("f = %10.6f\n", ggml_get_f32_1d(f, 0));
 
-        if (xnorm < 1.0) {
-            xnorm = 1.0;
+        if (xnorm < 1.0f) {
+            xnorm = 1.0f;
         }
         if (gnorm/xnorm <= params.lbfgs.eps) {
             // converged
@@ -10159,7 +10299,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
             if (params.past <= k) {
                 const float rate = (pf[k%params.past] - fx)/fx;
 
-                if (fabs(rate) < params.delta) {
+                if (fabsf(rate) < params.delta) {
                     return GGML_OPT_OK;
                 }
             }
@@ -10309,6 +10449,7 @@ enum ggml_opt_result ggml_opt(
         struct ggml_init_params params_ctx = {
             .mem_size   = 16*1024*1024,
             .mem_buffer = NULL,
+            .no_alloc   = false,
         };
 
         ctx = ggml_init(params_ctx);
@@ -10355,64 +10496,50 @@ enum ggml_opt_result ggml_opt(
 
 ////////////////////////////////////////////////////////////////////////////////
 
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist) {
-    const int nb = k / qk;
-    const size_t bs = (sizeof(float) + sizeof(uint8_t)*qk/2);
-    const size_t row_size = nb*bs;
-
-    assert(k % qk == 0);
-
-    char * pdst = (char *) dst;
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
+    assert(k % QK == 0);
+    const int nb = k / QK;
 
     for (int j = 0; j < n; j += k) {
-        uint8_t * pd = (uint8_t *) (pdst + (j/k)*row_size + 0*bs);
-        uint8_t * pb = (uint8_t *) (pdst + (j/k)*row_size + 0*bs + sizeof(float));
+        block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK;
 
-        quantize_row_q4_0_reference(src + j, pd, k);
+        quantize_row_q4_0_reference(src + j, y, k);
 
         for (int i = 0; i < nb; i++) {
-            for (int l = 0; l < qk; l += 2) {
-                const uint8_t vi0 = pb[l/2] & 0xF;
-                const uint8_t vi1 = pb[l/2] >> 4;
+            for (int l = 0; l < QK; l += 2) {
+                const uint8_t vi0 = y[i].qs[l/2] & 0xF;
+                const uint8_t vi1 = y[i].qs[l/2] >> 4;
 
                 hist[vi0]++;
                 hist[vi1]++;
             }
-            pb += bs;
         }
     }
 
-    return (n/k)*row_size;
+    return (n/QK*sizeof(block_q4_0));
 }
 
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist) {
-    const int nb = k / qk;
-    const size_t bs = (2*sizeof(float) + sizeof(uint8_t)*qk/2);
-    const size_t row_size = nb*bs;
-
-    assert(k % qk == 0);
-
-    char * pdst = (char *) dst;
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) {
+    assert(k % QK == 0);
+    const int nb = k / QK;
 
     for (int j = 0; j < n; j += k) {
-        uint8_t * pd = (uint8_t *) (pdst + (j/k)*row_size + 0*bs);
-        uint8_t * pb = (uint8_t *) (pdst + (j/k)*row_size + 0*bs + 2*sizeof(float));
+        block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK;
 
-        quantize_row_q4_1(src + j, pd, k);
+        quantize_row_q4_1_reference(src + j, y, k);
 
         for (int i = 0; i < nb; i++) {
-            for (int l = 0; l < qk; l += 2) {
-                const uint8_t vi0 = pb[l/2] & 0xF;
-                const uint8_t vi1 = pb[l/2] >> 4;
+            for (int l = 0; l < QK; l += 2) {
+                const uint8_t vi0 = y[i].qs[l/2] & 0xF;
+                const uint8_t vi1 = y[i].qs[l/2] >> 4;
 
                 hist[vi0]++;
                 hist[vi1]++;
             }
-            pb += bs;
         }
     }
 
-    return (n/k)*row_size;
+    return (n/QK*sizeof(block_q4_1));
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/ggml.h b/ggml.h
index ddb97318b..ad962b109 100644
--- a/ggml.h
+++ b/ggml.h
@@ -258,11 +258,11 @@ struct ggml_tensor {
     enum ggml_type type;
 
     int    n_dims;
-    int    ne[GGML_MAX_DIMS]; // number of elements
-    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
-                              // nb[0] = sizeof(type)
-                              // nb[1] = nb[0]   * ne[0] + padding
-                              // nb[i] = nb[i-1] * ne[i-1]
+    int64_t ne[GGML_MAX_DIMS]; // number of elements
+    size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
+                               // nb[0] = sizeof(type)
+                               // nb[1] = nb[0]   * ne[0] + padding
+                               // nb[i] = nb[i-1] * ne[i-1]
 
     // compute data
     enum ggml_op op;
@@ -316,6 +316,7 @@ struct ggml_init_params {
     // memory pool
     size_t mem_size;   // bytes
     void * mem_buffer; // if NULL, memory will be allocated internally
+    bool   no_alloc;   // don't allocate memory for the tensor data
 };
 
 void    ggml_time_init(void); // call this once at the beginning of the program
@@ -327,8 +328,8 @@ int64_t ggml_cycles_per_ms(void);
 void ggml_print_object (const struct ggml_object * obj);
 void ggml_print_objects(const struct ggml_context * ctx);
 
-int    ggml_nelements(const struct ggml_tensor * tensor);
-size_t ggml_nbytes   (const struct ggml_tensor * tensor);
+int64_t ggml_nelements(const struct ggml_tensor * tensor);
+size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
 
 int    ggml_blck_size (enum ggml_type type);
 size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
@@ -344,39 +345,43 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
 bool ggml_mlock_supported(void);
-bool ggml_mlock(struct ggml_context * ctx, char ** err_p);
+bool ggml_mlock(
+        struct ggml_context * ctx,
+        const void *opt_extra_addr,
+        size_t opt_extra_len,
+        char **err_p);
 
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int *ne);
+        const int64_t *ne);
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0);
+        int64_t ne0);
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1);
+        int64_t ne0,
+        int64_t ne1);
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2);
 
 struct ggml_tensor * ggml_new_tensor_4d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2,
-        int    ne3);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2,
+        int64_t ne3);
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
@@ -526,30 +531,30 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1);
+        int64_t               ne0,
+        int64_t               ne1);
 
 // return view(a)
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape_3d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2);
 
 // offset in bytes
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
+        int64_t               ne0,
         size_t                offset);
 
 struct ggml_tensor * ggml_view_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
+        int64_t               ne0,
+        int64_t               ne1,
         size_t                nb1, // row stride in bytes
         size_t                offset);
 
@@ -748,8 +753,8 @@ enum ggml_opt_result ggml_opt(
 // quantization
 //
 
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
 
 //
 // system info
diff --git a/llama.cpp b/llama.cpp
index 2bd520353..878907185 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -12,6 +12,19 @@
 #include <cassert>
 #include <cstring>
 
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#else
+#include <sys/types.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <fcntl.h>
+#endif
+
+#define Min(X, Y) ((Y) > (X) ? (X) : (Y))
+#define Max(X, Y) ((Y) < (X) ? (X) : (Y))
+
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
 
@@ -142,6 +155,10 @@ struct llama_model {
     // the model memory buffer
     std::vector<uint8_t> buf;
 
+    // model memory mapped file
+    void * mm_addr = NULL;
+    uint64_t mm_length = 0;
+
     // tensors
     int n_loaded;
     std::unordered_map<std::string, struct ggml_tensor *> tensors;
@@ -165,6 +182,7 @@ struct llama_context {
 
     int64_t t_load_us = 0;
     int64_t t_start_us = 0;
+    bool has_evaluated_once = false;
 
     int64_t t_sample_us = 0;
     int64_t t_eval_us   = 0;
@@ -206,7 +224,7 @@ struct llama_context {
         }
 
         if (buf_last >= 0) {
-            buf_max_size[buf_last] = std::max(buf_max_size[buf_last], last_size);
+            buf_max_size[buf_last] = Max(buf_max_size[buf_last], last_size);
         }
 
         buf_last = i;
@@ -238,14 +256,15 @@ static bool kv_cache_init(
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
-    const int n_mem      = n_layer*n_ctx;
-    const int n_elements = n_embd*n_mem;
+    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
+    const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size();
     params.mem_buffer = cache.buf.data();
+    params.no_alloc   = false;
 
     cache.ctx = ggml_init(params);
 
@@ -288,6 +307,58 @@ struct llama_context_params llama_context_default_params() {
 // model loading
 //
 
+static void *mmap_file(const char *fname, uint64_t *mm_length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    HANDLE hFile = CreateFileA(fname,
+                               GENERIC_READ,
+                               FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE,
+                               NULL,
+                               OPEN_EXISTING,
+                               FILE_ATTRIBUTE_NORMAL | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED,
+                               NULL);
+    if (hFile == INVALID_HANDLE_VALUE) return 0;
+    LARGE_INTEGER fileSize;
+    fileSize.QuadPart = -1;
+    GetFileSizeEx(hFile, &fileSize);
+    int64_t length = fileSize.QuadPart;
+    HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+    CloseHandle(hFile);
+    if (!hMapping) return 0;
+    void *addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+    CloseHandle(hMapping);
+    if (!addr) return 0;
+#else
+    int fd = open(fname, O_RDONLY);
+    if (fd == -1) return 0;
+    int64_t length = lseek(fd, 0, SEEK_END);
+    void *addr = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, 0);
+    close(fd);
+    if (addr == MAP_FAILED) return 0;
+#endif
+    *mm_length = length;
+    return addr;
+}
+
+static void munmap_file(void * addr, size_t length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    UnmapViewOfFile(addr);
+#else
+    munmap(addr, length);
+#endif
+}
+
+static bool report_bad_magic(const char *path, uint32_t got, uint32_t want) {
+    fprintf(stderr,
+            "%s: invalid model file (bad magic [got %#x want %#x])\n"
+            "\tyou most likely need to regenerate your ggml files\n"
+            "\tthe benefit is you'll get 10-100x faster load times\n"
+            "\tsee https://github.com/ggerganov/llama.cpp/issues/91\n"
+            "\tuse convert-pth-to-ggml.py to regenerate from original pth\n"
+            "\tuse migrate-ggml-2023-03-30-pr613.py if you deleted originals\n",
+            path, got, want);
+    return false;
+}
+
 static bool llama_model_load(
         const std::string & fname,
         llama_context & lctx,
@@ -299,34 +370,35 @@ static bool llama_model_load(
         void *progress_callback_user_data) {
     fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
 
-    const int64_t t_start_us = ggml_time_us();
-
-    lctx.t_start_us = t_start_us;
-
-    std::vector<char> f_buf(1024*1024);
+    lctx.t_start_us = ggml_time_us();
 
     auto & model = lctx.model;
     auto & vocab = lctx.vocab;
 
     auto fin = std::ifstream(fname, std::ios::binary);
-    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
     if (!fin) {
         fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
         return false;
     }
 
+    std::vector<char> f_buf(1024*1024);
+    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+
+    fin.seekg(0, fin.end);
+    const size_t file_size = fin.tellg();
+    fin.seekg(0);
+
     // verify magic
     {
         uint32_t magic;
         fin.read((char *) &magic, sizeof(magic));
         if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) {
-            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files!)\n",
+            fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files or convert them with convert-unversioned-ggml-to-ggml.py!)\n",
                     __func__, fname.c_str());
             return false;
         }
         if (magic != LLAMA_FILE_MAGIC) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
+            return report_bad_magic(fname.c_str(), magic, LLAMA_FILE_MAGIC);
         }
 
         uint32_t format_version;
@@ -449,43 +521,24 @@ static bool llama_model_load(
                 }
     }
 
+    // map model into memory
+    char *mm_addr = NULL;
+    model.mm_addr = mmap_file(fname.c_str(), &model.mm_length);
+    if (model.mm_addr == NULL) {
+        fprintf(stderr, "%s: failed to mmap '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+    mm_addr = (char *)model.mm_addr;
+    fprintf(stderr, "%s: ggml map size = %6.2f MB\n", __func__, model.mm_length/(1024.0*1024.0));
+
     auto & ctx = model.ctx;
 
     size_t ctx_size = 0;
-
     {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
+        const auto &hparams = model.hparams;
         const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // tok_embeddings
-
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm
-
-        ctx_size += n_embd*n_vocab*ggml_type_sizef(vtype); // output
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm
-
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo
-
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm
-
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
-        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
-
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
-
         ctx_size += (5 + 10*n_layer)*256; // object overhead
-
-        fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+        fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0);
     }
 
     // print memory requirements
@@ -495,6 +548,7 @@ static bool llama_model_load(
         // this is the total memory required to run the inference
         const size_t mem_required =
             ctx_size +
+            model.mm_length +
             MEM_REQ_SCRATCH0.at(model.type) +
             MEM_REQ_SCRATCH1.at(model.type) +
             MEM_REQ_EVAL.at    (model.type);
@@ -514,6 +568,7 @@ static bool llama_model_load(
         struct ggml_init_params params = {
             /*.mem_size   =*/ lctx.model.buf.size(),
             /*.mem_buffer =*/ lctx.model.buf.data(),
+            /*.no_alloc   =*/ true,
         };
 
         model.ctx = ggml_init(params);
@@ -576,234 +631,106 @@ static bool llama_model_load(
         }
     }
 
-    const size_t file_offset = fin.tellg();
-
-    fin.close();
-
     std::vector<uint8_t> tmp;
 
     if (progress_callback) {
         progress_callback(0.0, progress_callback_user_data);
     }
 
-    for (int i = 0; i < n_parts; ++i) {
-        const int part_id = i;
-        //const int part_id = n_parts - i - 1;
+    fprintf(stderr, "%s: loading tensors from '%s'\n", __func__, fname.c_str());
 
-        std::string fname_part = fname;
-        if (i > 0) {
-            fname_part += "." + std::to_string(i);
-        }
+    // load weights
+    {
+        size_t total_size = 0;
+        model.n_loaded = 0;
 
-        fprintf(stderr, "%s: loading model part %d/%d from '%s'\n", __func__, i+1, n_parts, fname_part.c_str());
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
 
-        fin = std::ifstream(fname_part, std::ios::binary);
-        fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
 
-        fin.seekg(0, fin.end);
-        const size_t file_size = fin.tellg();
-
-        fin.seekg(file_offset);
-
-        // load weights
-        {
-            size_t total_size = 0;
-
-            model.n_loaded = 0;
-
-            fprintf(stderr, "%s: ", __func__);
-
-            while (true) {
-                int32_t n_dims;
-                int32_t length;
-                int32_t ftype;
-
-                fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
-                fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-                fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
-
-                if (fin.eof()) {
-                    break;
-                }
-
-                int32_t nelements = 1;
-                int32_t ne[2] = { 1, 1 };
-                for (int i = 0; i < n_dims; ++i) {
-                    fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                    nelements *= ne[i];
-                }
-
-                std::string name(length, 0);
-                fin.read(&name[0], length);
-
-                if (model.tensors.find(name.data()) == model.tensors.end()) {
-                    fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                    return false;
-                }
-
-                // split_type = 0: split by columns
-                // split_type = 1: split by rows
-                int split_type = 0;
-
-                // split_type = 0:
-                // regex:
-                //   - tok_embeddings.*
-                //   - layers.*.attention.wo.weight
-                //   - layers.*.feed_forward.w2.weight
-
-                // split_type = 1:
-                // regex:
-                //   - output.*
-                //   - layers.*.attention.wq.weight
-                //   - layers.*.attention.wk.weight
-                //   - layers.*.attention.wv.weight
-                //   - layers.*.feed_forward.w1.weight
-                //   - layers.*.feed_forward.w3.weight
-                if (name.find("tok_embeddings") != std::string::npos) {
-                    split_type = 0;
-                } else if (name.find("layers") != std::string::npos) {
-                    if (name.find("attention.wo.weight") != std::string::npos) {
-                        split_type = 0;
-                    } else if (name.find("feed_forward.w2.weight") != std::string::npos) {
-                        split_type = 0;
-                    } else {
-                        split_type = 1;
-                    }
-                } else if (name.find("output") != std::string::npos) {
-                    split_type = 1;
-                }
-
-                auto tensor = model.tensors[name.data()];
-
-                if (n_dims == 1) {
-                    if (ggml_nelements(tensor) != nelements) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                        return false;
-                    }
-                } else {
-                    if (ggml_nelements(tensor)/n_parts != nelements) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                        return false;
-                    }
-                }
-
-                if (n_dims == 1) {
-                    if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                                __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
-                        return false;
-                    }
-                } else {
-                    if (split_type == 0) {
-                        if (tensor->ne[0]/n_parts != ne[0] || tensor->ne[1] != ne[1]) {
-                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                                    __func__, name.data(), tensor->ne[0]/n_parts, tensor->ne[1], ne[0], ne[1]);
-                            return false;
-                        }
-                    } else {
-                        if (tensor->ne[0] != ne[0] || tensor->ne[1]/n_parts != ne[1]) {
-                            fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                                    __func__, name.data(), tensor->ne[0], tensor->ne[1]/n_parts, ne[0], ne[1]);
-                            return false;
-                        }
-                    }
-                }
-
-                if (0) {
-                    static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
-                    fprintf(stderr, "%24s - [%5d, %5d], type = %6s, split = %d\n", name.data(), ne[0], ne[1], ftype_str[ftype], split_type);
-                }
-
-                size_t bpe = 0;
-
-                switch (ftype) {
-                    case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
-                    case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
-                    case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
-                    case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
-                    default:
-                            {
-                                fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
-                                return false;
-                            }
-                };
-
-                if (n_dims == 1 || n_parts == 1) {
-                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                                __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
-                        return false;
-                    }
-
-                    if (part_id == 0) {
-                        fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-                    } else {
-                        fin.seekg(ggml_nbytes(tensor), std::ios::cur);
-                    }
-
-                    total_size += ggml_nbytes(tensor);
-                } else {
-                    if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)/n_parts) {
-                        fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
-                                __func__, name.data(), ggml_nbytes(tensor)/n_parts, nelements*bpe);
-                        return false;
-                    }
-
-                    if (split_type == 0) {
-                        const int np0 = ne[0];
-
-                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-                        assert(row_size == tensor->nb[1]);
-
-                        for (int i1 = 0; i1 < ne[1]; ++i1) {
-                            const size_t offset_row = i1*row_size;
-                            const size_t offset = offset_row + ((part_id*np0)/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-                            fin.read(reinterpret_cast<char *>(tensor->data) + offset, row_size/n_parts);
-                        }
-                    } else {
-                        const int np1 = ne[1];
-
-                        const size_t row_size = (tensor->ne[0]/ggml_blck_size(tensor->type))*ggml_type_size(tensor->type);
-
-                        for (int i1 = 0; i1 < ne[1]; ++i1) {
-                            const size_t offset_row = (i1 + part_id*np1)*row_size;
-                            fin.read(reinterpret_cast<char *>(tensor->data) + offset_row, row_size);
-                        }
-                    }
-
-                    total_size += ggml_nbytes(tensor)/n_parts;
-                }
-
-                //fprintf(stderr, "%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
-                model.n_loaded++;
-
-                // progress
-                if (progress_callback) {
-                    double current_file_progress = double(size_t(fin.tellg()) - file_offset) / double(file_size - file_offset);
-                    double current_progress = (double(i) + current_file_progress) / double(n_parts);
-                    progress_callback(current_progress, progress_callback_user_data);
-                }
-                if (model.n_loaded % 8 == 0) {
-                    fprintf(stderr, ".");
-                    fflush(stderr);
-                }
+            if (fin.eof()) {
+                break;
             }
 
-            fprintf(stderr, " done\n");
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
 
-            fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, model.n_loaded);
-            if (model.n_loaded == 0) {
-                fprintf(stderr, "%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__);
-            } else if (model.n_loaded != (int) model.tensors.size()) {
-                fprintf(stderr, "%s: ERROR not all tensors loaded from model file - expected %zu, got %d\n", __func__, model.tensors.size(), model.n_loaded);
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
                 return false;
             }
+
+            auto tensor = model.tensors[name.data()];
+
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
+                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+            if (0) {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                fprintf(stderr, "%24s - [%5d, %5d], type = %6s\n", name.data(), ne[0], ne[1], ftype_str[ftype]);
+            }
+
+            switch (ftype) {
+                case 0:  // f32
+                case 1:  // f16
+                    break;
+                case 2:  // q4_0
+                case 3:  // q4_1
+                    assert(ne[0] % 64 == 0);
+                    break;
+                default:
+                    fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                    return false;
+            };
+
+            // load the tensor data into memory without copying or reading it
+            size_t offset = fin.tellg();
+            size_t tensor_data_size = ggml_nbytes(tensor);
+            offset = (offset + 31) & -32;
+            tensor->data = mm_addr + offset;
+            fin.seekg(offset + tensor_data_size);
+            total_size += tensor_data_size;
+            model.n_loaded++;
+
+            // progress
+            if (progress_callback) {
+                double current_progress = size_t(fin.tellg()) / double(file_size);
+                progress_callback(current_progress, progress_callback_user_data);
+            }
         }
 
         fin.close();
+
+        fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, model.n_loaded);
+        if (model.n_loaded == 0) {
+            fprintf(stderr, "%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__);
+        } else if (model.n_loaded != (int) model.tensors.size()) {
+            fprintf(stderr, "%s: ERROR not all tensors loaded from model file - expected %zu, got %d\n", __func__, model.tensors.size(), model.n_loaded);
+            return false;
+        }
     }
 
-    lctx.t_load_us = ggml_time_us() - t_start_us;
+    // loading time will be recalculate after the first eval, so
+    // we take page faults deferred by mmap() into consideration
+    lctx.t_load_us = ggml_time_us() - lctx.t_start_us;
 
     if (progress_callback) {
         progress_callback(1.0, progress_callback_user_data);
@@ -849,6 +776,7 @@ static bool llama_eval_internal(
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_compute.size(),
         /*.mem_buffer =*/ buf_compute.data(),
+        /*.no_alloc   =*/ false,
     };
 
     struct ggml_context * ctx0 = ggml_init(params);
@@ -856,7 +784,7 @@ static bool llama_eval_internal(
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
     ggml_cgraph gf = {};
-    gf.n_threads = N > 255 && ggml_cpu_has_blas() ? 1 : n_threads;
+    gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
@@ -922,7 +850,7 @@ static bool llama_eval_internal(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)));
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
@@ -1126,7 +1054,7 @@ struct llama_tokenizer {
         size_t offs = 0;
         while (offs < text.size()) {
             llama_sp_symbol sym;
-            size_t char_len = std::min(text.size() - offs, utf8_len(text[offs]));
+            size_t char_len = Min(text.size() - offs, utf8_len(text[offs]));
             sym.text = text.c_str() + offs;
             sym.n = char_len;
             offs += char_len;
@@ -1240,12 +1168,12 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
 // sampling
 //
 
-static void sample_top_k(std::vector<std::pair<double, llama_vocab::id>> & logits_id, int top_k) {
+static void sample_top_k(std::vector<std::pair<float, llama_vocab::id>> & logits_id, int top_k) {
     // find the top k tokens
     std::partial_sort(
             logits_id.begin(),
             logits_id.begin() + top_k, logits_id.end(),
-            [](const std::pair<double, llama_vocab::id> & a, const std::pair<double, llama_vocab::id> & b) {
+            [](const std::pair<float, llama_vocab::id> & a, const std::pair<float, llama_vocab::id> & b) {
         return a.first > b.first;
     });
 
@@ -1256,9 +1184,9 @@ static llama_vocab::id llama_sample_top_p_top_k(
         llama_context & lctx,
         const std::vector<llama_vocab::id> & last_n_tokens,
         int top_k,
-        double top_p,
-        double temp,
-        double repeat_penalty) {
+        float top_p,
+        float temp,
+        float repeat_penalty) {
     auto & rng = lctx.rng;
 
     const int n_logits = lctx.model.hparams.n_vocab;
@@ -1266,17 +1194,17 @@ static llama_vocab::id llama_sample_top_p_top_k(
     const auto & logits = lctx.logits;
     const auto * plogits = logits.data() + logits.size() - n_logits;
 
-    std::vector<std::pair<double, llama_vocab::id>> logits_id;
+    std::vector<std::pair<float, llama_vocab::id>> logits_id;
     logits_id.reserve(n_logits);
 
     {
-        const double scale = 1.0/temp;
+        const float scale = 1.0f/temp;
         for (int i = 0; i < n_logits; ++i) {
             // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
             // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
             if (std::find(last_n_tokens.begin(), last_n_tokens.end(), i) != last_n_tokens.end()) {
                 // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
-                if (plogits[i] < 0.0) {
+                if (plogits[i] < 0.0f) {
                     logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
                 } else {
                     logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
@@ -1289,18 +1217,18 @@ static llama_vocab::id llama_sample_top_p_top_k(
 
     sample_top_k(logits_id, top_k);
 
-    double maxl = -std::numeric_limits<double>::infinity();
+    float maxl = -std::numeric_limits<float>::infinity();
     for (const auto & kv : logits_id) {
-        maxl = std::max(maxl, kv.first);
+        maxl = Max(maxl, kv.first);
     }
 
     // compute probs for the top k tokens
-    std::vector<double> probs;
+    std::vector<float> probs;
     probs.reserve(logits_id.size());
 
     double sum = 0.0;
     for (const auto & kv : logits_id) {
-        double p = exp(kv.first - maxl);
+        const float p = expf(kv.first - maxl);
         probs.push_back(p);
         sum += p;
     }
@@ -1310,8 +1238,8 @@ static llama_vocab::id llama_sample_top_p_top_k(
         p /= sum;
     }
 
-    if (top_p < 1.0f) {
-        double cumsum = 0.0f;
+    if (top_p < 1.0) {
+        double cumsum = 0.0;
         for (int i = 0; i < (int) probs.size(); i++) {
             cumsum += probs[i];
             if (cumsum >= top_p) {
@@ -1345,7 +1273,7 @@ static llama_vocab::id llama_sample_top_p_top_k(
 //
 
 // TODO: reuse code from the llama_model_load() somehow
-bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype, int qk) {
+static bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype) {
     ggml_type type = GGML_TYPE_Q4_1;
 
     switch (itype) {
@@ -1385,8 +1313,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             return false;
         }
         if (magic != LLAMA_FILE_MAGIC) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
-            return false;
+            return report_bad_magic(fname_inp.c_str(), magic, LLAMA_FILE_MAGIC);
         }
 
         fout.write((char *) &magic, sizeof(magic));
@@ -1444,7 +1371,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             return false;
         }
 
-        std::string word;
+        std::vector<char> word(32);
         vocab.id_to_token.resize(n_vocab);
         for (int i = 0; i < n_vocab; i++) {
             uint32_t len;
@@ -1452,17 +1379,17 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             fout.write((char *) &len, sizeof(len));
 
             word.resize(len);
-            finp.read ((char *) word.data(), len);
-            fout.write((char *) word.data(), len);
+            finp.read ((char *) &word[0], len);
+            fout.write((char *) &word[0], len);
 
             float score;
             finp.read ((char *) &score, sizeof(score));
             fout.write((char *) &score, sizeof(score));
 
-            vocab.token_to_id[word] = i;
+            vocab.token_to_id[word.data()] = i;
 
             auto &tok_score = vocab.id_to_token[i];
-            tok_score.tok = word;
+            tok_score.tok = word.data();
             tok_score.score = score;
         }
     }
@@ -1503,6 +1430,13 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             std::string name(length, 0);
             finp.read (&name[0], length);
 
+            {
+                // ensure tensor data is aligned
+                uint64_t offset = finp.tellg();
+                offset = (offset + 31) & -32;
+                finp.seekg(offset);
+            }
+
             {
                 static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
                 printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
@@ -1558,6 +1492,13 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
             }
             fout.write(&name[0], length);
 
+            {
+                // ensure tensor data is aligned
+                uint64_t offset = fout.tellp();
+                offset = (offset + 31) & -32;
+                fout.seekp(offset);
+            }
+
             if (quantize) {
                 printf("quantizing .. ");
                 work.resize(nelements); // for quantization
@@ -1568,11 +1509,11 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
                 switch (type) {
                     case GGML_TYPE_Q4_0:
                         {
-                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                         } break;
                     case GGML_TYPE_Q4_1:
                         {
-                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], qk, hist_cur.data());
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                         } break;
                     default:
                         {
@@ -1590,7 +1531,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
                 }
 
                 for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                    printf("%5.3f ", hist_cur[i] / float(nelements));
                 }
                 printf("\n");
             } else {
@@ -1613,7 +1554,7 @@ bool llama_model_quantize_internal(const std::string & fname_inp, const std::str
 
             printf("%s: hist: ", __func__);
             for (int i = 0; i < (int) hist_all.size(); ++i) {
-                printf("%5.3f ", hist_all[i] / (float)sum_all);
+                printf("%5.3f ", hist_all[i] / float(sum_all));
             }
             printf("\n");
         }
@@ -1655,7 +1596,10 @@ struct llama_context * llama_init_from_file(
 
     if (params.use_mlock) {
         char *err;
-        if (!ggml_mlock(ctx->model.ctx, &err)) {
+        if (!ggml_mlock(ctx->model.ctx,
+                        ctx->model.mm_addr,
+                        ctx->model.mm_length,
+                        &err)) {
             fprintf(stderr, "%s\n", err);
             free(err);
             llama_free(ctx);
@@ -1664,7 +1608,7 @@ struct llama_context * llama_init_from_file(
     }
 
     // reserve memory for context buffers
-    {
+    if (!params.vocab_only) {
         if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -1705,15 +1649,18 @@ void llama_free(struct llama_context * ctx) {
         ggml_free(ctx->model.ctx);
     }
 
+    if (ctx->model.mm_addr) {
+        munmap_file(ctx->model.mm_addr, ctx->model.mm_length);
+    }
+
     delete ctx;
 }
 
 int llama_model_quantize(
         const char * fname_inp,
         const char * fname_out,
-               int   itype,
-               int   qk) {
-    if (!llama_model_quantize_internal(fname_inp, fname_out, itype, qk)) {
+               int   itype) {
+    if (!llama_model_quantize_internal(fname_inp, fname_out, itype)) {
         fprintf(stderr, "%s: failed to quantize\n", __func__);
         return 1;
     }
@@ -1721,6 +1668,33 @@ int llama_model_quantize(
     return 0;
 }
 
+// Returns the KV cache that will contain the context for the
+// ongoing prediction with the model.
+const uint8_t * llama_get_kv_cache(struct llama_context * ctx) {
+    return ctx->model.kv_self.buf.data();
+}
+
+// Returns the size of the KV cache
+size_t llama_get_kv_cache_size(struct llama_context * ctx) {
+    return ctx->model.kv_self.buf.size();
+}
+
+int llama_get_kv_cache_token_count(struct llama_context * ctx) {
+    return ctx->model.kv_self.n;
+}
+
+// Sets the KV cache containing the current context for the model
+void llama_set_kv_cache(
+        struct llama_context * ctx,
+               const uint8_t * kv_cache,
+                      size_t   n_size,
+                         int   n_token_count) {
+    // Make sure we have the same kv cache setup
+    LLAMA_ASSERT(ctx->model.kv_self.buf.size() == n_size);
+    memcpy(ctx->model.kv_self.buf.data(), kv_cache, n_size);
+    ctx->model.kv_self.n = n_token_count;
+}
+
 int llama_eval(
         struct llama_context * ctx,
            const llama_token * tokens,
@@ -1731,7 +1705,11 @@ int llama_eval(
         fprintf(stderr, "%s: failed to eval\n", __func__);
         return 1;
     }
-
+    // get a more accurate load time, upon first eval
+    if (!ctx->has_evaluated_once) {
+        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
+        ctx->has_evaluated_once = true;
+    }
     return 0;
 }
 
@@ -1796,9 +1774,9 @@ llama_token llama_sample_top_p_top_k(
       const llama_token * last_n_tokens_data,
                     int   last_n_tokens_size,
                     int   top_k,
-                 double   top_p,
-                 double   temp,
-                 double   repeat_penalty) {
+                  float   top_p,
+                  float   temp,
+                  float   repeat_penalty) {
     const int64_t t_start_sample_us = ggml_time_us();
 
     llama_token result = 0;
@@ -1824,21 +1802,20 @@ llama_token llama_sample_top_p_top_k(
 void llama_print_timings(struct llama_context * ctx) {
     const int64_t t_end_us = ggml_time_us();
 
-    const int32_t n_sample = std::max(1, ctx->n_sample);
-    const int32_t n_eval   = std::max(1, ctx->n_eval);
-    const int32_t n_p_eval = std::max(1, ctx->n_p_eval);
+    const int32_t n_sample = Max(1, ctx->n_sample);
+    const int32_t n_eval   = Max(1, ctx->n_eval);
+    const int32_t n_p_eval = Max(1, ctx->n_p_eval);
 
     fprintf(stderr, "\n");
-    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0f);
-    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_sample_us, n_sample, 1e-3f * ctx->t_sample_us / n_sample);
-    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3f * ctx->t_p_eval_us, n_p_eval, 1e-3f * ctx->t_p_eval_us / n_p_eval);
-    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3f * ctx->t_eval_us,   n_eval,   1e-3f * ctx->t_eval_us   / n_eval);
-    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
+    fprintf(stderr, "%s:        load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0);
+    fprintf(stderr, "%s:      sample time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3 * ctx->t_sample_us, n_sample, 1e-3 * ctx->t_sample_us / n_sample);
+    fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token)\n", __func__, 1e-3 * ctx->t_p_eval_us, n_p_eval, 1e-3 * ctx->t_p_eval_us / n_p_eval);
+    fprintf(stderr, "%s:        eval time = %8.2f ms / %5d runs   (%8.2f ms per run)\n",   __func__, 1e-3 * ctx->t_eval_us,   n_eval,   1e-3 * ctx->t_eval_us   / n_eval);
+    fprintf(stderr, "%s:       total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0);
 }
 
 void llama_reset_timings(struct llama_context * ctx) {
     ctx->t_start_us = ggml_time_us();
-
     ctx->t_sample_us = ctx->n_sample = 0;
     ctx->t_eval_us   = ctx->n_eval   = 0;
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
diff --git a/llama.h b/llama.h
index ebf55f41c..04e2bf71c 100644
--- a/llama.h
+++ b/llama.h
@@ -6,7 +6,7 @@
 #include <stdbool.h>
 
 #ifdef LLAMA_SHARED
-#    ifdef _WIN32
+#    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
 #            define LLAMA_API __declspec(dllexport)
 #        else
@@ -20,7 +20,7 @@
 #endif
 
 #define LLAMA_FILE_VERSION 1
-#define LLAMA_FILE_MAGIC 0x67676d66 // 'ggmf' in hex
+#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
 #define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
 
 #ifdef __cplusplus
@@ -45,7 +45,7 @@ extern "C" {
 
     } llama_token_data;
 
-    typedef void (*llama_progress_callback)(double progress, void *ctx);
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     struct llama_context_params {
         int n_ctx;   // text context
@@ -81,8 +81,24 @@ extern "C" {
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-                   int   itype,
-                   int   qk);
+                   int   itype);
+
+    // Returns the KV cache that will contain the context for the
+    // ongoing prediction with the model.
+    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
+
+    // Returns the size of the KV cache
+    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+
+    // Sets the KV cache containing the current context for the model
+    LLAMA_API void llama_set_kv_cache(
+            struct llama_context * ctx,
+                   const uint8_t * kv_cache,
+                          size_t   n_size,
+                             int   n_token_count);
 
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
@@ -135,9 +151,9 @@ extern "C" {
           const llama_token * last_n_tokens_data,
                         int   last_n_tokens_size,
                         int   top_k,
-                     double   top_p,
-                     double   temp,
-                     double   repeat_penalty);
+                      float   top_p,
+                      float   temp,
+                      float   repeat_penalty);
 
     // Performance information
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
diff --git a/migrate-ggml-2023-03-30-pr613.py b/migrate-ggml-2023-03-30-pr613.py
new file mode 100644
index 000000000..b6ef2476e
--- /dev/null
+++ b/migrate-ggml-2023-03-30-pr613.py
@@ -0,0 +1,311 @@
+# Migrate ggml file(s) with ggmf magic to ggml file with ggjt magic
+#
+# We caused a breaking change to the file format on 2023-03-30 in:
+#     https://github.com/ggerganov/llama.cpp/pull/613
+#
+# (1) If you still have the Meta LLaMA .pth files, then close this
+#     file now; you can just run `convert-pth-to-ggml.py` again to
+#     migrate to the new format. The tool is easier to use too. It
+#     isn't necessary anymore to manage split output files because
+#     the new format always combines things into a single file.
+#
+# (2) If you deleted the Meta LLaMA .pth files due to save on disk
+#     space, then this tool is intended to help you.  Please check
+#     out the instructions below.
+#
+# USAGE
+#
+#     python migrate-ggml-2023-03-30-pr613.py INPUT OUTPUT
+#
+# PREREQUISITES
+#
+#     pip install numpy
+#     cd llama.cpp
+#     make -j4
+#
+# EXAMPLE (7B MODEL)
+#
+#     # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
+#     python migrate-ggml-2023-03-30-pr613.py models/7B/ggml-model-f16.bin models/7B/ggml-model-f16-ggjt.bin
+#
+#     # check that it works
+#     ./main -m models/7B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
+#
+#     # you can delete the old files
+#     rm -f models/7B/ggml-model-f16.bin
+#     mv models/7B/ggml-model-f16-ggjt.bin models/7B/ggml-model-f16.bin
+#
+# EXAMPLE (13B MODEL)
+#
+#     # you can replace all the 'f16' with 'q4_0' if you're using quantized weights
+#     python migrate-ggml-2023-03-30-pr613.py models/13B/ggml-model-f16.bin models/13B/ggml-model-f16-ggjt.bin
+#
+#     # check that it works
+#     ./main -m models/13B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?'
+#
+#     # you can delete the old files
+#     rm -f models/13B/ggml-model-f16.bin*
+#     mv models/13B/ggml-model-f16-ggjt.bin models/13B/ggml-model-f16.bin
+#
+
+import argparse
+import os
+import sys
+import json
+import struct
+import numpy as np
+
+QK = 32
+
+GGML_TYPE_Q4_0  = 0
+GGML_TYPE_Q4_1  = 1
+GGML_TYPE_I8    = 2
+GGML_TYPE_I16   = 3
+GGML_TYPE_I32   = 4
+GGML_TYPE_F16   = 5
+GGML_TYPE_F32   = 6
+
+WTYPE_NAMES = {
+    0: "F32",
+    1: "F16",
+    2: "Q4_0",
+    3: "Q4_1",
+}
+
+WTYPES = {
+    0: GGML_TYPE_F32,
+    1: GGML_TYPE_F16,
+    2: GGML_TYPE_Q4_0,
+    3: GGML_TYPE_Q4_1,
+}
+
+GGML_BLCK_SIZE = {
+    GGML_TYPE_Q4_0:  QK,
+    GGML_TYPE_Q4_1:  QK,
+    GGML_TYPE_I8:    1,
+    GGML_TYPE_I16:   1,
+    GGML_TYPE_I32:   1,
+    GGML_TYPE_F16:   1,
+    GGML_TYPE_F32:   1,
+}
+
+GGML_TYPE_SIZE = {
+    GGML_TYPE_Q4_0: 4   + QK//2,
+    GGML_TYPE_Q4_1: 4*2 + QK//2,
+    GGML_TYPE_I8:   1,
+    GGML_TYPE_I16:  2,
+    GGML_TYPE_I32:  4,
+    GGML_TYPE_F16:  2,
+    GGML_TYPE_F32:  4,
+}
+
+HPARAMS = [
+    'magic',    # int32
+    'version',  # int32
+    'n_vocab',  # int32
+    'n_embd',   # int32
+    'n_mult',   # int32
+    'n_head',   # int32
+    'n_layer',  # int32
+    'n_rot',    # int32
+    'f16',      # int32
+]
+
+def read_hparams(fin):
+    struct_fmt = "i" * len(HPARAMS)
+    struct_size = struct.calcsize(struct_fmt)
+    buf = fin.read(struct_size)
+    ints = struct.unpack(struct_fmt, buf)
+    hparams = dict(zip(HPARAMS, ints))
+    return hparams
+
+def write_hparams(fout, hparams):
+    struct_fmt = "i" * len(HPARAMS)
+    struct_size = struct.calcsize(struct_fmt)
+    ints = [hparams[h] for h in HPARAMS]
+    fout.write(struct.pack(struct_fmt, *ints))
+
+def read_tokens(fin, hparams):
+    tokens = []
+    for i in range(hparams['n_vocab']):
+        len_b = fin.read(4)
+        (length,) = struct.unpack("i", len_b)
+        word = fin.read(length)
+        score_b = fin.read(4)
+        (score,) = struct.unpack("f", score_b)
+        tokens.append((word, score))
+    return tokens
+
+def write_tokens(fout, tokens):
+    for word, score in tokens:
+        fout.write(struct.pack("i", len(word)))
+        fout.write(word)
+        fout.write(struct.pack("f", score))
+
+def ggml_nelements(shape):
+    r = 1
+    for i in shape:
+        r *= i
+    return r
+
+def ggml_nbytes(shape, ftype):
+    x = ggml_nelements(shape)
+    t = WTYPES[ftype]
+    x *= GGML_TYPE_SIZE[t]
+    x //= GGML_BLCK_SIZE[t]
+    return x
+
+def copy_tensors(fin, fout, part_id, n_parts):
+    while True:
+
+        b = fin.read(4)
+        if not b: break
+        (n_dims,) = struct.unpack("i", b)
+        b = fin.read(4)
+        (length,) = struct.unpack("i", b)
+        b = fin.read(4)
+        (ftype,) = struct.unpack("i", b)
+
+        assert n_dims in (1, 2)
+
+        partshape = list(range(n_dims))
+        for i in range(n_dims):
+            b = fin.read(4)
+            partshape[i] = struct.unpack("i", b)[0]
+        partshape = list(reversed(partshape))
+
+        name = fin.read(length)
+        data = fin.read(ggml_nbytes(partshape, ftype))
+
+        blck_size = GGML_BLCK_SIZE[WTYPES[ftype]]
+        type_size = GGML_TYPE_SIZE[WTYPES[ftype]]
+
+        print(f"Processing tensor {name} with shape: {partshape} and type: {WTYPE_NAMES[ftype]}")
+
+        # determine dimension along which multipart tensor is sharded
+        #
+        # split_dim 0 regex:
+        #   - output.*
+        #   - layers.*.attention.wq.weight
+        #   - layers.*.attention.wk.weight
+        #   - layers.*.attention.wv.weight
+        #   - layers.*.feed_forward.w1.weight
+        #   - layers.*.feed_forward.w3.weight
+        #
+        # split_dim 1 regex:
+        #   - tok_embeddings.*
+        #   - layers.*.attention.wo.weight
+        #   - layers.*.feed_forward.w2.weight
+        #
+        if n_dims > 1:
+            split_dim = 1
+            if b"tok_embeddings" in name:
+                split_dim = 1
+            elif b"layers" in name:
+                if b"attention.wo.weight" in name:
+                    split_dim = 1
+                elif b"feed_forward.w2.weight" in name:
+                    split_dim = 1
+                else:
+                    split_dim = 0
+            elif b"output" in name:
+                split_dim = 0
+
+        # output tensor header
+        fullshape = list(partshape)
+        if n_dims > 1:
+            fullshape[split_dim] *= n_parts
+        fout.write(struct.pack("iii", n_dims, len(name), ftype))
+        for dim in reversed(fullshape):
+            fout.write(struct.pack("i", dim))
+        fout.write(name)
+
+        # ensure tensor data is aligned
+        tensor_data_offset = fout.tell()
+        while tensor_data_offset % QK != 0:
+            fout.write(struct.pack("B", 0))
+            tensor_data_offset += 1
+
+        # output unified mappable tensor data
+        if n_dims == 1 or n_parts == 1:
+            # copy tensor which we thankfully received in one piece
+            if part_id == 0:
+                fout.write(data)
+        elif split_dim == 0:
+            # reassemble multifile tensor containing some of the rows
+            rows_per_chunk = partshape[0]
+            current_row = part_id * rows_per_chunk
+            bytes_per_row = fullshape[1] // blck_size * type_size
+            offset = current_row * bytes_per_row
+            fout.seek(tensor_data_offset + offset)
+            fout.write(data)
+        elif split_dim == 1:
+            # reassemble multifile tensor containing some of the cols
+            cols_per_chunk = partshape[1]
+            current_col = part_id * cols_per_chunk
+            bpr = partshape[1] // blck_size * type_size
+            bytes_per_row = fullshape[1] // blck_size * type_size
+            offset_current_col = current_col // blck_size * type_size
+            for row in range(partshape[0]):
+                offset_row = row * bytes_per_row
+                offset = offset_row + offset_current_col
+                fout.seek(tensor_data_offset + offset)
+                fout.write(data[row * bpr:row * bpr + bpr])
+
+        # advance file position to next tensor
+        fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype))
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Migrate from GGML to new GGJT file format')
+    parser.add_argument('fin_path', help='your old ggml file (leave out the .1 .2 etc.)')
+    parser.add_argument('fout_path', help='your new ggjt file name')
+    return parser.parse_args()
+
+def main():
+    args = parse_args()
+    assert args.fin_path
+    assert args.fout_path
+    assert args.fin_path != args.fout_path
+
+    with open(args.fin_path, "rb") as fin:
+        hparams = read_hparams(fin)
+        tokens = read_tokens(fin, hparams)
+
+    if hparams['magic'] == 0x67676a74:  # ggjt
+        print(f"{args.fin_path}: input ggml has already been converted to 'ggjt' magic\n")
+        sys.exit(1)
+
+    if hparams['magic'] != 0x67676d66:  # ggmf
+        print(f"{args.fin_path}: input ggml file doesn't have expected 'ggmf' magic: {hparams['magic']:#x}\n")
+        sys.exit(1)
+
+    hparams['magic'] = 0x67676a74  # ggjt
+
+    # count number of multipart files by convention
+    n_parts = 1
+    while True:
+        if os.path.exists(f"{args.fin_path}.{n_parts}"):
+            n_parts += 1
+        else:
+            break
+
+    # we output a single file for ggml
+    with open(args.fout_path, "wb") as fout:
+        write_hparams(fout, hparams)
+        write_tokens(fout, tokens)
+        offset_of_tensors = fout.tell()
+        # the tensors we load could be split across multiple files
+        for part_id in range(n_parts):
+            fout.seek(offset_of_tensors)
+            print(f"Processing part {part_id+1} of {n_parts}\n")
+            fin_path = args.fin_path
+            if part_id > 0:
+                fin_path += f".{part_id}"
+            with open(fin_path, "rb") as fin:
+                read_tokens(fin, read_hparams(fin))
+                copy_tensors(fin, fout, part_id, n_parts)
+
+    print(f"Done. Output file: {args.fout_path}\n")
+
+if __name__ == "__main__":
+    main()
diff --git a/models/ggml-vocab.bin b/models/ggml-vocab.bin
index 3651f708e..38f63493a 100644
Binary files a/models/ggml-vocab.bin and b/models/ggml-vocab.bin differ
diff --git a/prompts/reason-act.txt b/prompts/reason-act.txt
new file mode 100644
index 000000000..872016631
--- /dev/null
+++ b/prompts/reason-act.txt
@@ -0,0 +1,18 @@
+You run in a loop of Thought, Action, Observation.
+At the end of the loop either Answer or restate your Thought and Action.
+Use Thought to describe your thoughts about the question you have been asked.
+Use Action to run one of these actions available to you:
+- calculate[python math expression]
+Observation will be the result of running those actions
+
+
+Question: What is 4 * 7 / 3?
+Thought: Do I need to use an action? Yes, I use calculate to do math
+Action: calculate[4 * 7 / 3]
+Observation: 9.3333333333
+Thought: Do I need to use an action? No, have the result
+Answer: The calculate tool says it is 9.3333333333
+Question: What is capital of france?
+Thought: Do I need to use an action? No, I know the answer
+Answer: Paris is the capital of France
+Question:
diff --git a/quantize.py b/quantize.py
deleted file mode 100644
index 16b5963d3..000000000
--- a/quantize.py
+++ /dev/null
@@ -1,127 +0,0 @@
-#!/usr/bin/env python3
-
-"""Script to execute the "quantize" script on a given set of models."""
-
-import subprocess
-import argparse
-import glob
-import sys
-import os
-
-
-def main():
-    """Update the quantize binary name depending on the platform and parse
-    the command line arguments and execute the script.
-    """
-
-    if "linux" in sys.platform or "darwin" in sys.platform:
-        quantize_script_binary = "quantize"
-
-    elif "win32" in sys.platform or "cygwin" in sys.platform:
-        quantize_script_binary = "quantize.exe"
-
-    else:
-        print("WARNING: Unknown platform. Assuming a UNIX-like OS.\n")
-        quantize_script_binary = "quantize"
-
-    parser = argparse.ArgumentParser(
-        prog='python3 quantize.py',
-        description='This script quantizes the given models by applying the '
-        f'"{quantize_script_binary}" script on them.'
-    )
-    parser.add_argument(
-        'models', nargs='+', choices=('7B', '13B', '30B', '65B'),
-        help='The models to quantize.'
-    )
-    parser.add_argument(
-        '-r', '--remove-16', action='store_true', dest='remove_f16',
-        help='Remove the f16 model after quantizing it.'
-    )
-    parser.add_argument(
-        '-m', '--models-path', dest='models_path',
-        default=os.path.join(os.getcwd(), "models"),
-        help='Specify the directory where the models are located.'
-    )
-    parser.add_argument(
-        '-q', '--quantize-script-path', dest='quantize_script_path',
-        default=os.path.join(os.getcwd(), quantize_script_binary),
-        help='Specify the path to the "quantize" script.'
-    )
-
-    # TODO: Revise this code
-    # parser.add_argument(
-    #     '-t', '--threads', dest='threads', type='int',
-    #     default=os.cpu_count(),
-    #     help='Specify the number of threads to use to quantize many models at '
-    #     'once. Defaults to os.cpu_count().'
-    # )
-
-    args = parser.parse_args()
-    args.models_path = os.path.abspath(args.models_path)
-
-    if not os.path.isfile(args.quantize_script_path):
-        print(
-            f'The "{quantize_script_binary}" script was not found in the '
-            "current location.\nIf you want to use it from another location, "
-            "set the --quantize-script-path argument from the command line."
-        )
-        sys.exit(1)
-
-    for model in args.models:
-        # The model is separated in various parts
-        # (ggml-model-f16.bin, ggml-model-f16.bin.0, ggml-model-f16.bin.1...)
-        f16_model_path_base = os.path.join(
-            args.models_path, model, "ggml-model-f16.bin"
-        )
-
-        f16_model_parts_paths = map(
-            lambda filename: os.path.join(f16_model_path_base, filename),
-            glob.glob(f"{f16_model_path_base}*")
-        )
-
-        for f16_model_part_path in f16_model_parts_paths:
-            if not os.path.isfile(f16_model_part_path):
-                print(
-                    f"The f16 model {os.path.basename(f16_model_part_path)} "
-                    f"was not found in {args.models_path}{os.path.sep}{model}"
-                    ". If you want to use it from another location, set the "
-                    "--models-path argument from the command line."
-                )
-                sys.exit(1)
-
-            __run_quantize_script(
-                args.quantize_script_path, f16_model_part_path
-            )
-
-            if args.remove_f16:
-                os.remove(f16_model_part_path)
-
-
-# This was extracted to a top-level function for parallelization, if
-# implemented. See https://github.com/ggerganov/llama.cpp/pull/222/commits/f8db3d6cd91bf1a1342db9d29e3092bc12dd783c#r1140496406
-
-def __run_quantize_script(script_path, f16_model_part_path):
-    """Run the quantize script specifying the path to it and the path to the
-    f16 model to quantize.
-    """
-
-    new_quantized_model_path = f16_model_part_path.replace("f16", "q4_0")
-    subprocess.run(
-        [script_path, f16_model_part_path, new_quantized_model_path, "2"],
-        check=True
-    )
-
-
-if __name__ == "__main__":
-    try:
-        main()
-
-    except subprocess.CalledProcessError:
-        print("\nAn error ocurred while trying to quantize the models.")
-        sys.exit(1)
-
-    except KeyboardInterrupt:
-        sys.exit(0)
-
-    else:
-        print("\nSuccesfully quantized all models.")
diff --git a/spm-headers/llama.h b/spm-headers/llama.h
new file mode 120000
index 000000000..9acceb980
--- /dev/null
+++ b/spm-headers/llama.h
@@ -0,0 +1 @@
+../llama.h
\ No newline at end of file
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index b44d7fe7e..157d7336e 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -5,5 +5,6 @@ function(llama_add_test source)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()
 
+# llama_add_test(test-double-float.c) # SLOW
 llama_add_test(test-quantize.c)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
diff --git a/tests/test-double-float.c b/tests/test-double-float.c
new file mode 100644
index 000000000..89dafc9f2
--- /dev/null
+++ b/tests/test-double-float.c
@@ -0,0 +1,53 @@
+// These tests may take a long time!
+// They are to prove that conversion from double to float of various functions in ggml.c doesn't affect the result.
+// This is done by checking all finite (non-NaN, non-infinite) floats.
+
+#undef NDEBUG
+#include <assert.h>
+#include <immintrin.h>
+#include <math.h>
+#include <stdint.h>
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdouble-promotion"
+
+// ggml.c::quantize_row_q4_0_reference
+inline static uint8_t round_orig(float v0) { return ((int8_t) (round(v0))) + 8; }
+
+// ggml.c::ggml_silu_f32
+inline static float silu_orig(float x) {
+    return x/(1.0 + exp(-x));
+}
+
+#pragma GCC diagnostic pop
+
+// ggml.c::quantize_row_q4_0_reference
+inline static uint8_t round_float(float v0) { return (int8_t)roundf(v0) + 8; }
+
+// ggml.c::ggml_silu_f32
+inline static float silu_float(float x) {
+    return x/(1.0f + expf(-x));
+}
+
+int main(void) {
+    uint32_t x = UINT32_MAX;
+    do {
+        float f = *(float *)&x;
+        assert(!isfinite(f) || (round_orig(f) == round_float(f)));
+    } while (x--);
+
+#ifdef __F16C__
+    // GELU and SILU implementations are used with a FP16 lookup table.
+    // The original and float-only results are not equal for all inputs after converting to FP16.
+    // GELU is an approximation anyway (tanh), not tested here.
+    // For SILU, verify that the results are at least the closest floating point numbers, if the FP16 values don't match.
+    for (x = 0; x <= UINT16_MAX; x++) {
+        float f = _cvtsh_ss(x);
+        const float so = silu_orig(f);
+        const float sf = silu_float(f);
+        assert(   (_cvtss_sh(so, 0) == _cvtss_sh(sf, 0))
+               || (nextafterf(so, sf) == sf)
+               || (nextafterf(sf, so) == so));
+    }
+#endif
+}
diff --git a/tests/test-quantize.c b/tests/test-quantize.c
index d59ecb8ab..993e9dcc3 100644
--- a/tests/test-quantize.c
+++ b/tests/test-quantize.c
@@ -13,7 +13,7 @@ int main(void) {
         src[i] = (float)(i + 1);
     }
 
-    size_t size = ggml_quantize_q4_0(src, dst, QK, QK, QK, hist);
+    size_t size = ggml_quantize_q4_0(src, dst, QK, QK, hist);
     assert(size == 20);
     float max_result = ((float *)dst)[0];
     float max_expected = src[31] / ((1 << 3) - 1);
@@ -24,7 +24,7 @@ int main(void) {
         assert(q4_result == q4_expected);
     }
 
-    size = ggml_quantize_q4_1(src, dst, QK, QK, QK, hist);
+    size = ggml_quantize_q4_1(src, dst, QK, QK, hist);
     assert(size == 24);
     float delta_result = ((float *)dst)[0];
     float delta_expected = (src[31] - src[0]) / ((1 << 4) - 1);
diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp
index 382055324..55b086dae 100644
--- a/tests/test-tokenizer-0.cpp
+++ b/tests/test-tokenizer-0.cpp
@@ -77,5 +77,7 @@ int main(int argc, char **argv) {
         }
     }
 
+    llama_free(ctx);
+
     return 0;
 }