release : v1.7.3

* ggml : fix cpy op for IQ-quants to use reference impl

ggml-ci

* ggml : disable tests involving i-matrix quantization

* ggml : update ggml_backend_cpu_device_supports_op

ggml-ci

.devops/cublas.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libsdl2-dev
+    apt-get install -y build-essential git cmake libsdl2-dev wget
 
 WORKDIR /app
 
@@ -23,6 +23,6 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
 ENV GGML_CUDA=1
 
-RUN make
+RUN make base.en
 
 ENTRYPOINT ["/app/main"]

.devops/main-cuda.Dockerfile

@@ -17,7 +17,7 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 ENV GGML_CUDA=1
 
 RUN apt-get update && \
-    apt-get install -y build-essential libsdl2-dev \
+    apt-get install -y build-essential libsdl2-dev wget cmake \
     && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 # Ref: https://stackoverflow.com/a/53464012
@@ -25,7 +25,7 @@ ENV CUDA_MAIN_VERSION=12.3
 ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 
 COPY .. .
-RUN make
+RUN make base.en
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 ENV CUDA_MAIN_VERSION=12.3
@@ -33,7 +33,7 @@ ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
+  apt-get install -y curl ffmpeg wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app

.devops/main.Dockerfile

@@ -2,17 +2,17 @@ FROM ubuntu:22.04 AS build
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y build-essential \
+  apt-get install -y build-essential wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY .. .
-RUN make
+RUN make base.en
 
 FROM ubuntu:22.04 AS runtime
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg libsdl2-dev \
+  apt-get install -y curl ffmpeg libsdl2-dev wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app

.github/workflows/bindings-ruby.yml

@@ -3,61 +3,41 @@ on:
   push:
     paths:
       - bindings/ruby/**
-      - src/whisper.cpp
-      - include/whisper.h
-      - ggml/src/ggml.c
-      - ggml/src/ggml-impl.h
-      - ggml/src/ggml-aarch64.h
-      - ggml/src/ggml-aarch64.c
-      - ggml/src/ggml-alloc.c
-      - ggml/src/ggml-backend-impl.h
-      - ggml/src/ggml-backend.cpp
-      - ggml/src/ggml-common.h
-      - ggml/src/ggml-quants.h
-      - ggml/src/ggml-quants.c
-      - ggml/src/ggml-cpu-impl.h
-      - ggml/src/ggml-metal.m
-      - ggml/src/ggml-metal.metal
-      - ggml/src/ggml-blas.cpp
-      - ggml/include/ggml.h
-      - ggml/include/ggml-alloc.h
-      - ggml/include/ggml-backend.h
-      - ggml/include/ggml-cuda.h
-      - ggml/include/ggml-kompute.h
-      - ggml/include/ggml-metal.h
-      - ggml/include/ggml-sycl.h
-      - ggml/include/ggml-vulkan.h
-      - ggml/include/ggml-blas.h
+      - src/**/*.c
+      - src/**/*.cpp
+      - src/**/*.h
+      - src/**/*.m
+      - src/**/*.metal
+      - include/**/*.c
+      - include/**/*.cpp
+      - include/**/*.h
+      - include/**/*.m
+      - include/**/*.metal
+      - ggml/**/*.c
+      - ggml/**/*.cpp
+      - ggml/**/*.h
+      - ggml/**/*.m
+      - ggml/**/*.metal
       - scripts/get-flags.mk
       - examples/dr_wav.h
   pull_request:
     paths:
       - bindings/ruby/**
-      - src/whisper.cpp
-      - include/whisper.h
-      - ggml/src/ggml.c
-      - ggml/src/ggml-impl.h
-      - ggml/src/ggml-aarch64.h
-      - ggml/src/ggml-aarch64.c
-      - ggml/src/ggml-alloc.c
-      - ggml/src/ggml-backend-impl.h
-      - ggml/src/ggml-backend.cpp
-      - ggml/src/ggml-common.h
-      - ggml/src/ggml-quants.h
-      - ggml/src/ggml-quants.c
-      - ggml/src/ggml-cpu-impl.h
-      - ggml/src/ggml-metal.m
-      - ggml/src/ggml-metal.metal
-      - ggml/src/ggml-blas.cpp
-      - ggml/include/ggml.h
-      - ggml/include/ggml-alloc.h
-      - ggml/include/ggml-backend.h
-      - ggml/include/ggml-cuda.h
-      - ggml/include/ggml-kompute.h
-      - ggml/include/ggml-metal.h
-      - ggml/include/ggml-sycl.h
-      - ggml/include/ggml-vulkan.h
-      - ggml/include/ggml-blas.h
+      - src/**/*.c
+      - src/**/*.cpp
+      - src/**/*.h
+      - src/**/*.m
+      - src/**/*.metal
+      - include/**/*.c
+      - include/**/*.cpp
+      - include/**/*.h
+      - include/**/*.m
+      - include/**/*.metal
+      - ggml/**/*.c
+      - ggml/**/*.cpp
+      - ggml/**/*.h
+      - ggml/**/*.m
+      - ggml/**/*.metal
       - scripts/get-flags.mk
       - examples/dr_wav.h
 
@@ -70,6 +50,6 @@ jobs:
     steps:
       - uses: ruby/setup-ruby@v1
         with:
-          ruby-version: '3.0'
+          ruby-version: '3.1'
       - uses: actions/checkout@v4
       - run: rake test

.github/workflows/build.yml

@@ -3,6 +3,7 @@ on: [push, pull_request]
 
 env:
   ubuntu_image: "ubuntu:22.04"
+  VCPKG_BINARY_SOURCES: "clear;x-gha,readwrite"
 
 jobs:
   ubuntu-latest:
@@ -27,9 +28,9 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential libsdl2-dev
-            make
-            make stream'
+            apt install -y build-essential libsdl2-dev cmake
+            cmake -B build
+            cmake --build build --config Release -j $(nproc)'
 
   macOS-latest:
     runs-on: macOS-latest
@@ -41,30 +42,30 @@ jobs:
       - name: Dependencies
         run: |
           brew update
-          brew install sdl2
+          brew install sdl2 cmake
 
       - name: Build
         run: |
-          make
-          make stream
+          cmake -B build
+          cmake --build build --config Release
 
-  freeBSD-latest:
-    runs-on: macos-12
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Build
-        uses: cross-platform-actions/action@v0.24.0
-        with:
-          operating_system: freebsd
-          version: '13.3'
-          run: |
-            sudo pkg update
-            sudo pkg install -y gmake sdl2
-            gmake
-            gmake stream
+#  freeBSD-latest:
+#    runs-on: macos-12
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Build
+#        uses: cross-platform-actions/action@v0.24.0
+#        with:
+#          operating_system: freebsd
+#          version: '13.3'
+#          run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake sdl2 cmake
+#            cmake -B build
+#            cmake --build build --config Release
 
   ubuntu-latest-gcc:
     runs-on: ubuntu-latest
@@ -279,25 +280,10 @@ jobs:
             mingw-w64-${{matrix.env}}-SDL2
             mingw-w64-${{matrix.env}}-openblas
 
-      - name: Build using make
-        shell: msys2 {0}
-        run: |
-            make -j $(nproc)
-
-      - name: Clean after building using make
-        shell: msys2 {0}
-        run: |
-            make clean
-
-      - name: Build using make w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            make GGML_OPENBLAS=1 -j $(nproc)
-
       - name: Build using CMake
         shell: msys2 {0}
         run: |
-            cmake -B build
+            cmake -B build -DWHISPER_SDL2=ON
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
       - name: Clean after building using CMake
@@ -308,7 +294,7 @@ jobs:
       - name: Build using CMake w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            cmake -B build -DGGML_OPENBLAS=ON
+            cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
@@ -382,10 +368,8 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
           - arch: x64
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
           - sdl2: ON
             s2ver: 2.28.5
@@ -394,17 +378,21 @@ jobs:
       - name: Clone
         uses: actions/checkout@v4
 
+      - name: Export GitHub Actions cache environment variables
+        uses: actions/github-script@v7
+        with:
+          script: |
+            core.exportVariable('ACTIONS_CACHE_URL', process.env.ACTIONS_CACHE_URL || '');
+            core.exportVariable('ACTIONS_RUNTIME_TOKEN', process.env.ACTIONS_RUNTIME_TOKEN || '');
+
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v2
 
-      - name: Fetch OpenBLAS
+      - name: Install OpenBLAS and pkgconfiglite
         if: matrix.blas == 'ON'
         run: |
-          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
-          7z x blas.zip -oblas -y
-          copy blas/include/cblas.h .
-          copy blas/include/openblas_config.h .
-          echo "OPENBLAS_PATH=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+          vcpkg install --triplet=${{ matrix.s2arc }}-windows openblas
+          choco install pkgconfiglite
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -416,9 +404,10 @@ jobs:
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_TOOLCHAIN_FILE="$env:VCPKG_INSTALLATION_ROOT/scripts/buildsystems/vcpkg.cmake"
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
+          -DGGML_BLAS=${{ matrix.blas }}
+          -DGGML_BLAS_VENDOR=OpenBLAS
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
@@ -426,9 +415,9 @@ jobs:
           cd ./build
           msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 
-      - name: Copy libopenblas.dll
+      - name: Copy openblas.dll
         if: matrix.blas == 'ON'
-        run: copy "$env:OPENBLAS_PATH/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+        run: copy "C:/vcpkg/packages/openblas_${{ matrix.s2arc }}-windows/bin/openblas.dll" build/bin/${{ matrix.build }}
 
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
@@ -441,71 +430,72 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
-  windows-cublas:
-    runs-on: windows-2019
-
-    strategy:
-      matrix:
-        build: [Release]
-        arch: [x64]
-        cublas: [ON]
-        sdl2: [ON]
-        cuda-toolkit: [12.2.0, 11.8.0]
-        include:
-          - arch: x64
-            s2arc: x64
-          - sdl2: ON
-            s2ver: 2.28.5
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
-
-      - name: Install CUDA Toolkit
-        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
-        with:
-          cuda: '${{ matrix.cuda-toolkit }}'
-
-      - name: Fetch SDL2 and set SDL2_DIR
-        if: matrix.sdl2 == 'ON'
-        run: |
-          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
-          7z x sdl2.zip
-          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
-
-      - name: Configure
-        run: >
-          cmake -S . -B ./build -A ${{ matrix.arch }}
-          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_CUDA=${{ matrix.cublas }}
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
-
-      - name: Build ${{ matrix.cuda-toolkit }}
-        run: |
-          cd ./build
-          cmake --build . --config ${{ matrix.build }}
-
-      - name: Copy CUDA DLLs
-        run: >
-          Copy-Item -PassThru
-          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
-          -Include cudart64_*,cublas64_*,cublasLt64_*
-          -Destination build/bin/${{ matrix.build }}
-
-      - name: Copy SDL2.dll
-        if: matrix.sdl2 == 'ON'
-        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
-
-      - name: Upload binaries
-        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
-        with:
-          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
-          path: build/bin/${{ matrix.build }}
+# TODO: fix and re-enable
+#  windows-cublas:
+#    runs-on: windows-2019
+#
+#    strategy:
+#      matrix:
+#        build: [Release]
+#        arch: [x64]
+#        cublas: [ON]
+#        sdl2: [ON]
+#        cuda-toolkit: [12.2.0, 11.8.0]
+#        include:
+#          - arch: x64
+#            s2arc: x64
+#          - sdl2: ON
+#            s2ver: 2.28.5
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Add msbuild to PATH
+#        uses: microsoft/setup-msbuild@v2
+#
+#      - name: Install CUDA Toolkit
+#        id: cuda-toolkit
+#        uses: Jimver/cuda-toolkit@v0.2.15
+#        with:
+#          cuda: '${{ matrix.cuda-toolkit }}'
+#
+#      - name: Fetch SDL2 and set SDL2_DIR
+#        if: matrix.sdl2 == 'ON'
+#        run: |
+#          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
+#          7z x sdl2.zip
+#          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
+#
+#      - name: Configure
+#        run: >
+#          cmake -S . -B ./build -A ${{ matrix.arch }}
+#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#          -DGGML_CUDA=${{ matrix.cublas }}
+#          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+#
+#      - name: Build ${{ matrix.cuda-toolkit }}
+#        run: |
+#          cd ./build
+#          cmake --build . --config ${{ matrix.build }}
+#
+#      - name: Copy CUDA DLLs
+#        run: >
+#          Copy-Item -PassThru
+#          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
+#          -Include cudart64_*,cublas64_*,cublasLt64_*
+#          -Destination build/bin/${{ matrix.build }}
+#
+#      - name: Copy SDL2.dll
+#        if: matrix.sdl2 == 'ON'
+#        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+#
+#      - name: Upload binaries
+#        if: matrix.sdl2 == 'ON'
+#        uses: actions/upload-artifact@v4
+#        with:
+#          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
+#          path: build/bin/${{ matrix.build }}
 
   emscripten:
     runs-on: ubuntu-latest
@@ -529,7 +519,7 @@ jobs:
           emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           make
 
-  ios:
+  ios-xcode-build:
     runs-on: macos-latest
 
     strategy:
@@ -537,7 +527,7 @@ jobs:
         build: [Release]
 
     steps:
-      - name: Clone
+      - name: Checkout code
         uses: actions/checkout@v4
 
       - name: Configure
@@ -545,11 +535,34 @@ jobs:
           cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
           mkdir models/ggml-base.en-encoder.mlmodelc
 
-      - name: Build objc example
-        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DWHISPER_BUILD_EXAMPLES=OFF \
+            -DWHISPER_BUILD_TESTS=OFF \
+            -DWHISPER_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+          sudo cmake --install . --config Release
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme whisper-Package -destination 'generic/platform=iOS'
+
+#- name: Build objc example
+#  run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphoneos build
 
       - name: Build swiftui example
-        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
+        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
   android:
     runs-on: ubuntu-latest
@@ -560,12 +573,6 @@ jobs:
         with:
           path: whisper
 
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          repository: ggerganov/ggml
-          path: ggml
-
       - name: Install Java
         uses: actions/setup-java@v4
         with:
@@ -584,7 +591,7 @@ jobs:
         run: |
           export PATH_TO_GGML=$PWD/ggml
           cd whisper/examples/whisper.android
-          ./gradlew assembleRelease --no-daemon -PGGML_HOME=$PATH_TO_GGML
+          ./gradlew assembleRelease --no-daemon
 
 # TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602
 #  android_java:
@@ -666,5 +673,6 @@ jobs:
       - name: Test quantize
         run: |
           ./models/download-ggml-model.sh tiny.en
-          make quantize
-          ./quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0
+          cmake -B build
+          cmake --build build --config Release
+          ./build/bin/quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0

.github/workflows/docker.yml

@@ -45,7 +45,7 @@ jobs:
         with:
           context: .
           push: true
-          platforms: ${{ matrix.config.platforms }}
+          platforms: ${{ matrix.config.platform }}
           tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
@@ -54,6 +54,6 @@ jobs:
         with:
           context: .
           push: ${{ github.event_name == 'push' }}
-          platforms: ${{ matrix.config.platforms }}
+          platforms: ${{ matrix.config.platform }}
           tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}

.gitignore

@@ -1,5 +1,6 @@
 *.o
 *.a
+*.d
 .cache/
 .coreml/
 .test/
@@ -19,6 +20,9 @@ build-*/
 .swiftpm
 *.metallib
 
+ggml-metal-embed.metal
+ggml-metal-embed.metal.tmp
+
 /main
 /stream
 /command

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
 project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.7.1)
+project("whisper.cpp" VERSION 1.7.3)
 include(CheckIncludeFileCXX)
 
 set(SOVERSION 1)

Package.swift

@@ -14,47 +14,6 @@ let package = Package(
         .library(name: "whisper", targets: ["whisper"]),
     ],
     targets: [
-        .target(
-            name: "whisper",
-            path: ".",
-            exclude: [
-               "bindings",
-               "cmake",
-               "coreml",
-               "examples",
-               "extra",
-               "models",
-               "samples",
-               "tests",
-               "CMakeLists.txt",
-               "Makefile"
-            ],
-            sources: [
-                "ggml/src/ggml.c",
-                "src/whisper.cpp",
-                "ggml/src/ggml-aarch64.c",
-                "ggml/src/ggml-alloc.c",
-                "ggml/src/ggml-backend.cpp",
-                "ggml/src/ggml-quants.c",
-                "ggml/src/ggml-metal.m"
-            ],
-            resources: [.process("ggml-metal.metal")],
-            publicHeadersPath: "spm-headers",
-            cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL")
-                // NOTE: NEW_LAPACK will required iOS version 16.4+
-                // We should consider add this in the future when we drop support for iOS 14
-                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-                // .define("ACCELERATE_NEW_LAPACK"),
-                // .define("ACCELERATE_LAPACK_ILP64")
-            ],
-            linkerSettings: [
-                .linkedFramework("Accelerate")
-            ]
-        )
-    ],
-    cxxLanguageStandard: .cxx11
+        .systemLibrary(name: "whisper", pkgConfig: "whisper"),
+    ]
 )

README.md

@@ -7,7 +7,7 @@
 [![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.7.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.7.3](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.3) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -16,7 +16,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
-- [4-bit and 5-bit integer quantization support](#quantization)
+- [Integer quantization support](#quantization)
 - Zero memory allocations at runtime
 - [Vulkan support](#vulkan-gpu-support)
 - Support for CPU-only inference
@@ -89,10 +89,11 @@ Now build the [main](examples/main) example and transcribe an audio file like th
 
 ```bash
 # build the main example
-make -j
+cmake -B build
+cmake --build build --config Release
 
 # transcribe an audio file
-./main -f samples/jfk.wav
+./build/bin/main -f samples/jfk.wav
 ```
 
 ---
@@ -265,11 +266,12 @@ Here are the steps for creating and using a quantized model:
 
 ```bash
 # quantize a model with Q5_0 method
-make -j quantize
-./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
+cmake -B build
+cmake --build build --config Release
+./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
 
 # run the examples as usual, specifying the quantized model file
-./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
+./build/bin/main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
 ```
 
 ## Core ML support
@@ -303,10 +305,6 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 - Build `whisper.cpp` with Core ML support:
 
   ```bash
-  # using Makefile
-  make clean
-  WHISPER_COREML=1 make -j
-
   # using CMake
   cmake -B build -DWHISPER_COREML=1
   cmake --build build -j --config Release
@@ -426,8 +424,8 @@ First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-do
 Now build `whisper.cpp` with CUDA support:
 
 ```
-make clean
-GGML_CUDA=1 make -j
+cmake -B build -DGGML_CUDA=1
+cmake --build build -j --config Release
 ```
 
 ## Vulkan GPU support
@@ -436,8 +434,8 @@ First, make sure your graphics card driver provides support for Vulkan API.
 
 Now build `whisper.cpp` with Vulkan support:
 ```
-make clean
-make GGML_VULKAN=1 -j
+cmake -B build -DGGML_VULKAN=1
+cmake --build build -j --config Release
 ```
 
 ## BLAS CPU support via OpenBLAS
@@ -448,28 +446,13 @@ First, make sure you have installed `openblas`: https://www.openblas.net/
 Now build `whisper.cpp` with OpenBLAS support:
 
 ```
-make clean
-GGML_OPENBLAS=1 make -j
-```
-
-## BLAS CPU support via Intel MKL
-
-Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
-First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
-
-Now build `whisper.cpp` with Intel MKL BLAS support:
-
-```
-source /opt/intel/oneapi/setvars.sh
-mkdir build
-cd build
-cmake -DWHISPER_MKL=ON ..
-WHISPER_MKL=1 make -j
+cmake -B build -DGGML_BLAS=1
+cmake --build build -j --config Release
 ```
 
 ## Ascend NPU support
 
-Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores. 
+Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores.
 
 First, check if your Ascend NPU device is supported:
 
@@ -483,10 +466,8 @@ Then, make sure you have installed [`CANN toolkit`](https://www.hiascend.com/en/
 Now build `whisper.cpp` with CANN support:
 
 ```
-mkdir build
-cd build
-cmake .. -D GGML_CANN=on
-make -j
+cmake -B build -DGGML_CANN=1
+cmake --build build -j --config Release
 ```
 
 Run the inference examples as usual, for example:
@@ -636,8 +617,9 @@ The [stream](examples/stream) tool samples the audio every half a second and run
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-make stream -j
-./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+cmake -B build
+cmake --build build --config Release
+./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4

Sources/whisper/module.modulemap

@@ -0,0 +1,5 @@
+module whisper [system] {
+    header "whisper.h"
+    link "whisper"
+    export *
+}

Sources/whisper/whisper.h

@@ -0,0 +1,4 @@
+#pragma once
+
+#include <whisper.h>
+

bindings/java/README.md

@@ -67,5 +67,5 @@ copy /y ..\..\build\bin\Release\whisper.dll build\generated\resources\main\win32
 
 ## License
 
-The license for the Go bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
+The license for the Java bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
 

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.7.1",
+  "version": "1.7.3",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/.gitignore

@@ -1,3 +1,5 @@
 LICENSE
 pkg/
-lib/whisper.*
+lib/whisper.so
+lib/whisper.bundle
+lib/whisper.dll

bindings/ruby/README.md

@@ -22,7 +22,7 @@ Usage
 ```ruby
 require "whisper"
 
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 
 params = Whisper::Params.new
 params.language = "en"
@@ -31,7 +31,7 @@ params.duration = 60_000
 params.max_text_tokens = 300
 params.translate = true
 params.print_timestamps = false
-params.prompt = "Initial prompt here."
+params.initial_prompt = "Initial prompt here."
 
 whisper.transcribe("path/to/audio.wav", params) do |whole_text|
   puts whole_text
@@ -41,21 +41,66 @@ end
 
 ### Preparing model ###
 
-Use script to download model file(s):
+Some models are prepared up-front:
 
-```bash
-git clone https://github.com/ggerganov/whisper.cpp.git
-cd whisper.cpp
-sh ./models/download-ggml-model.sh base.en
+```ruby
+base_en = Whisper::Model.pre_converted_models["base.en"]
+whisper = Whisper::Context.new(base_en)
 ```
 
-There are some types of models. See [models][] page for details.
+At first time you use a model, it is downloaded automatically. After that, downloaded cached file is used. To clear cache, call `#clear_cache`:
+
+```ruby
+Whisper::Model.pre_converted_models["base"].clear_cache
+```
+
+You also can use shorthand for pre-converted models:
+
+```ruby
+whisper = Whisper::Context.new("base.en")
+```
+
+You can see the list of prepared model names by `Whisper::Model.preconverted_models.keys`:
+
+```ruby
+puts Whisper::Model.preconverted_model_names
+# tiny
+# tiny.en
+# tiny-q5_1
+# tiny.en-q5_1
+# tiny-q8_0
+# base
+# base.en
+# base-q5_1
+# base.en-q5_1
+# base-q8_0
+#   :
+#   :
+```
+
+You can also use local model files you prepared:
+
+```ruby
+whisper = Whisper::Context.new("path/to/your/model.bin")
+```
+
+Or, you can download model files:
+
+```ruby
+model_uri = Whisper::Model::URI.new("http://example.net/uri/of/your/model.bin")
+whisper = Whisper::Context.new(model_uri)
+```
+
+See [models][] page for details.
 
 ### Preparing audio file ###
 
 Currently, whisper.cpp accepts only 16-bit WAV files.
 
-### API ###
+API
+---
+
+### Segments ###
 
 Once `Whisper::Context#transcribe` called, you can retrieve segments by `#each_segment`:
 
@@ -85,13 +130,6 @@ end
 You can also add hook to params called on new segment:
 
 ```ruby
-def format_time(time_ms)
-  sec, decimal_part = time_ms.divmod(1000)
-  min, sec = sec.divmod(60)
-  hour, min = min.divmod(60)
-  "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
-end
-
 # Add hook before calling #transcribe
 params.on_new_segment do |segment|
   line = "[%{st} --> %{ed}] %{text}" % {
@@ -107,5 +145,87 @@ whisper.transcribe("path/to/audio.wav", params)
 
 ```
 
+### Models ###
+
+You can see model information:
+
+```ruby
+whisper = Whisper::Context.new("base")
+model = whisper.model
+
+model.n_vocab # => 51864
+model.n_audio_ctx # => 1500
+model.n_audio_state # => 512
+model.n_audio_head # => 8
+model.n_audio_layer # => 6
+model.n_text_ctx # => 448
+model.n_text_state # => 512
+model.n_text_head # => 8
+model.n_text_layer # => 6
+model.n_mels # => 80
+model.ftype # => 1
+model.type # => "base"
+
+```
+
+### Logging ###
+
+You can set log callback:
+
+```ruby
+prefix = "[MyApp] "
+log_callback = ->(level, buffer, user_data) {
+  case level
+  when Whisper::LOG_LEVEL_NONE
+    puts "#{user_data}none: #{buffer}"
+  when Whisper::LOG_LEVEL_INFO
+    puts "#{user_data}info: #{buffer}"
+  when Whisper::LOG_LEVEL_WARN
+    puts "#{user_data}warn: #{buffer}"
+  when Whisper::LOG_LEVEL_ERROR
+    puts "#{user_data}error: #{buffer}"
+  when Whisper::LOG_LEVEL_DEBUG
+    puts "#{user_data}debug: #{buffer}"
+  when Whisper::LOG_LEVEL_CONT
+    puts "#{user_data}same to previous: #{buffer}"
+  end
+}
+Whisper.log_set log_callback, prefix
+```
+
+Using this feature, you are also able to suppress log:
+
+```ruby
+Whisper.log_set ->(level, buffer, user_data) {
+  # do nothing
+}, nil
+Whisper::Context.new("base")
+```
+
+### Low-level API to transcribe ###
+
+You can also call `Whisper::Context#full` and `#full_parallel` with a Ruby array as samples. Although `#transcribe` with audio file path is recommended because it extracts PCM samples in C++ and is fast, `#full` and `#full_parallel` give you flexibility.
+
+```ruby
+require "whisper"
+require "wavefile"
+
+reader = WaveFile::Reader.new("path/to/audio.wav", WaveFile::Format.new(:mono, :float, 16000))
+samples = reader.enum_for(:each_buffer).map(&:samples).flatten
+
+whisper = Whisper::Context.new("base")
+whisper.full(Whisper::Params.new, samples)
+whisper.each_segment do |segment|
+  puts segment.text
+end
+```
+
+The second argument `samples` may be an array, an object with `length` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
+
+License
+-------
+
+The same to [whisper.cpp][].
+
 [whisper.cpp]: https://github.com/ggerganov/whisper.cpp
 [models]: https://github.com/ggerganov/whisper.cpp/tree/master/models

bindings/ruby/Rakefile

@@ -1,59 +1,64 @@
 require 'rake/clean'
 require "bundler/gem_tasks"
-require "pathname"
-require "yaml"
 require "rake/testtask"
+require_relative "extsources"
 
-extsources = YAML.load_file("extsources.yaml")
 SOURCES = FileList[]
-extsources.each do |src|
+
+EXTSOURCES.each do |src|
   basename = src.pathmap("%f")
-  dest = basename == "LICENSE" ? basename : basename.pathmap("ext/%f")
+  dest = basename == "LICENSE" ? basename : src.pathmap("%{../..,ext}p")
+  dir = dest.pathmap("%d")
   file src
-  file dest => src do |t|
+  directory dir
+  file dest => [src, dir] do |t|
     cp t.source, t.name
   end
   SOURCES.include dest
 end
-CLEAN.include SOURCES
-CLEAN.include FileList[
-                "ext/*.o",
-                "ext/*.metal",
-                "ext/whisper.{so,bundle,dll}",
-                "ext/depend"
-              ]
 
-task build: SOURCES + FileList[
-                        "ext/extconf.rb",
-                        "ext/ruby_whisper.h",
-                        "ext/ruby_whisper.cpp",
-                        "whispercpp.gemspec",
-                      ]
+CLEAN.include SOURCES
+CLEAN.include FileList["ext/*.o", "ext/*.metal", "ext/whisper.{so,bundle,dll}"]
+
+task build: ["ext/Makefile", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp", "whispercpp.gemspec"]
 
 directory "pkg"
 CLOBBER.include "pkg"
 
-TEST_MODEL = "../../models/ggml-base.en.bin"
 LIB_NAME = "whisper".ext(RbConfig::CONFIG["DLEXT"])
+SO_FILE = File.join("ext", LIB_NAME)
 LIB_FILE = File.join("lib", LIB_NAME)
 
-directory "lib"
-task LIB_FILE => SOURCES + ["lib"] do |t|
+file "ext/Makefile" => ["ext/extconf.rb", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp"] + SOURCES do |t|
+  Dir.chdir "ext" do
+    ruby "extconf.rb"
+  end
+end
+
+file SO_FILE => "ext/Makefile" do |t|
   Dir.chdir "ext" do
-    sh "ruby extconf.rb"
     sh "make"
   end
-  mv "ext/#{LIB_NAME}", t.name
+end
+CLEAN.include SO_FILE
+
+directory "lib"
+file LIB_FILE => [SO_FILE, "lib"] do |t|
+  copy t.source, t.name
 end
 CLEAN.include LIB_FILE
 
 Rake::TestTask.new do |t|
   t.test_files = FileList["tests/test_*.rb"]
 end
-task test: [TEST_MODEL, LIB_FILE]
 
-file TEST_MODEL do
-  Dir.chdir "../.." do
-    sh "./models/download-ggml-model.sh base.en"
+TEST_MEMORY_VIEW = "tests/jfk_reader/jfk_reader.#{RbConfig::CONFIG['DLEXT']}"
+file TEST_MEMORY_VIEW => "tests/jfk_reader/jfk_reader.c" do |t|
+  Dir.chdir "tests/jfk_reader" do
+    ruby "extconf.rb"
+    sh "make"
   end
 end
+CLEAN.include "tests/jfk_reader/jfk_reader.{o,#{RbConfig::CONFIG['DLEXT']}}"
+
+task test: [LIB_FILE, TEST_MEMORY_VIEW]

bindings/ruby/ext/.gitignore

@@ -1,35 +1,13 @@
 Makefile
-ggml.c
-ggml.h
-ggml-alloc.c
-ggml-alloc.h
-ggml-aarch64.c
-ggml-aarch64.h
-ggml-backend.cpp
-ggml-backend-impl.h
-ggml-backend.c
-ggml-backend.h
-ggml-common.h
-ggml-cpu-impl.h
-ggml-metal.m
-ggml-metal.metal
-ggml-metal-embed.metal
-ggml-blas.cpp
-ggml-cuda.h
-ggml-impl.h
-ggml-kompute.h
-ggml-metal.h
-ggml-opencl.h
-ggml-quants.c
-ggml-quants.h
-ggml-sycl.h
-ggml-vulkan.h
-ggml-blas.h
-get-flags.mk
-whisper.cpp
-whisper.h
-dr_wav.h
-depend
-whisper.bundle
 whisper.so
+whisper.bundle
 whisper.dll
+scripts/get-flags.mk
+*.o
+*.c
+*.cpp
+*.h
+*.m
+*.metal
+!ruby_whisper.cpp
+!ruby_whisper.h

bindings/ruby/ext/cpu.mk

@@ -0,0 +1,9 @@
+ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
+	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/include/ggml-backend.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-cpu.h \
+	ggml/src/ggml-impl.h
+	$(CXX) $(CXXFLAGS)   -c $< -o $@

bindings/ruby/ext/extconf.rb

@@ -1,7 +1,10 @@
 require 'mkmf'
 
 # need to use c++ compiler flags
-$CXXFLAGS << ' -std=c++11'
+$CXXFLAGS << ' -std=c++17'
+
+$LDFLAGS << ' -lstdc++'
+
 # Set to true when building binary gems
 if enable_config('static-stdlib', false)
   $LDFLAGS << ' -static-libgcc -static-libstdc++'
@@ -12,34 +15,6 @@ if enable_config('march-tune-native', false)
   $CXXFLAGS << ' -march=native -mtune=native'
 end
 
-def with_disabling_unsupported_files
-  disabled_files = []
-
-  unless $GGML_METAL
-    disabled_files << 'ggml-metal.h' << 'ggml-metal.m'
-  end
-
-  unless $GGML_METAL_EMBED_LIBRARY
-    disabled_files << 'ggml-metal.metal'
-  end
-
-  unless $OBJ_ALL&.include? 'ggml-blas.o'
-    disabled_files << 'ggml-blas.h' << 'ggml-blas.cpp'
-  end
-
-  disabled_files.filter! {|file| File.exist? file}
-
-  disabled_files.each do |file|
-    File.rename file, "#{file}.disabled"
-  end
-
-  yield
-
-  disabled_files.each do |file|
-    File.rename "#{file}.disabled", file
-  end
-end
-
 if ENV['WHISPER_METAL']
   $GGML_METAL ||= true
   $DEPRECATE_WARNING ||= true
@@ -60,16 +35,16 @@ if $GGML_METAL
   $GGML_METAL_EMBED_LIBRARY = true
 end
 
-$MK_CPPFLAGS = ''
+$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples'
 $MK_CFLAGS   = '-std=c11   -fPIC'
-$MK_CXXFLAGS = '-std=c++11 -fPIC'
-$MK_NVCCFLAGS = '-std=c++11'
+$MK_CXXFLAGS = '-std=c++17 -fPIC'
+$MK_NVCCFLAGS = '-std=c++17'
 $MK_LDFLAGS = ''
 
-$OBJ_GGML = ''
-$OBJ_WHISPER = ''
-$OBJ_COMMON = ''
-$OBJ_SDL = ''
+$OBJ_GGML = []
+$OBJ_WHISPER = []
+$OBJ_COMMON = []
+$OBJ_SDL = []
 
 $MK_CPPFLAGS << ' -D_XOPEN_SOURCE=600'
 
@@ -136,11 +111,6 @@ unless ENV['RISCV']
     $MK_CFLAGS     << ' -march=native -mtune=native'
     $HOST_CXXFLAGS << ' -march=native -mtune=native'
   end
-
-  if $UNAME_M.match? /aarch64.*/
-    $MK_CFLAGS   << ' -mcpu=native'
-    $MK_CXXFLAGS << ' -mcpu=native'
-  end
 else
   $MK_CFLAGS   << ' -march=rv64gcv -mabi=lp64d'
   $MK_CXXFLAGS << ' -march=rv64gcv -mabi=lp64d'
@@ -148,11 +118,11 @@ end
 
 unless ENV['GGML_NO_ACCELERATE']
   if $UNAME_S == 'Darwin'
-    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS'
+    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE'
     $MK_CPPFLAGS << ' -DACCELERATE_NEW_LAPACK'
     $MK_CPPFLAGS << ' -DACCELERATE_LAPACK_ILP64'
     $MK_LDFLAGS  << ' -framework Accelerate'
-    $OBJ_GGML    << ' ggml-blas.o'
+    $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
   end
 end
 
@@ -160,20 +130,20 @@ if ENV['GGML_OPENBLAS']
   $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas`.chomp}"
   $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas)`.chomp}"
   $MK_LDFLAGS  << " #{`pkg-config --libs openblas`}"
-  $OBJ_GGML    << ' ggml-blas.o'
+  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
 end
 
 if ENV['GGML_OPENBLAS64']
   $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas64`.chomp}"
   $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas64)`.chomp}"
   $MK_LDFLAGS  << " #{`pkg-config --libs openblas64`}"
-  $OBJ_GGML    << ' ggml-blas.o'
+  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
 end
 
 if $GGML_METAL
   $MK_CPPFLAGS << ' -DGGML_USE_METAL'
   $MK_LDFLAGS  << ' -framework Foundation -framework Metal -framework MetalKit'
-  $OBJ_GGML    << ' ggml-metal.o'
+  $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal.o'
 
   if ENV['GGML_METAL_NDEBUG']
     $MK_CPPFLAGS << ' -DGGML_METAL_NDEBUG'
@@ -181,21 +151,30 @@ if $GGML_METAL
 
   if $GGML_METAL_EMBED_LIBRARY
     $MK_CPPFLAGS << ' -DGGML_METAL_EMBED_LIBRARY'
-    $OBJ_GGML    << ' ggml-metal-embed.o'
+    $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal-embed.o'
   end
 end
 
 $OBJ_GGML <<
-  ' ggml.o' <<
-  ' ggml-alloc.o' <<
-  ' ggml-backend.o' <<
-  ' ggml-quants.o' <<
-  ' ggml-aarch64.o'
+  'ggml/src/ggml.o' <<
+  'ggml/src/ggml-alloc.o' <<
+  'ggml/src/ggml-backend.o' <<
+  'ggml/src/ggml-backend-reg.o' <<
+  'ggml/src/ggml-opt.o' <<
+  'ggml/src/ggml-quants.o' <<
+  'ggml/src/ggml-threading.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-cpp.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-aarch64.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-hbm.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-quants.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-traits.o'
 
 $OBJ_WHISPER <<
-  ' whisper.o'
+  'src/whisper.o'
 
-$OBJ_ALL = "#{$OBJ_GGML} #{$OBJ_WHISPER} #{$OBJ_COMMON} #{$OBJ_SDL}"
+$objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
+$objs << "ruby_whisper.o"
 
 $CPPFLAGS  = "#{$MK_CPPFLAGS} #{$CPPFLAGS}"
 $CFLAGS    = "#{$CPPFLAGS} #{$MK_CFLAGS} #{$GF_CFLAGS} #{$CFLAGS}"
@@ -204,26 +183,16 @@ $CXXFLAGS  = "#{$BASE_CXXFLAGS} #{$HOST_CXXFLAGS} #{$GF_CXXFLAGS} #{$CPPFLAGS}"
 $NVCCFLAGS = "#{$MK_NVCCFLAGS} #{$NVCCFLAGS}"
 $LDFLAGS   = "#{$MK_LDFLAGS} #{$LDFLAGS}"
 
-if $GGML_METAL_EMBED_LIBRARY
-  File.write 'depend', "$(OBJS): $(OBJS) ggml-metal-embed.o\n"
-end
-
-with_disabling_unsupported_files do
-
-  create_makefile('whisper')
-
-end
+create_makefile('whisper')
 
 File.open 'Makefile', 'a' do |file|
-  file.puts 'include get-flags.mk'
+  file.puts 'include scripts/get-flags.mk'
+  file.puts 'include cpu.mk'
 
   if $GGML_METAL
-    if $GGML_METAL_EMBED_LIBRARY
-      # mkmf determines object files to compile dependent on existing *.{c,cpp,m} files
-      # but ggml-metal-embed.c doesn't exist on creating Makefile.
-      file.puts "objs := $(OBJS)"
-      file.puts "OBJS = $(objs) 'ggml-metal-embed.o'"
+    file.puts 'include metal.mk'
 
+    if $GGML_METAL_EMBED_LIBRARY
       file.puts 'include metal-embed.mk'
     end
   end

bindings/ruby/ext/metal-embed.mk

@@ -1,14 +1,17 @@
-ggml-metal-embed.o: \
-	ggml-metal.metal \
-	ggml-common.h
+ggml/src/ggml-metal/ggml-metal-embed.o: \
+	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
+	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml-metal-embed.metal\""          >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
+	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}

bindings/ruby/ext/metal.mk

@@ -0,0 +1,6 @@
+ggml/src/ggml-metal/ggml-metal.o: \
+	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h
+	$(CC) $(CFLAGS) -c $< -o $@

bindings/ruby/ext/ruby_whisper.cpp

@@ -1,4 +1,5 @@
 #include <ruby.h>
+#include <ruby/memory_view.h>
 #include "ruby_whisper.h"
 #define DR_WAV_IMPLEMENTATION
 #include "dr_wav.h"
@@ -35,11 +36,22 @@ extern "C" {
 VALUE mWhisper;
 VALUE cContext;
 VALUE cParams;
+VALUE eError;
+
+VALUE cSegment;
+VALUE cModel;
 
 static ID id_to_s;
 static ID id_call;
 static ID id___method__;
 static ID id_to_enum;
+static ID id_length;
+static ID id_next;
+static ID id_new;
+static ID id_to_path;
+static ID id_pre_converted_models;
+
+static bool is_log_callback_finalized = false;
 
 /*
  * call-seq:
@@ -88,6 +100,39 @@ static VALUE ruby_whisper_s_lang_str_full(VALUE self, VALUE id) {
   return rb_str_new2(str_full);
 }
 
+static VALUE ruby_whisper_s_finalize_log_callback(VALUE self, VALUE id) {
+  is_log_callback_finalized = true;
+  return Qnil;
+}
+
+/*
+ * call-seq:
+ *   log_set ->(level, buffer, user_data) { ... }, user_data -> nil
+ */
+static VALUE ruby_whisper_s_log_set(VALUE self, VALUE log_callback, VALUE user_data) {
+  VALUE old_callback = rb_iv_get(self, "log_callback");
+  if (!NIL_P(old_callback)) {
+    rb_undefine_finalizer(old_callback);
+  }
+
+  rb_iv_set(self, "log_callback", log_callback);
+  rb_iv_set(self, "user_data", user_data);
+
+  VALUE finalize_log_callback = rb_funcall(mWhisper, rb_intern("method"), 1, rb_str_new2("finalize_log_callback"));
+  rb_define_finalizer(log_callback, finalize_log_callback);
+
+  whisper_log_set([](ggml_log_level level, const char * buffer, void * user_data) {
+    if (is_log_callback_finalized) {
+      return;
+    }
+    VALUE log_callback = rb_iv_get(mWhisper, "log_callback");
+    VALUE udata = rb_iv_get(mWhisper, "user_data");
+    rb_funcall(log_callback, id_call, 3, INT2NUM(level), rb_str_new2(buffer), udata);
+  }, nullptr);
+
+  return Qnil;
+}
+
 static void ruby_whisper_free(ruby_whisper *rw) {
   if (rw->context) {
     whisper_free(rw->context);
@@ -146,6 +191,7 @@ static VALUE ruby_whisper_params_allocate(VALUE klass) {
   ruby_whisper_params *rwp;
   rwp = ALLOC(ruby_whisper_params);
   rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  rwp->diarize = false;
   rwp->new_segment_callback_container = rb_whisper_callback_container_allocate();
   rwp->progress_callback_container = rb_whisper_callback_container_allocate();
   rwp->abort_callback_container = rb_whisper_callback_container_allocate();
@@ -154,7 +200,9 @@ static VALUE ruby_whisper_params_allocate(VALUE klass) {
 
 /*
  * call-seq:
+ *   new("base.en") -> Whisper::Context
  *   new("path/to/model.bin") -> Whisper::Context
+ *   new(Whisper::Model::URI.new("https://example.net/uri/of/model.bin")) -> Whisper::Context
  */
 static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   ruby_whisper *rw;
@@ -164,6 +212,14 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   rb_scan_args(argc, argv, "01", &whisper_model_file_path);
   Data_Get_Struct(self, ruby_whisper, rw);
 
+  VALUE pre_converted_models = rb_funcall(cModel, id_pre_converted_models, 0);
+  VALUE pre_converted_model = rb_hash_aref(pre_converted_models, whisper_model_file_path);
+  if (!NIL_P(pre_converted_model)) {
+    whisper_model_file_path = pre_converted_model;
+  }
+  if (rb_respond_to(whisper_model_file_path, id_to_path)) {
+    whisper_model_file_path = rb_funcall(whisper_model_file_path, id_to_path, 0);
+  }
   if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
     rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
   }
@@ -389,6 +445,288 @@ static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
   return self;
 }
 
+/*
+ * call-seq:
+ *   model_n_vocab -> Integer
+ */
+VALUE ruby_whisper_model_n_vocab(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_vocab(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_ctx -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_ctx(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_state -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_state(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_head -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_head(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_layer -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_layer(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_ctx -> Integer
+ */
+VALUE ruby_whisper_model_n_text_ctx(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_state -> Integer
+ */
+VALUE ruby_whisper_model_n_text_state(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_head -> Integer
+ */
+VALUE ruby_whisper_model_n_text_head(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_layer -> Integer
+ */
+VALUE ruby_whisper_model_n_text_layer(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_mels -> Integer
+ */
+VALUE ruby_whisper_model_n_mels(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_mels(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_ftype -> Integer
+ */
+VALUE ruby_whisper_model_ftype(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_ftype(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_type -> String
+ */
+VALUE ruby_whisper_model_type(VALUE self) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return rb_str_new2(whisper_model_type_readable(rw->context));
+}
+
+/*
+ * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+ * Not thread safe for same context
+ * Uses the specified decoding strategy to obtain the text.
+ *
+ * call-seq:
+ *   full(params, samples, n_samples) -> nil
+ *   full(params, samples) -> nil
+ *
+ * The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
+ */
+VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self) {
+  if (argc < 2 || argc > 3) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  if (argc == 3) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (memory_view_available_p) {
+      if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+        view.obj = Qnil;
+        rb_raise(rb_eArgError, "unable to get a memory view");
+      }
+      n_samples = view.byte_size / view.item_size;
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p)  {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // TODO: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError appropriately
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  const int result = whisper_full(rw->context, rwp->params, c_samples, n_samples);
+  if (0 == result) {
+    return Qnil;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
+/*
+ * Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+ * Result is stored in the default state of the context
+ * Not thread safe if executed in parallel on the same context.
+ * It seems this approach can offer some speedup in some cases.
+ * However, the transcription accuracy can be worse at the beginning and end of each chunk.
+ *
+ * call-seq:
+ *   full_parallel(params, samples) -> nil
+ *   full_parallel(params, samples, n_samples) -> nil
+ *   full_parallel(params, samples, n_samples, n_processors) -> nil
+ *   full_parallel(params, samples, nil, n_processors) -> nil
+ */
+static VALUE ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self) {
+  if (argc < 2 || argc > 4) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  int n_processors;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  switch (argc) {
+  case 2:
+    n_processors = 1;
+    break;
+  case 3:
+    n_processors = 1;
+    break;
+  case 4:
+    n_processors = NUM2INT(argv[3]);
+    break;
+  }
+  if (argc >= 3 && !NIL_P(argv[2])) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else if (memory_view_available_p) {
+    if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+      view.obj = Qnil;
+      rb_raise(rb_eArgError, "unable to get a memory view");
+    }
+    n_samples = view.byte_size / view.item_size;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p) {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // FIXME: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
+  if (0 == result) {
+    return Qnil;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
 /*
  * Number of segments.
  *
@@ -923,6 +1261,25 @@ static VALUE ruby_whisper_params_set_logprob_thold(VALUE self, VALUE value) {
   rwp->params.logprob_thold = RFLOAT_VALUE(value);
   return value;
 }
+/*
+ * call-seq:
+ *   no_speech_thold -> Float
+ */
+static VALUE ruby_whisper_params_get_no_speech_thold(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return DBL2NUM(rwp->params.no_speech_thold);
+}
+/*
+ * call-seq:
+ *   no_speech_thold = threshold -> threshold
+ */
+static VALUE ruby_whisper_params_set_no_speech_thold(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.no_speech_thold = RFLOAT_VALUE(value);
+  return value;
+}
 /*
  * Sets new segment callback, called for every newly generated text segment.
  *
@@ -1015,7 +1372,9 @@ typedef struct {
   int index;
 } ruby_whisper_segment;
 
-VALUE cSegment;
+typedef struct {
+  VALUE context;
+} ruby_whisper_model;
 
 static void rb_whisper_segment_mark(ruby_whisper_segment *rws) {
   rb_gc_mark(rws->context);
@@ -1188,31 +1547,270 @@ static VALUE ruby_whisper_segment_get_text(VALUE self) {
   return rb_str_new2(text);
 }
 
+static void rb_whisper_model_mark(ruby_whisper_model *rwm) {
+  rb_gc_mark(rwm->context);
+}
+
+static VALUE ruby_whisper_model_allocate(VALUE klass) {
+  ruby_whisper_model *rwm;
+  rwm = ALLOC(ruby_whisper_model);
+  return Data_Wrap_Struct(klass, rb_whisper_model_mark, RUBY_DEFAULT_FREE, rwm);
+}
+
+static VALUE rb_whisper_model_initialize(VALUE context) {
+  ruby_whisper_model *rwm;
+  const VALUE model = ruby_whisper_model_allocate(cModel);
+  Data_Get_Struct(model, ruby_whisper_model, rwm);
+  rwm->context = context;
+  return model;
+};
+
+/*
+ * call-seq:
+ *   model -> Whisper::Model
+ */
+static VALUE ruby_whisper_get_model(VALUE self) {
+  return rb_whisper_model_initialize(self);
+}
+
+/*
+ * call-seq:
+ *   n_vocab -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_vocab(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_vocab(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_ctx -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_audio_ctx(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_state -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_audio_state(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_head -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_audio_head(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_layer -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_audio_layer(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_ctx -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_text_ctx(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_state -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_text_state(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_head -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_text_head(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_layer -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_text_layer(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_mels -> Integer
+ */
+static VALUE ruby_whisper_c_model_n_mels(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_mels(rw->context));
+}
+
+/*
+ * call-seq:
+ *   ftype -> Integer
+ */
+static VALUE ruby_whisper_c_model_ftype(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_ftype(rw->context));
+}
+
+/*
+ * call-seq:
+ *   type -> String
+ */
+static VALUE ruby_whisper_c_model_type(VALUE self) {
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return rb_str_new2(whisper_model_type_readable(rw->context));
+}
+
+static VALUE ruby_whisper_error_initialize(VALUE self, VALUE code) {
+  const int c_code = NUM2INT(code);
+  char *raw_message;
+  switch (c_code) {
+  case -2:
+    raw_message = "failed to compute log mel spectrogram";
+    break;
+  case -3:
+    raw_message = "failed to auto-detect language";
+    break;
+  case -4:
+    raw_message = "too many decoders requested";
+    break;
+  case -5:
+    raw_message = "audio_ctx is larger than the maximum allowed";
+    break;
+  case -6:
+    raw_message = "failed to encode";
+    break;
+  case -7:
+    raw_message = "whisper_kv_cache_init() failed for self-attention cache";
+    break;
+  case -8:
+    raw_message = "failed to decode";
+    break;
+  case -9:
+    raw_message = "failed to decode";
+    break;
+  default:
+    raw_message = "unknown error";
+    break;
+  }
+  const VALUE message = rb_str_new2(raw_message);
+  rb_call_super(1, &message);
+  rb_iv_set(self, "@code", code);
+
+  return self;
+}
+
+
 void Init_whisper() {
   id_to_s = rb_intern("to_s");
   id_call = rb_intern("call");
   id___method__ = rb_intern("__method__");
   id_to_enum = rb_intern("to_enum");
+  id_length = rb_intern("length");
+  id_next = rb_intern("next");
+  id_new = rb_intern("new");
+  id_to_path = rb_intern("to_path");
+  id_pre_converted_models = rb_intern("pre_converted_models");
 
   mWhisper = rb_define_module("Whisper");
   cContext = rb_define_class_under(mWhisper, "Context", rb_cObject);
   cParams  = rb_define_class_under(mWhisper, "Params", rb_cObject);
+  eError   = rb_define_class_under(mWhisper, "Error", rb_eStandardError);
+
+  rb_define_const(mWhisper, "LOG_LEVEL_NONE", INT2NUM(GGML_LOG_LEVEL_NONE));
+  rb_define_const(mWhisper, "LOG_LEVEL_INFO", INT2NUM(GGML_LOG_LEVEL_INFO));
+  rb_define_const(mWhisper, "LOG_LEVEL_WARN", INT2NUM(GGML_LOG_LEVEL_WARN));
+  rb_define_const(mWhisper, "LOG_LEVEL_ERROR", INT2NUM(GGML_LOG_LEVEL_ERROR));
+  rb_define_const(mWhisper, "LOG_LEVEL_DEBUG", INT2NUM(GGML_LOG_LEVEL_DEBUG));
+  rb_define_const(mWhisper, "LOG_LEVEL_CONT", INT2NUM(GGML_LOG_LEVEL_CONT));
 
   rb_define_singleton_method(mWhisper, "lang_max_id", ruby_whisper_s_lang_max_id, 0);
   rb_define_singleton_method(mWhisper, "lang_id", ruby_whisper_s_lang_id, 1);
   rb_define_singleton_method(mWhisper, "lang_str", ruby_whisper_s_lang_str, 1);
   rb_define_singleton_method(mWhisper, "lang_str_full", ruby_whisper_s_lang_str_full, 1);
+  rb_define_singleton_method(mWhisper, "log_set", ruby_whisper_s_log_set, 2);
+  rb_define_singleton_method(mWhisper, "finalize_log_callback", ruby_whisper_s_finalize_log_callback, 1);
 
   rb_define_alloc_func(cContext, ruby_whisper_allocate);
   rb_define_method(cContext, "initialize", ruby_whisper_initialize, -1);
 
   rb_define_method(cContext, "transcribe", ruby_whisper_transcribe, -1);
+  rb_define_method(cContext, "model_n_vocab", ruby_whisper_model_n_vocab, 0);
+  rb_define_method(cContext, "model_n_audio_ctx", ruby_whisper_model_n_audio_ctx, 0);
+  rb_define_method(cContext, "model_n_audio_state", ruby_whisper_model_n_audio_state, 0);
+  rb_define_method(cContext, "model_n_audio_head", ruby_whisper_model_n_audio_head, 0);
+  rb_define_method(cContext, "model_n_audio_layer", ruby_whisper_model_n_audio_layer, 0);
+  rb_define_method(cContext, "model_n_text_ctx", ruby_whisper_model_n_text_ctx, 0);
+  rb_define_method(cContext, "model_n_text_state", ruby_whisper_model_n_text_state, 0);
+  rb_define_method(cContext, "model_n_text_head", ruby_whisper_model_n_text_head, 0);
+  rb_define_method(cContext, "model_n_text_layer", ruby_whisper_model_n_text_layer, 0);
+  rb_define_method(cContext, "model_n_mels", ruby_whisper_model_n_mels, 0);
+  rb_define_method(cContext, "model_ftype", ruby_whisper_model_ftype, 0);
+  rb_define_method(cContext, "model_type", ruby_whisper_model_type, 0);
   rb_define_method(cContext, "full_n_segments", ruby_whisper_full_n_segments, 0);
   rb_define_method(cContext, "full_lang_id", ruby_whisper_full_lang_id, 0);
   rb_define_method(cContext, "full_get_segment_t0", ruby_whisper_full_get_segment_t0, 1);
   rb_define_method(cContext, "full_get_segment_t1", ruby_whisper_full_get_segment_t1, 1);
   rb_define_method(cContext, "full_get_segment_speaker_turn_next", ruby_whisper_full_get_segment_speaker_turn_next, 1);
   rb_define_method(cContext, "full_get_segment_text", ruby_whisper_full_get_segment_text, 1);
+  rb_define_method(cContext, "full", ruby_whisper_full, -1);
+  rb_define_method(cContext, "full_parallel", ruby_whisper_full_parallel, -1);
 
   rb_define_alloc_func(cParams, ruby_whisper_params_allocate);
 
@@ -1264,6 +1862,8 @@ void Init_whisper() {
   rb_define_method(cParams, "entropy_thold=", ruby_whisper_params_set_entropy_thold, 1);
   rb_define_method(cParams, "logprob_thold", ruby_whisper_params_get_logprob_thold, 0);
   rb_define_method(cParams, "logprob_thold=", ruby_whisper_params_set_logprob_thold, 1);
+  rb_define_method(cParams, "no_speech_thold", ruby_whisper_params_get_no_speech_thold, 0);
+  rb_define_method(cParams, "no_speech_thold=", ruby_whisper_params_set_no_speech_thold, 1);
 
   rb_define_method(cParams, "new_segment_callback=", ruby_whisper_params_set_new_segment_callback, 1);
   rb_define_method(cParams, "new_segment_callback_user_data=", ruby_whisper_params_set_new_segment_callback_user_data, 1);
@@ -1272,6 +1872,9 @@ void Init_whisper() {
   rb_define_method(cParams, "abort_callback=", ruby_whisper_params_set_abort_callback, 1);
   rb_define_method(cParams, "abort_callback_user_data=", ruby_whisper_params_set_abort_callback_user_data, 1);
 
+  rb_define_attr(eError, "code", true, false);
+  rb_define_method(eError, "initialize", ruby_whisper_error_initialize, 1);
+
   // High leve
   cSegment  = rb_define_class_under(mWhisper, "Segment", rb_cObject);
 
@@ -1284,6 +1887,22 @@ void Init_whisper() {
   rb_define_method(cSegment, "end_time", ruby_whisper_segment_get_end_time, 0);
   rb_define_method(cSegment, "speaker_next_turn?", ruby_whisper_segment_get_speaker_turn_next, 0);
   rb_define_method(cSegment, "text", ruby_whisper_segment_get_text, 0);
+
+  cModel = rb_define_class_under(mWhisper, "Model", rb_cObject);
+  rb_define_alloc_func(cModel, ruby_whisper_model_allocate);
+  rb_define_method(cContext, "model", ruby_whisper_get_model, 0);
+  rb_define_method(cModel, "n_vocab", ruby_whisper_c_model_n_vocab, 0);
+  rb_define_method(cModel, "n_audio_ctx", ruby_whisper_c_model_n_audio_ctx, 0);
+  rb_define_method(cModel, "n_audio_state", ruby_whisper_c_model_n_audio_state, 0);
+  rb_define_method(cModel, "n_audio_head", ruby_whisper_c_model_n_audio_head, 0);
+  rb_define_method(cModel, "n_audio_layer", ruby_whisper_c_model_n_audio_layer, 0);
+  rb_define_method(cModel, "n_text_ctx", ruby_whisper_c_model_n_text_ctx, 0);
+  rb_define_method(cModel, "n_text_state", ruby_whisper_c_model_n_text_state, 0);
+  rb_define_method(cModel, "n_text_head", ruby_whisper_c_model_n_text_head, 0);
+  rb_define_method(cModel, "n_text_layer", ruby_whisper_c_model_n_text_layer, 0);
+  rb_define_method(cModel, "n_mels", ruby_whisper_c_model_n_mels, 0);
+  rb_define_method(cModel, "ftype", ruby_whisper_c_model_ftype, 0);
+  rb_define_method(cModel, "type", ruby_whisper_c_model_type, 0);
 }
 #ifdef __cplusplus
 }

bindings/ruby/extsources.rb

@@ -0,0 +1,6 @@
+require "yaml"
+
+sources = `git ls-files -z ../..`.split("\x0")
+paths = YAML.load_file("../../.github/workflows/bindings-ruby.yml")[true]["push"]["paths"]
+paths.delete "bindings/ruby/**"
+EXTSOURCES = (Dir.glob(paths, base: "../..").collect {|path| "../../#{path}"} << "../../LICENSE") & sources

bindings/ruby/extsources.yaml

@@ -1,29 +0,0 @@
----
-- ../../src/whisper.cpp
-- ../../include/whisper.h
-- ../../ggml/src/ggml.c
-- ../../ggml/src/ggml-impl.h
-- ../../ggml/src/ggml-aarch64.h
-- ../../ggml/src/ggml-aarch64.c
-- ../../ggml/src/ggml-alloc.c
-- ../../ggml/src/ggml-backend-impl.h
-- ../../ggml/src/ggml-backend.cpp
-- ../../ggml/src/ggml-common.h
-- ../../ggml/src/ggml-quants.h
-- ../../ggml/src/ggml-quants.c
-- ../../ggml/src/ggml-cpu-impl.h
-- ../../ggml/src/ggml-metal.m
-- ../../ggml/src/ggml-metal.metal
-- ../../ggml/src/ggml-blas.cpp
-- ../../ggml/include/ggml.h
-- ../../ggml/include/ggml-alloc.h
-- ../../ggml/include/ggml-backend.h
-- ../../ggml/include/ggml-cuda.h
-- ../../ggml/include/ggml-kompute.h
-- ../../ggml/include/ggml-metal.h
-- ../../ggml/include/ggml-sycl.h
-- ../../ggml/include/ggml-vulkan.h
-- ../../ggml/include/ggml-blas.h
-- ../../scripts/get-flags.mk
-- ../../examples/dr_wav.h
-- ../../LICENSE

bindings/ruby/lib/whisper.rb

@@ -0,0 +1,2 @@
+require "whisper.so"
+require "whisper/model/uri"

bindings/ruby/lib/whisper/model/uri.rb

@@ -0,0 +1,157 @@
+require "whisper.so"
+require "uri"
+require "net/http"
+require "time"
+require "pathname"
+require "io/console/size"
+
+class Whisper::Model
+  class URI
+    def initialize(uri)
+      @uri = URI(uri)
+    end
+
+    def to_path
+      cache
+      cache_path.to_path
+    end
+
+    def clear_cache
+      path = cache_path
+      path.delete if path.exist?
+    end
+
+    private
+
+    def cache_path
+      base_cache_dir/@uri.host/@uri.path[1..]
+    end
+
+    def base_cache_dir
+      base = case RUBY_PLATFORM
+             when /mswin|mingw/
+               ENV.key?("LOCALAPPDATA") ? Pathname(ENV["LOCALAPPDATA"]) : Pathname(Dir.home)/"AppData/Local"
+             when /darwin/
+               Pathname(Dir.home)/"Library/Caches"
+             else
+               ENV.key?("XDG_CACHE_HOME") ? ENV["XDG_CACHE_HOME"] : Pathname(Dir.home)/".cache"
+             end
+      base/"whisper.cpp"
+    end
+
+    def cache
+      path = cache_path
+      headers = {}
+      headers["if-modified-since"] = path.mtime.httpdate if path.exist?
+      request @uri, headers
+      path
+    end
+
+    def request(uri, headers)
+      Net::HTTP.start uri.host, uri.port, use_ssl: uri.scheme == "https" do |http|
+        request = Net::HTTP::Get.new(uri, headers)
+        http.request request do |response|
+          case response
+          when Net::HTTPNotModified
+            # noop
+          when Net::HTTPOK
+            download response
+          when Net::HTTPRedirection
+            request URI(response["location"]), headers
+          else
+            return if headers.key?("if-modified-since") # Use cache file
+
+            raise "#{response.code} #{response.message}\n#{response.body}"
+          end
+        end
+      end
+    end
+
+    def download(response)
+      path = cache_path
+      path.dirname.mkpath unless path.dirname.exist?
+      downloading_path = Pathname("#{path}.downloading")
+      size = response.content_length
+      downloading_path.open "wb" do |file|
+        downloaded = 0
+        response.read_body do |chunk|
+          file << chunk
+          downloaded += chunk.bytesize
+          show_progress downloaded, size
+        end
+      end
+      downloading_path.rename path
+    end
+
+    def show_progress(current, size)
+      return unless $stderr.tty?
+      return unless size
+
+      unless @prev
+        @prev = Time.now
+        $stderr.puts "Downloading #{@uri}"
+      end
+
+      now = Time.now
+      return if now - @prev < 1 && current < size
+
+      progress_width = 20
+      progress = current.to_f / size
+      arrow_length = progress * progress_width
+      arrow = "=" * (arrow_length - 1) + ">" + " " * (progress_width - arrow_length)
+      line = "[#{arrow}] (#{format_bytesize(current)} / #{format_bytesize(size)})"
+      padding = ' ' * ($stderr.winsize[1] - line.size)
+      $stderr.print "\r#{line}#{padding}"
+      $stderr.puts if current >= size
+      @prev = now
+    end
+
+    def format_bytesize(bytesize)
+      return "0.0 B" if bytesize.zero?
+
+      units = %w[B KiB MiB GiB TiB]
+      exp = (Math.log(bytesize) / Math.log(1024)).to_i
+      format("%.1f %s", bytesize.to_f / 1024 ** exp, units[exp])
+    end
+  end
+
+  @pre_converted_models = {}
+  %w[
+    tiny
+    tiny.en
+    tiny-q5_1
+    tiny.en-q5_1
+    tiny-q8_0
+    base
+    base.en
+    base-q5_1
+    base.en-q5_1
+    base-q8_0
+    small
+    small.en
+    small.en-tdrz
+    small-q5_1
+    small.en-q5_1
+    small-q8_0
+    medium
+    medium.en
+    medium-q5_0
+    medium.en-q5_0
+    medium-q8_0
+    large-v1
+    large-v2
+    large-v2-q5_0
+    large-v2-q8_0
+    large-v3
+    large-v3-q5_0
+    large-v3-turbo
+    large-v3-turbo-q5_0
+    large-v3-turbo-q8_0
+  ].each do |name|
+    @pre_converted_models[name] = URI.new("https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-#{name}.bin")
+  end
+
+  class << self
+    attr_reader :pre_converted_models
+  end
+end

bindings/ruby/tests/helper.rb

@@ -0,0 +1,7 @@
+require "test/unit"
+require "whisper"
+require_relative "jfk_reader/jfk_reader"
+
+class TestBase < Test::Unit::TestCase
+  AUDIO = File.join(__dir__, "..", "..", "..", "samples", "jfk.wav")
+end

bindings/ruby/tests/jfk_reader/.gitignore

@@ -0,0 +1,5 @@
+Makefile
+jfk_reader.o
+jfk_reader.so
+jfk_reader.bundle
+jfk_reader.dll

bindings/ruby/tests/jfk_reader/extconf.rb

@@ -0,0 +1,3 @@
+require "mkmf"
+
+create_makefile("jfk_reader")

bindings/ruby/tests/jfk_reader/jfk_reader.c

@@ -0,0 +1,68 @@
+#include <ruby.h>
+#include <ruby/memory_view.h>
+#include <ruby/encoding.h>
+
+static VALUE
+jfk_reader_initialize(VALUE self, VALUE audio_path)
+{
+  rb_iv_set(self, "audio_path", audio_path);
+  return Qnil;
+}
+
+static bool
+jfk_reader_get_memory_view(const VALUE obj, rb_memory_view_t *view, int flags)
+{
+  VALUE audio_path = rb_iv_get(obj, "audio_path");
+  const char *audio_path_str = StringValueCStr(audio_path);
+  const int n_samples = 176000;
+  float *data = (float *)malloc(n_samples * sizeof(float));
+  short *samples = (short *)malloc(n_samples * sizeof(short));
+  FILE *file = fopen(audio_path_str, "rb");
+
+  fseek(file, 78, SEEK_SET);
+  fread(samples, sizeof(short), n_samples, file);
+  fclose(file);
+  for (int i = 0; i < n_samples; i++) {
+    data[i] = samples[i]/32768.0;
+  }
+
+  view->obj = obj;
+  view->data = (void *)data;
+  view->byte_size = sizeof(float) * n_samples;
+  view->readonly = true;
+  view->format = "f";
+  view->item_size = sizeof(float);
+  view->item_desc.components = NULL;
+  view->item_desc.length = 0;
+  view->ndim = 1;
+  view->shape = NULL;
+  view->sub_offsets = NULL;
+  view->private_data = NULL;
+
+  return true;
+}
+
+static bool
+jfk_reader_release_memory_view(const VALUE obj, rb_memory_view_t *view)
+{
+  return true;
+}
+
+static bool
+jfk_reader_memory_view_available_p(const VALUE obj)
+{
+  return true;
+}
+
+static const rb_memory_view_entry_t jfk_reader_view_entry = {
+  jfk_reader_get_memory_view,
+  jfk_reader_release_memory_view,
+  jfk_reader_memory_view_available_p
+};
+
+void Init_jfk_reader(void)
+{
+  VALUE cJFKReader = rb_define_class("JFKReader", rb_cObject);
+  rb_memory_view_register(cJFKReader, &jfk_reader_view_entry);
+  rb_define_method(cJFKReader, "initialize", jfk_reader_initialize, 1);
+}

bindings/ruby/tests/test_callback.rb

@@ -1,14 +1,11 @@
-require "test/unit"
-require "whisper"
-
-class TestCallback < Test::Unit::TestCase
-  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+require_relative "helper"
 
+class TestCallback < TestBase
   def setup
     GC.start
     @params = Whisper::Params.new
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
-    @audio = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+    @whisper = Whisper::Context.new("base.en")
+    @audio = File.join(AUDIO)
   end
 
   def test_new_segment_callback

bindings/ruby/tests/test_error.rb

@@ -0,0 +1,20 @@
+require_relative "helper"
+
+class TestError < TestBase
+  def test_error
+    error = Whisper::Error.new(-2)
+    assert_equal "failed to compute log mel spectrogram", error.message
+    assert_equal -2, error.code
+  end
+
+  def test_unknown_error
+    error = Whisper::Error.new(-20)
+    assert_equal "unknown error", error.message
+  end
+
+  def test_non_int_code
+    assert_raise TypeError do
+      error = Whisper::Error.new("non int")
+    end
+  end
+end

bindings/ruby/tests/test_model.rb

@@ -0,0 +1,71 @@
+require_relative "helper"
+require "pathname"
+
+class TestModel < TestBase
+  def test_model
+    whisper = Whisper::Context.new("base.en")
+    assert_instance_of Whisper::Model, whisper.model
+  end
+
+  def test_attributes
+    whisper = Whisper::Context.new("base.en")
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_gc
+    model = Whisper::Context.new("base.en").model
+    GC.start
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_pathname
+    path = Pathname(Whisper::Model.pre_converted_models["base.en"].to_path)
+    whisper = Whisper::Context.new(path)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_auto_download
+    path = Whisper::Model.pre_converted_models["base.en"].to_path
+
+    assert_path_exist path
+    assert_equal 147964211, File.size(path)
+  end
+end

bindings/ruby/tests/test_package.rb

@@ -1,9 +1,9 @@
-require 'test/unit'
+require_relative "helper"
 require 'tempfile'
 require 'tmpdir'
 require 'shellwords'
 
-class TestPackage < Test::Unit::TestCase
+class TestPackage < TestBase
   def test_build
     Tempfile.create do |file|
       assert system("gem", "build", "whispercpp.gemspec", "--output", file.to_path.shellescape, exception: true)

bindings/ruby/tests/test_params.rb

@@ -1,7 +1,6 @@
-require 'test/unit'
-require 'whisper'
+require_relative "helper"
 
-class TestParams < Test::Unit::TestCase
+class TestParams < TestBase
   def setup
     @params  = Whisper::Params.new
   end
@@ -152,4 +151,10 @@ class TestParams < Test::Unit::TestCase
     @params.logprob_thold = -0.5
     assert_in_delta -0.5, @params.logprob_thold
   end
+
+  def test_no_speech_thold
+    assert_in_delta 0.6, @params.no_speech_thold
+    @params.no_speech_thold = 0.2
+    assert_in_delta 0.2, @params.no_speech_thold
+  end
 end

bindings/ruby/tests/test_segment.rb

@@ -1,18 +1,14 @@
-require "test/unit"
-require "whisper"
-
-class TestSegment < Test::Unit::TestCase
-  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+require_relative "helper"
 
+class TestSegment < TestBase
   class << self
     attr_reader :whisper
 
     def startup
-      @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+      @whisper = Whisper::Context.new("base.en")
       params = Whisper::Params.new
       params.print_timestamps = false
-      jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
-      @whisper.transcribe(jfk, params)
+      @whisper.transcribe(TestBase::AUDIO, params)
     end
   end
 
@@ -60,7 +56,7 @@ class TestSegment < Test::Unit::TestCase
       end
       index += 1
     end
-    whisper.transcribe(File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav'), params)
+    whisper.transcribe(AUDIO, params)
     assert_equal 0, seg.start_time
     assert_match /ask not what your country can do for you, ask what you can do for your country/, seg.text
   end
@@ -76,7 +72,7 @@ class TestSegment < Test::Unit::TestCase
       assert_same seg, segment
       return
     end
-    whisper.transcribe(File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav'), params)
+    whisper.transcribe(AUDIO, params)
   end
 
   private

bindings/ruby/tests/test_whisper.rb

@@ -1,20 +1,21 @@
-require 'whisper'
-require 'test/unit'
+require_relative "helper"
+require "stringio"
+require "etc"
 
-class TestWhisper < Test::Unit::TestCase
-  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+# Exists to detect memory-related bug
+Whisper.log_set ->(level, buffer, user_data) {}, nil
 
+class TestWhisper < TestBase
   def setup
     @params  = Whisper::Params.new
   end
 
   def test_whisper
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    @whisper = Whisper::Context.new("base.en")
     params  = Whisper::Params.new
     params.print_timestamps = false
 
-    jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
-    @whisper.transcribe(jfk, params) {|text|
+    @whisper.transcribe(AUDIO, params) {|text|
       assert_match /ask not what your country can do for you, ask what you can do for your country/, text
     }
   end
@@ -24,11 +25,10 @@ class TestWhisper < Test::Unit::TestCase
       attr_reader :whisper
 
       def startup
-        @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+        @whisper = Whisper::Context.new("base.en")
         params = Whisper::Params.new
         params.print_timestamps = false
-        jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
-        @whisper.transcribe(jfk, params)
+        @whisper.transcribe(TestBase::AUDIO, params)
       end
     end
 
@@ -96,4 +96,131 @@ class TestWhisper < Test::Unit::TestCase
       Whisper.lang_str_full(Whisper.lang_max_id + 1)
     end
   end
+
+  def test_log_set
+    user_data = Object.new
+    logs = []
+    log_callback = ->(level, buffer, udata) {
+      logs << [level, buffer, udata]
+    }
+    Whisper.log_set log_callback, user_data
+    Whisper::Context.new("base.en")
+
+    assert logs.length > 30
+    logs.each do |log|
+      assert_include [Whisper::LOG_LEVEL_DEBUG, Whisper::LOG_LEVEL_INFO, Whisper::LOG_LEVEL_WARN], log[0]
+      assert_same user_data, log[2]
+    end
+  end
+
+  def test_log_suppress
+    stderr = $stderr
+    Whisper.log_set ->(level, buffer, user_data) {
+      # do nothing
+    }, nil
+    dev = StringIO.new("")
+    $stderr = dev
+    Whisper::Context.new("base.en")
+    assert_empty dev.string
+  ensure
+    $stderr = stderr
+  end
+
+  sub_test_case "full" do
+    def setup
+      super
+      @whisper = Whisper::Context.new("base.en")
+      @samples = File.read(AUDIO, nil, 78).unpack("s<*").collect {|i| i.to_f / 2**15}
+    end
+
+    def test_full
+      @whisper.full(@params, @samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_without_length
+      @whisper.full(@params, @samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_enumerator
+      samples = @samples.each
+      @whisper.full(@params, samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_enumerator_without_length
+      samples = @samples.each
+      assert_raise ArgumentError do
+        @whisper.full(@params, samples)
+      end
+    end
+
+    def test_full_enumerator_with_too_large_length
+      samples = @samples.each.take(10).to_enum
+      assert_raise StopIteration do
+        @whisper.full(@params, samples, 11)
+      end
+    end
+
+    def test_full_with_memory_view
+      samples = JFKReader.new(AUDIO)
+      @whisper.full(@params, samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_parallel
+      @whisper.full_parallel(@params, @samples, @samples.length, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_with_memory_view
+      samples = JFKReader.new(AUDIO)
+      @whisper.full_parallel(@params, samples, nil, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_length_and_n_processors
+      @whisper.full_parallel(@params, @samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_length
+      @whisper.full_parallel(@params, @samples, nil, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_n_processors
+      @whisper.full_parallel(@params, @samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+  end
 end

bindings/ruby/whispercpp.gemspec

@@ -1,4 +1,4 @@
-require "yaml"
+require_relative "extsources"
 
 Gem::Specification.new do |s|
   s.name    = "whispercpp"
@@ -10,24 +10,24 @@ Gem::Specification.new do |s|
   s.extra_rdoc_files = ['LICENSE', 'README.md']
   
   s.files = `git ls-files . -z`.split("\x0") +
-              YAML.load_file("extsources.yaml").collect {|file|
+              EXTSOURCES.collect {|file|
                 basename = File.basename(file)
                 if s.extra_rdoc_files.include?(basename)
                   basename
                 else
-                  File.join("ext", basename)
+                  file.sub("../..", "ext")
                 end
               }
 
   s.summary = %q{Ruby whisper.cpp bindings}
-  s.test_files = ["tests/test_whisper.rb"]
+  s.test_files = s.files.select {|file| file.start_with? "tests/"}
   
   s.extensions << 'ext/extconf.rb'
-  
+  s.required_ruby_version = '>= 3.1.0'
 
   #### Documentation and testing.
   s.homepage = 'https://github.com/ggerganov/whisper.cpp'
-  s.rdoc_options = ['--main', '../../README.md']
+  s.rdoc_options = ['--main', 'README.md']
 
   
     s.platform = Gem::Platform::RUBY

cmake/whisper.pc.in

@@ -1,10 +1,10 @@
 prefix=@CMAKE_INSTALL_PREFIX@
 exec_prefix=${prefix}
-libdir=@CMAKE_INSTALL_FULL_LIBDIR@
+libdir=${exec_prefix}/lib
 includedir=${prefix}/include
 
 Name: whisper
 Description: Port of OpenAI's Whisper model in C/C++
 Version: @PROJECT_VERSION@
-Libs: -L${libdir} -lwhisper
+Libs: -L${libdir} -lggml  -lggml-base -lwhisper
 Cflags: -I${includedir}

examples/CMakeLists.txt

@@ -137,7 +137,7 @@ if (WHISPER_SDL2)
     set_target_properties(lsp PROPERTIES FOLDER "examples")
     if (GGML_SYCL)
         add_subdirectory(sycl)
-        set_target_properties(sycl PROPERTIES FOLDER "examples")
+        set_target_properties(ls-sycl-device PROPERTIES FOLDER "examples")
     endif()
 endif (WHISPER_SDL2)
 endif()

examples/common-ggml.cpp

@@ -72,9 +72,6 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_XS:
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_4:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_8:
-        case GGML_FTYPE_MOSTLY_Q4_0_8_8:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -212,9 +209,6 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_IQ4_XS:
                 case GGML_TYPE_IQ1_M:
                 case GGML_TYPE_BF16:
-                case GGML_TYPE_Q4_0_4_4:
-                case GGML_TYPE_Q4_0_4_8:
-                case GGML_TYPE_Q4_0_8_8:
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
                 case GGML_TYPE_COUNT:

examples/common-sdl.cpp

@@ -1,5 +1,7 @@
 #include "common-sdl.h"
 
+#include <cstdio>
+
 audio_async::audio_async(int len_ms) {
     m_len_ms = len_ms;
 

examples/ffmpeg-transcode.cpp

@@ -204,8 +204,6 @@ static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
     const size_t errbuffsize = 1024;
     char errbuff[errbuffsize];
 
-    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
-
     fmt_ctx = avformat_alloc_context();
     avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
     LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);

examples/stream/README.md

@@ -5,7 +5,7 @@ The `stream` tool samples the audio every half a second and runs the transcripti
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
@@ -15,7 +15,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Setting the `--step` argument to `0` enables the sliding window mode:
 
 ```bash
- ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
+ ./build/bin/stream -m ./models/ggml-base.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
 In this mode, the tool will transcribe only after some speech activity is detected. A very
@@ -40,21 +40,10 @@ sudo dnf install SDL2 SDL2-devel
 # Install SDL2 on Mac OS
 brew install sdl2
 
-make stream
-```
+cmake -B build -DWHISPER_SDL2=ON
+cmake --build build --config Release
 
-Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
-`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
-
-```bash
-whisper.cpp/examples/stream$ make stream
-g++     stream.cpp   -o stream
-stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
-    6 | #include "common/sdl.h"
-      |          ^~~~~~~~~~~~~~
-compilation terminated.
-make: *** [<builtin>: stream] Error 1
+./build/bin/stream
 ```
 
 ## Web version

examples/sycl/CMakeLists.txt

@@ -5,5 +5,5 @@
 set(TARGET ls-sycl-device)
 add_executable(${TARGET} ls-sycl-device.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/build.sh

@@ -7,13 +7,16 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
 
 #for FP32
-cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#for other features from the examples, e.g. stream and talk link with SDL2:
+#cmake .. -DGGML_SYCL=ON -DWHISPER_SDL2=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 #build example/main only
 #cmake --build . --config Release --target main
 
 #build all binary
-cmake --build . --config Release -v
+cmake --build . --config Release -v

examples/talk-llama/llama-sampling.h

@@ -4,8 +4,6 @@
 
 #include "llama-grammar.h"
 
-#include <unordered_map>
-
 struct llama_vocab;
 struct llama_grammar;
 
@@ -27,3 +25,24 @@ struct llama_sampler * llama_sampler_init_grammar_impl(
         const struct llama_vocab & vocab,
                       const char * grammar_str,
                       const char * grammar_root);
+
+struct llama_sampler * llama_sampler_init_infill_impl(
+        const struct llama_vocab & vocab);
+
+struct llama_sampler * llama_sampler_init_dry_impl(
+        const struct llama_vocab &  vocab,
+                         int32_t    context_size,
+                           float    dry_multiplier,
+                           float    dry_base,
+                         int32_t    dry_allowed_length,
+                         int32_t    dry_penalty_last_n,
+                      const char ** seq_breakers,
+                          size_t    num_breakers);
+
+struct llama_sampler * llama_sampler_init_dry_testing(
+                         int32_t   context_size,
+                           float   dry_multiplier,
+                           float   dry_base,
+                         int32_t   dry_allowed_length,
+                         int32_t   dry_penalty_last_n,
+  const std::vector<std::vector<llama_token>>& seq_breakers);

examples/talk-llama/llama-vocab.cpp

@@ -221,7 +221,7 @@ struct llm_tokenizer_spm_session {
         }
 
         // seed the work queue with all possible 2-character tokens.
-        for (size_t i = 1; i < symbols.size(); ++i) {
+        for (int i = 1; i < (int) symbols.size(); ++i) {
             try_add_bigram(i - 1, i);
         }
 
@@ -418,6 +418,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
             case LLAMA_VOCAB_PRE_TYPE_EXAONE:
+            case LLAMA_VOCAB_PRE_TYPE_MINERVA:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -563,7 +564,7 @@ struct llm_tokenizer_bpe_session {
                 index++;
                 symbols.emplace_back(sym);
             }
-            for (size_t i = 1; i < symbols.size(); ++i) {
+            for (int i = 1; i < (int) symbols.size(); ++i) {
                 add_new_bigram(i - 1, i);
             }
 
@@ -737,7 +738,7 @@ struct llm_tokenizer_wpm_session {
         std::vector<std::string> words(1, "");
 
         for (const uint32_t cpt : cpts_nfd) {
-            const auto flags = unicode_cpt_flags(cpt);
+            const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
                 if (words.back().size()) {  // finish previous word if any
@@ -1663,6 +1664,14 @@ llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {
     return vocab.special_eos_id;
 }
 
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eot_id;
+}
+
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
+    return vocab.special_eom_id;
+}
+
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab) {
     return vocab.special_cls_id;
 }
@@ -1688,23 +1697,39 @@ bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
 }
 
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_prefix_id;
+    return vocab.special_fim_pre_id;
 }
 
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab) {
-    return vocab.special_middle_id;
+    return vocab.special_fim_mid_id;
 }
 
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab) {
-    return vocab.special_suffix_id;
+    return vocab.special_fim_suf_id;
 }
 
-llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eot_id;
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pre_id;
 }
 
-llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
-    return vocab.special_eom_id;
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_suf_id;
+}
+
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_mid_id;
+}
+
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_pad_id;
+}
+
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_rep_id;
+}
+
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab) {
+    return vocab.special_fim_sep_id;
 }
 
 int32_t llama_tokenize_impl(
@@ -1942,3 +1967,19 @@ int32_t llama_detokenize_impl(
 
     return total <= text_len_max ? total : -total;
 }
+
+std::string llama_detokenize(const struct llama_vocab & vocab, const std::vector<llama_token> & tokens, bool special) {
+    std::string text;
+    text.resize(std::max(text.capacity(), tokens.size()));
+    int32_t n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    if (n_chars < 0) {
+        text.resize(-n_chars);
+        n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
+    }
+
+    text.resize(n_chars);
+
+    // NOTE: the original tokenizer decodes bytes after collecting the pieces.
+    return text;
+}

examples/talk-llama/llama-vocab.h

@@ -37,20 +37,26 @@ struct llama_vocab {
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
     // default LLaMA special tokens
+    // TODO: should we set all of these to LLAMA_TOKEN_NULL?
     id special_bos_id  = 1;
     id special_eos_id  = 2;
+    id special_eot_id  = LLAMA_TOKEN_NULL;
+    id special_eom_id  = LLAMA_TOKEN_NULL;
     id special_unk_id  = 0;
-    id special_sep_id  = -1;
-    id special_pad_id  = -1;
-    id special_cls_id  = -1;
-    id special_mask_id = -1;
+    id special_sep_id  = LLAMA_TOKEN_NULL;
+    id special_pad_id  = LLAMA_TOKEN_NULL;
+    id special_cls_id  = LLAMA_TOKEN_NULL;
+    id special_mask_id = LLAMA_TOKEN_NULL;
 
-    id linefeed_id       = 13;
-    id special_prefix_id = -1;
-    id special_suffix_id = -1;
-    id special_middle_id = -1;
-    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
-    id special_eom_id    = -1;
+    id linefeed_id = 13;
+
+    // fim tokens
+    id special_fim_pre_id = LLAMA_TOKEN_NULL;
+    id special_fim_suf_id = LLAMA_TOKEN_NULL;
+    id special_fim_mid_id = LLAMA_TOKEN_NULL;
+    id special_fim_pad_id = LLAMA_TOKEN_NULL;
+    id special_fim_rep_id = LLAMA_TOKEN_NULL; // repo
+    id special_fim_sep_id = LLAMA_TOKEN_NULL; // file separator
 
     // set of all tokens that cause "end of generation"
     std::set<id> special_eog_ids;
@@ -104,19 +110,26 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
 
 llama_token llama_token_bos_impl(const struct llama_vocab & vocab);
 llama_token llama_token_eos_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eot_impl(const struct llama_vocab & vocab);
+llama_token llama_token_eom_impl(const struct llama_vocab & vocab);
 llama_token llama_token_cls_impl(const struct llama_vocab & vocab);
 llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
 
-bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
-bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
-
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
 llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
-llama_token llama_token_eot_impl   (const struct llama_vocab & vocab);
-llama_token llama_token_eom_impl   (const struct llama_vocab & vocab);
+
+llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab);
+llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab);
+
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
 
 int32_t llama_tokenize_impl(
         const struct llama_vocab & vocab,
@@ -136,6 +149,12 @@ int32_t llama_token_to_piece_impl(
                          int32_t   lstrip,
                             bool   special);
 
+// check if token0 is contained as a prefix in token1
+bool llama_token_is_prefix_impl(
+        const struct llama_vocab & vocab,
+                     llama_token   token0,
+                     llama_token   token1);
+
 int32_t llama_detokenize_impl(
         const struct llama_vocab & vocab,
                const llama_token * tokens,
@@ -144,3 +163,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special);
+
+std::string llama_detokenize(
+        const struct llama_vocab & vocab,
+  const std::vector<llama_token> & tokens,
+                            bool   special);

examples/talk-llama/llama.h

@@ -2,6 +2,7 @@
 #define LLAMA_H
 
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-backend.h"
 
 #include <stddef.h>
@@ -103,12 +104,15 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
         LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
+        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
     };
 
     enum llama_rope_type {
-        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM = 0,
-        LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_NONE   = -1,
+        LLAMA_ROPE_TYPE_NORM   = 0,
+        LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
     enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -170,9 +174,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // except 1d tensors
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // removed from gguf files, use Q4_0 and runtime repack
         LLAMA_FTYPE_MOSTLY_TQ1_0         = 36, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_TQ2_0         = 37, // except 1d tensors
 
@@ -184,7 +188,8 @@ extern "C" {
         LLAMA_ROPE_SCALING_TYPE_NONE        = 0,
         LLAMA_ROPE_SCALING_TYPE_LINEAR      = 1,
         LLAMA_ROPE_SCALING_TYPE_YARN        = 2,
-        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_YARN,
+        LLAMA_ROPE_SCALING_TYPE_LONGROPE    = 3,
+        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_LONGROPE,
     };
 
     enum llama_pooling_type {
@@ -205,7 +210,7 @@ extern "C" {
     enum llama_split_mode {
         LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
         LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
+        LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
 
     // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
@@ -217,6 +222,7 @@ extern "C" {
 
     typedef struct llama_token_data_array {
         // TODO: consider SoA
+        // NOTE: this pointer can be modified by the samplers
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
@@ -232,8 +238,11 @@ extern "C" {
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
+    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
     // - seq_id : the sequence to which the respective token belongs
+    //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+    //            (if set to NULL, only the logits for last token will be returned)
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -244,15 +253,6 @@ extern "C" {
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
         int8_t       *  logits; // TODO: rename this to "output"
-
-        // NOTE: helpers for smooth API transition - can be deprecated in the future
-        //       for future-proof code, use the above fields instead and ignore everything below
-        //
-        // pos[i] = all_pos_0 + i*all_pos_1
-        //
-        llama_pos    all_pos_0;  // used if pos == NULL
-        llama_pos    all_pos_1;  // used if pos == NULL
-        llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
     enum llama_model_kv_override_type {
@@ -276,13 +276,13 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
+        ggml_backend_dev_t * devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
-        // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_MODE_LAYER: ignored
+        // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
         int32_t main_gpu;
 
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
@@ -433,6 +433,7 @@ extern "C" {
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
     LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API bool llama_supports_rpc        (void);
 
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
@@ -457,6 +458,7 @@ extern "C" {
     // Functions to access the model's GGUF metadata scalar values
     // - The functions return the length of the string on success, or -1 on failure
     // - The output string is always null-terminated and cleared on failure
+    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
     // - GGUF array values are not supported by these functions
 
     // Get metadata value as a string by key name
@@ -673,6 +675,9 @@ extern "C" {
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
     LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
 
+    // Check if the context supports KV cache shifting
+    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+
     //
     // State / sessions
     //
@@ -775,15 +780,15 @@ extern "C" {
     // Decoding
     //
 
-    // Return batch for single sequence of tokens starting at pos_0
+    // Return batch for single sequence of tokens
+    // The sequence ID will be fixed to 0
+    // The position of the tokens will be tracked automatically by llama_decode
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
     //
     LLAMA_API struct llama_batch llama_batch_get_one(
                   llama_token * tokens,
-                      int32_t   n_tokens,
-                    llama_pos   pos_0,
-                 llama_seq_id   seq_id);
+                      int32_t   n_tokens);
 
     // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
     // Each token can be assigned up to n_seq_max sequence ids
@@ -803,7 +808,7 @@ extern "C" {
     // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
     // Stores the encoder output internally for later use by the decoder cross-attention layers.
     //   0 - success
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -811,7 +816,7 @@ extern "C" {
     // Positive return values does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    // < 0 - error
+    // < 0 - error. the KV cache state is restored to the state before this call
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -896,6 +901,7 @@ extern "C" {
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
     LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
     LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
@@ -904,11 +910,17 @@ extern "C" {
     LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
     LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
 
-    // Codellama infill tokens
-    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
-    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
-    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
-    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
+    // infill tokens
+    DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
+
+    LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
+    LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
 
     //
     // Tokenization
@@ -983,6 +995,9 @@ extern "C" {
                                   char * buf,
                                int32_t   length);
 
+    // Get list of built-in chat templates
+    LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len);
+
     //
     // Sampling API
     //
@@ -1067,12 +1082,13 @@ extern "C" {
 
     // available samplers:
 
-    LLAMA_API struct llama_sampler * llama_sampler_init_greedy     (void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist       (uint32_t seed);
+    LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void);
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
+        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1083,16 +1099,18 @@ extern "C" {
     /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
-    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API struct llama_sampler * llama_sampler_init_tail_free  (float   z, size_t min_keep);
-
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
     LLAMA_API struct llama_sampler * llama_sampler_init_typical    (float   p, size_t min_keep);
+
+    /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
     LLAMA_API struct llama_sampler * llama_sampler_init_temp       (float   t);
 
     /// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
     LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext   (float   t, float   delta, float exponent);
 
+    /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
+    LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1121,22 +1139,50 @@ extern "C" {
                           const char * grammar_str,
                           const char * grammar_root);
 
+    /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
-                             int32_t   n_vocab,         // llama_n_vocab()
-                         llama_token   special_eos_id,  // llama_token_eos()
-                         llama_token   linefeed_id,     // llama_token_nl()
-                             int32_t   penalty_last_n,  // last n tokens to penalize (0 = disable penalty, -1 = context size)
-                               float   penalty_repeat,  // 1.0 = disabled
-                               float   penalty_freq,    // 0.0 = disabled
-                               float   penalty_present, // 0.0 = disabled
-                                bool   penalize_nl,     // consider newlines as a repeatable token
-                                bool   ignore_eos);     // ignore the end-of-sequence token
+                             int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
+                               float   penalty_repeat,   // 1.0 = disabled
+                               float   penalty_freq,     // 0.0 = disabled
+                               float   penalty_present); // 0.0 = disabled
+
+    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
+    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
+            const struct llama_model *  model,
+                               float    dry_multiplier,
+                               float    dry_base,
+                             int32_t    dry_allowed_length,
+                             int32_t    dry_penalty_last_n,
+                          const char ** seq_breakers,
+                              size_t    num_breakers);
 
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,
               const llama_logit_bias * logit_bias);
 
+    // this sampler is meant to be used for fill-in-the-middle infilling
+    // it's supposed to be used after top_k + top_p sampling
+    //
+    // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
+    // 2. combine probs of tokens that have the same prefix
+    //
+    // example:
+    //
+    // - before:
+    //   "hel":   0.5
+    //   "hell":  0.2
+    //   "hello": 0.1
+    //   "dummy": 0.1
+    //
+    // - after:
+    //   "hel":   0.8
+    //   "dummy": 0.1
+    //
+    // 3. discard non-EOG tokens with low prob
+    // 4. if no tokens are left -> pick EOT
+    //
+    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
 
     // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
     LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
@@ -1208,8 +1254,6 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
-    LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
-
 #ifdef __cplusplus
 }
 #endif

examples/talk-llama/unicode-data.cpp

@@ -2311,7 +2311,7 @@ const std::unordered_set<uint32_t> unicode_set_whitespace = {
 0x003000,
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase = {
 {0x000041, 0x000061},
 {0x000042, 0x000062},
@@ -3748,7 +3748,7 @@ const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_lowercase
 {0x01E921, 0x01E943},
 };
 
-// list is always in ascending order, to enable binary searh
+// list is always in ascending order, to enable binary search
 const std::initializer_list<std::pair<uint32_t, uint32_t>> unicode_map_uppercase = {
 {0x000061, 0x000041},
 {0x000062, 0x000042},

examples/talk-llama/unicode.cpp

@@ -71,15 +71,15 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
     throw std::invalid_argument("failed to convert utf8 to codepoint");
 }
 
-//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cp) {
+//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cpt) {
 //    std::vector<uint16_t> result;
-//    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
-//        result.emplace_back(cp);
+//    if (/* 0x0000 <= cpt && */ cpt <= 0xffff) {
+//        result.emplace_back(cpt);
 //        return result;
 //    }
-//    if (0x10000 <= cp && cp <= 0x10ffff) {
-//        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
-//        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+//    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+//        result.emplace_back(0xd800 | ((cpt - 0x10000) >> 10));
+//        result.emplace_back(0xdc00 | ((cpt - 0x10000) & 0x03ff));
 //        return result;
 //    }
 //    throw std::invalid_argument("failed to convert codepoint to utf16");
@@ -120,8 +120,8 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
 //    return result;
 //}
 
-static std::vector<codepoint_flags> unicode_cpt_flags_array() {
-    std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+static std::vector<unicode_cpt_flags> unicode_cpt_flags_array() {
+    std::vector<unicode_cpt_flags> cpt_flags(MAX_CODEPOINTS, unicode_cpt_flags::UNDEFINED);
 
     assert (unicode_ranges_flags.begin()[0].first == 0);
     assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS);
@@ -201,7 +201,18 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
 }
 
 static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+#if defined(__clang__)
+    // disable C++17 deprecation warning for std::codecvt_utf8
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
     return conv.from_bytes(s);
 }
 
@@ -242,8 +253,8 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -360,8 +371,8 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -561,29 +572,29 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
 // interface
 //
 
-std::string unicode_cpt_to_utf8(uint32_t cp) {
+std::string unicode_cpt_to_utf8(uint32_t cpt) {
     std::string result;
 
-    if (/* 0x00 <= cp && */ cp <= 0x7f) {
-        result.push_back(cp);
+    if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
+        result.push_back(cpt);
         return result;
     }
-    if (0x80 <= cp && cp <= 0x7ff) {
-        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x80 <= cpt && cpt <= 0x7ff) {
+        result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x800 <= cp && cp <= 0xffff) {
-        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x800 <= cpt && cpt <= 0xffff) {
+        result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x10000 <= cp && cp <= 0x10ffff) {
-        result.push_back(0xf0 | ((cp >> 18) & 0x07));
-        result.push_back(0x80 | ((cp >> 12) & 0x3f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+        result.push_back(0xf0 | ((cpt >> 18) & 0x07));
+        result.push_back(0x80 | ((cpt >> 12) & 0x3f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
 
@@ -613,19 +624,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     return result;
 }
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     static const auto cpt_flags = unicode_cpt_flags_array();
-    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
+    return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;
 }
 
-codepoint_flags unicode_cpt_flags(const std::string & utf8) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     if (utf8.empty()) {
         return undef;  // undefined
     }
     size_t offset = 0;
-    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
+    return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));
 }
 
 std::string unicode_byte_to_utf8(uint8_t byte) {
@@ -638,41 +649,41 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) {
     return map.at(utf8);
 }
 
-uint32_t unicode_tolower(uint32_t cp) {
+uint32_t unicode_tolower(uint32_t cpt) {
     // binary search
-    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cp,
+    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,
         [](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {
             return pair.first < value;
         });
-    if (it != unicode_map_lowercase.end() && it->first == cp) {
+    if (it != unicode_map_lowercase.end() && it->first == cpt) {
         return it->second;
     }
-    return cp;  // Return the original code point if no lowercase mapping is found
+    return cpt;  // Return the original code point if no lowercase mapping is found
 }
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {
-        { "\\p{N}", codepoint_flags::NUMBER },
-        { "\\p{L}", codepoint_flags::LETTER },
-        { "\\p{P}", codepoint_flags::PUNCTUATION },
+        { "\\p{N}", unicode_cpt_flags::NUMBER },
+        { "\\p{L}", unicode_cpt_flags::LETTER },
+        { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
     };
 
     static const std::map<int, int> k_ucat_cpt = {
-        { codepoint_flags::NUMBER,        0xD1 },
-        { codepoint_flags::LETTER,        0xD2 },
-        { codepoint_flags::PUNCTUATION,   0xD3 },
+        { unicode_cpt_flags::NUMBER,      0xD1 },
+        { unicode_cpt_flags::LETTER,      0xD2 },
+        { unicode_cpt_flags::PUNCTUATION, 0xD3 },
     };
 
     static const std::map<int, std::string> k_ucat_map = {
-        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
-        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
-        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+        { unicode_cpt_flags::NUMBER,      "\x30-\x39" }, // 0-9
+        { unicode_cpt_flags::LETTER,      "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
     };
 
     // compute collapsed codepoints only if needed by at least one regex
     bool need_collapse = false;
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // search for unicode categories
         for (const auto & ucat : k_ucat_enum) {
             if (std::string::npos != regex_expr.find(ucat.first)) {
@@ -698,7 +709,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const auto flags = unicode_cpt_flags(cpts[i]);
+            const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);
 
             if (flags.is_whitespace) {
                 //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
@@ -714,7 +725,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
 
     std::vector<size_t> bpe_offsets = { cpts.size() };
 
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // first, see if we have an efficient custom regex implementation
         auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
 
@@ -728,7 +739,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
             // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
             // with the corresponding collapsed representation
             bool use_collapsed = false;
-            for (auto & ucat : k_ucat_enum) {
+            for (const auto & ucat : k_ucat_enum) {
                 if (std::string::npos != regex_expr.find(ucat.first)) {
                     use_collapsed = true;
                     break;
@@ -794,7 +805,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
                 std::wstring wtext(cpts.begin(), cpts.end());
                 for (size_t i = 0; i < wtext.size(); ++i) {
-                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {
                         wtext[i] = 0x0B;
                     }
                 }

examples/talk-llama/unicode.h

@@ -4,9 +4,7 @@
 #include <string>
 #include <vector>
 
-// TODO: prefix all symbols with "llama_"
-
-struct codepoint_flags {
+struct unicode_cpt_flags {
     enum {
         UNDEFINED       = 0x0001,
         NUMBER          = 0x0002,  // regex: \p{N}
@@ -35,7 +33,7 @@ struct codepoint_flags {
     uint16_t is_nfd         : 1;
 
     // decode from uint16
-    inline codepoint_flags(const uint16_t flags=0) {
+    inline unicode_cpt_flags(const uint16_t flags = 0) {
         *reinterpret_cast<uint16_t*>(this) = flags;
     }
 
@@ -50,18 +48,19 @@ struct codepoint_flags {
 
 size_t unicode_len_utf8(char src);
 
-std::string unicode_cpt_to_utf8(uint32_t cp);
-uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+std::string unicode_cpt_to_utf8  (uint32_t cpt);
+uint32_t    unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp);
-codepoint_flags unicode_cpt_flags(const std::string & utf8);
+unicode_cpt_flags unicode_cpt_flags_from_cpt (uint32_t cpt);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8);
 
 std::string unicode_byte_to_utf8(uint8_t byte);
-uint8_t unicode_utf8_to_byte(const std::string & utf8);
+uint8_t     unicode_utf8_to_byte(const std::string & utf8);
 
-uint32_t unicode_tolower(uint32_t cp);
+uint32_t unicode_tolower(uint32_t cpt);
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -2,11 +2,11 @@ cmake_minimum_required(VERSION 3.10)
 
 project(whisper.cpp)
 
-set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_STANDARD 17)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../..)
 
 # Path to external GGML, otherwise uses the copy in whisper.cpp.
-option(GGML_HOME       "whisper: Path to external GGML source" OFF)
+option(GGML_HOME "whisper: Path to external GGML source" OFF)
 
 set(
     SOURCE_FILES
@@ -14,15 +14,24 @@ set(
     ${CMAKE_SOURCE_DIR}/jni.c
     )
 
+# TODO: this needs to be updated to work with the new ggml CMakeLists
+
 if (NOT GGML_HOME)
     set(
         SOURCE_FILES
         ${SOURCE_FILES}
         ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-aarch64.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-backend-reg.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-threading.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
         )
 endif()
 
@@ -80,3 +89,5 @@ include_directories(${WHISPER_LIB_DIR}/src)
 include_directories(${WHISPER_LIB_DIR}/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/src)
+include_directories(${WHISPER_LIB_DIR}/ggml/src/ggml-cpu)
+

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -24,6 +24,12 @@
 		18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.cpp */; };
 		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
+		18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */ = {isa = PBXBuildFile; fileRef = 18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */; };
+		18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */; };
+		18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */; };
+		18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */; };
+		18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */; };
+		18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
 		7FE3424C2A0C3FA20015A058 /* whisper-encoder.mm in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342472A0C3FA20015A058 /* whisper-encoder.mm */; };
 		7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */; };
@@ -49,8 +55,8 @@
 		18133C7F2C64E342005CEAAC /* ggml-aarch64.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-aarch64.c"; path = "../../../ggml/src/ggml-aarch64.c"; sourceTree = "<group>"; };
 		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml/src/ggml-alloc.c"; sourceTree = "<group>"; };
 		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml/include/ggml-alloc.h"; sourceTree = "<group>"; };
-		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal.m"; sourceTree = "<group>"; };
-		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal.metal"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal/ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal/ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@@ -76,6 +82,17 @@
 		18ABE1572AF556340044A204 /* ggml-backend.cpp */ = {isa = PBXFileReference; explicitFileType = sourcecode.cpp.cpp; fileEncoding = 4; name = "ggml-backend.cpp"; path = "../../../ggml/src/ggml-backend.cpp"; sourceTree = "<group>"; };
 		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
 		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
+		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.c"; sourceTree = "<group>"; };
+		18E864AA2CE73C580094B8B3 /* ggml-cpu.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu.h"; path = "../../../ggml/include/ggml-cpu.h"; sourceTree = "<group>"; };
+		18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-threading.h"; path = "../../../ggml/src/ggml-threading.h"; sourceTree = "<group>"; };
+		18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-threading.cpp"; path = "../../../ggml/src/ggml-threading.cpp"; sourceTree = "<group>"; };
+		18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-cpu.cpp"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.cpp"; sourceTree = "<group>"; };
+		18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-aarch64.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.h"; sourceTree = "<group>"; };
+		18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-aarch64.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.c"; sourceTree = "<group>"; };
+		18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-impl.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-impl.h"; sourceTree = "<group>"; };
+		18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-quants.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.h"; sourceTree = "<group>"; };
+		18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-quants.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.c"; sourceTree = "<group>"; };
+		18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-backend-reg.cpp"; path = "../../../ggml/src/ggml-backend-reg.cpp"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@@ -115,6 +132,17 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */,
+				18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */,
+				18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */,
+				18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */,
+				18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */,
+				18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */,
+				18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */,
+				18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */,
+				18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */,
+				18E864AA2CE73C580094B8B3 /* ggml-cpu.h */,
+				18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */,
 				18133C7F2C64E342005CEAAC /* ggml-aarch64.c */,
 				18133C7E2C64E342005CEAAC /* ggml-aarch64.h */,
 				18A275FF2C2A9563001C8D37 /* ggml-common.h */,
@@ -247,10 +275,16 @@
 				18627C9629052C5800BD2A04 /* ggml.c in Sources */,
 				18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
 				7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
+				18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */,
+				18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */,
 				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
+				18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */,
+				18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */,
+				18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */,
 				18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */,
 				18627C8C29052BE000BD2A04 /* main.m in Sources */,
 				18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
+				18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */,
 				1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,
 				7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */,
 			);
@@ -329,6 +363,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
 				MTL_FAST_MATH = YES;
@@ -382,6 +417,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
@@ -404,6 +440,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
@@ -433,6 +470,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;

examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift

@@ -1,4 +1,5 @@
 import Foundation
+import UIKit
 import whisper
 
 enum WhisperError: Error {
@@ -55,11 +56,91 @@ actor WhisperContext {
         return transcription
     }
 
+    static func benchMemcpy(nThreads: Int32) async -> String {
+        return String.init(cString: whisper_bench_memcpy_str(nThreads))
+    }
+
+    static func benchGgmlMulMat(nThreads: Int32) async -> String {
+        return String.init(cString: whisper_bench_ggml_mul_mat_str(nThreads))
+    }
+
+    private func systemInfo() -> String {
+        var info = ""
+        //if (ggml_cpu_has_neon() != 0) { info += "NEON " }
+        return String(info.dropLast())
+    }
+
+    func benchFull(modelName: String, nThreads: Int32) async -> String {
+        let nMels = whisper_model_n_mels(context)
+        if (whisper_set_mel(context, nil, 0, nMels) != 0) {
+            return "error: failed to set mel"
+        }
+
+        // heat encoder
+        if (whisper_encode(context, 0, nThreads) != 0) {
+            return "error: failed to encode"
+        }
+
+        var tokens = [whisper_token](repeating: 0, count: 512)
+
+        // prompt heat
+        if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
+            return "error: failed to decode"
+        }
+
+        // text-generation heat
+        if (whisper_decode(context, &tokens, 1, 256, nThreads) != 0) {
+            return "error: failed to decode"
+        }
+
+        whisper_reset_timings(context)
+
+        // actual run
+        if (whisper_encode(context, 0, nThreads) != 0) {
+            return "error: failed to encode"
+        }
+
+        // text-generation
+        for i in 0..<256 {
+            if (whisper_decode(context, &tokens, 1, Int32(i), nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        // batched decoding
+        for _ in 0..<64 {
+            if (whisper_decode(context, &tokens, 5, 0, nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        // prompt processing
+        for _ in 0..<16 {
+            if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
+                return "error: failed to decode"
+            }
+        }
+
+        whisper_print_timings(context)
+
+        let deviceModel = await UIDevice.current.model
+        let systemName = await UIDevice.current.systemName
+        let systemInfo = self.systemInfo()
+        let timings: whisper_timings = whisper_get_timings(context).pointee
+        let encodeMs = String(format: "%.2f", timings.encode_ms)
+        let decodeMs = String(format: "%.2f", timings.decode_ms)
+        let batchdMs = String(format: "%.2f", timings.batchd_ms)
+        let promptMs = String(format: "%.2f", timings.prompt_ms)
+        return "| \(deviceModel) | \(systemName) | \(systemInfo) | \(modelName) | \(nThreads) | 1 | \(encodeMs) | \(decodeMs) | \(batchdMs) | \(promptMs) | <todo> |"
+    }
+
     static func createContext(path: String) throws -> WhisperContext {
         var params = whisper_context_default_params()
 #if targetEnvironment(simulator)
         params.use_gpu = false
         print("Running on the simulator, using CPU")
+#else
+        params.flash_attn = true // Enabled by default for Metal
 #endif
         let context = whisper_init_from_file_with_params(path, params)
         if let context {

examples/whisper.swiftui/whisper.swiftui.demo/Models/Model.swift

@@ -0,0 +1,17 @@
+import Foundation
+
+struct Model: Identifiable {
+    var id = UUID()
+    var name: String
+    var info: String
+    var url: String
+
+    var filename: String
+    var fileURL: URL {
+        FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0].appendingPathComponent(filename)
+    }
+
+    func fileExists() -> Bool {
+        FileManager.default.fileExists(atPath: fileURL.path)
+    }
+}

examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift

@@ -14,7 +14,7 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     private var recordedFile: URL? = nil
     private var audioPlayer: AVAudioPlayer?
     
-    private var modelUrl: URL? {
+    private var builtInModelUrl: URL? {
         Bundle.main.url(forResource: "ggml-base.en", withExtension: "bin", subdirectory: "models")
     }
     
@@ -28,23 +28,59 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     
     override init() {
         super.init()
+        loadModel()
+    }
+    
+    func loadModel(path: URL? = nil, log: Bool = true) {
         do {
-            try loadModel()
+            whisperContext = nil
+            if (log) { messageLog += "Loading model...\n" }
+            let modelUrl = path ?? builtInModelUrl
+            if let modelUrl {
+                whisperContext = try WhisperContext.createContext(path: modelUrl.path())
+                if (log) { messageLog += "Loaded model \(modelUrl.lastPathComponent)\n" }
+            } else {
+                if (log) { messageLog += "Could not locate model\n" }
+            }
             canTranscribe = true
         } catch {
             print(error.localizedDescription)
-            messageLog += "\(error.localizedDescription)\n"
+            if (log) { messageLog += "\(error.localizedDescription)\n" }
         }
     }
-    
-    private func loadModel() throws {
-        messageLog += "Loading model...\n"
-        if let modelUrl {
-            whisperContext = try WhisperContext.createContext(path: modelUrl.path())
-            messageLog += "Loaded model \(modelUrl.lastPathComponent)\n"
-        } else {
-            messageLog += "Could not locate model\n"
+
+    func benchCurrentModel() async {
+        if whisperContext == nil {
+            messageLog += "Cannot bench without loaded model\n"
+            return
         }
+        messageLog += "Running benchmark for loaded model\n"
+        let result = await whisperContext?.benchFull(modelName: "<current>", nThreads: Int32(min(4, cpuCount())))
+        if (result != nil) { messageLog += result! + "\n" }
+    }
+
+    func bench(models: [Model]) async {
+        let nThreads = Int32(min(4, cpuCount()))
+
+//        messageLog += "Running memcpy benchmark\n"
+//        messageLog += await WhisperContext.benchMemcpy(nThreads: nThreads) + "\n"
+//
+//        messageLog += "Running ggml_mul_mat benchmark with \(nThreads) threads\n"
+//        messageLog += await WhisperContext.benchGgmlMulMat(nThreads: nThreads) + "\n"
+
+        messageLog += "Running benchmark for all downloaded models\n"
+        messageLog += "| CPU | OS | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |\n"
+        messageLog += "| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n"
+        for model in models {
+            loadModel(path: model.fileURL, log: false)
+            if whisperContext == nil {
+                messageLog += "Cannot bench without loaded model\n"
+                break
+            }
+            let result = await whisperContext?.benchFull(modelName: model.name, nThreads: nThreads)
+            if (result != nil) { messageLog += result! + "\n" }
+        }
+        messageLog += "Benchmarking completed\n"
     }
     
     func transcribeSample() async {
@@ -160,3 +196,8 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
         isRecording = false
     }
 }
+
+
+fileprivate func cpuCount() -> Int {
+    ProcessInfo.processInfo.processorCount
+}

examples/whisper.swiftui/whisper.swiftui.demo/UI/ContentView.swift

@@ -1,5 +1,6 @@
 import SwiftUI
 import AVFoundation
+import Foundation
 
 struct ContentView: View {
     @StateObject var whisperState = WhisperState()
@@ -29,15 +30,125 @@ struct ContentView: View {
                     Text(verbatim: whisperState.messageLog)
                         .frame(maxWidth: .infinity, alignment: .leading)
                 }
+                .font(.footnote)
+                .padding()
+                .background(Color.gray.opacity(0.1))
+                .cornerRadius(10)
+
+                HStack {
+                    Button("Clear Logs", action: {
+                        whisperState.messageLog = ""
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Copy Logs", action: {
+                        UIPasteboard.general.string = whisperState.messageLog
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+
+                    Button("Bench", action: {
+                        Task {
+                            await whisperState.benchCurrentModel()
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+
+                    Button("Bench All", action: {
+                        Task {
+                            await whisperState.bench(models: ModelsView.getDownloadedModels())
+                        }
+                    })
+                    .font(.footnote)
+                    .buttonStyle(.bordered)
+                    .disabled(!whisperState.canTranscribe)
+                }
+
+                NavigationLink(destination: ModelsView(whisperState: whisperState)) {
+                    Text("View Models")
+                }
+                .font(.footnote)
+                .padding()
             }
             .navigationTitle("Whisper SwiftUI Demo")
             .padding()
         }
     }
-}
 
-struct ContentView_Previews: PreviewProvider {
-    static var previews: some View {
-        ContentView()
+    struct ModelsView: View {
+        @ObservedObject var whisperState: WhisperState
+        @Environment(\.dismiss) var dismiss
+        
+        private static let models: [Model] = [
+            Model(name: "tiny", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin", filename: "tiny.bin"),
+            Model(name: "tiny-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q5_1.bin", filename: "tiny-q5_1.bin"),
+            Model(name: "tiny-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny-q8_0.bin", filename: "tiny-q8_0.bin"),
+            Model(name: "tiny.en", info: "(F16, 75 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en.bin", filename: "tiny.en.bin"),
+            Model(name: "tiny.en-q5_1", info: "(31 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q5_1.bin", filename: "tiny.en-q5_1.bin"),
+            Model(name: "tiny.en-q8_0", info: "(42 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q8_0.bin", filename: "tiny.en-q8_0.bin"),
+            Model(name: "base", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin", filename: "base.bin"),
+            Model(name: "base-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q5_1.bin", filename: "base-q5_1.bin"),
+            Model(name: "base-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base-q8_0.bin", filename: "base-q8_0.bin"),
+            Model(name: "base.en", info: "(F16, 142 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin", filename: "base.en.bin"),
+            Model(name: "base.en-q5_1", info: "(57 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q5_1.bin", filename: "base.en-q5_1.bin"),
+            Model(name: "base.en-q8_0", info: "(78 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en-q8_0.bin", filename: "base.en-q8_0.bin"),
+            Model(name: "small", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin", filename: "small.bin"),
+            Model(name: "small-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q5_1.bin", filename: "small-q5_1.bin"),
+            Model(name: "small-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small-q8_0.bin", filename: "small-q8_0.bin"),
+            Model(name: "small.en", info: "(F16, 466 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.bin", filename: "small.en.bin"),
+            Model(name: "small.en-q5_1", info: "(181 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q5_1.bin", filename: "small.en-q5_1.bin"),
+            Model(name: "small.en-q8_0", info: "(252 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en-q8_0.bin", filename: "small.en-q8_0.bin"),
+            Model(name: "medium", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.bin", filename: "medium.bin"),
+            Model(name: "medium-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q5_0.bin", filename: "medium-q5_0.bin"),
+            Model(name: "medium-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium-q8_0.bin", filename: "medium-q8_0.bin"),
+            Model(name: "medium.en", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en.bin", filename: "medium.en.bin"),
+            Model(name: "medium.en-q5_0", info: "(514 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q5_0.bin", filename: "medium.en-q5_0.bin"),
+            Model(name: "medium.en-q8_0", info: "(785 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en-q8_0.bin", filename: "medium.en-q8_0.bin"),
+            Model(name: "large-v1", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large.bin", filename: "large.bin"),
+            Model(name: "large-v2", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2.bin", filename: "large-v2.bin"),
+            Model(name: "large-v2-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q5_0.bin", filename: "large-v2-q5_0.bin"),
+            Model(name: "large-v2-q8_0", info: "(1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v2-q8_0.bin", filename: "large-v2-q8_0.bin"),
+            Model(name: "large-v3", info: "(F16, 2.9 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin", filename: "large-v3.bin"),
+            Model(name: "large-v3-q5_0", info: "(1.1 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-q5_0.bin", filename: "large-v3-q5_0.bin"),
+            Model(name: "large-v3-turbo", info: "(F16, 1.5 GiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin", filename: "large-v3-turbo.bin"),
+            Model(name: "large-v3-turbo-q5_0", info: "(547 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin", filename: "large-v3-turbo-q5_0.bin"),
+            Model(name: "large-v3-turbo-q8_0", info: "(834 MiB)", url: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q8_0.bin", filename: "large-v3-turbo-q8_0.bin"),
+        ]
+
+        static func getDownloadedModels() -> [Model] {
+            // Filter models that have been downloaded
+            return models.filter {
+                FileManager.default.fileExists(atPath: $0.fileURL.path())
+            }
+        }
+
+        func loadModel(model: Model) {
+            Task {
+                dismiss()
+                whisperState.loadModel(path: model.fileURL)
+            }
+        }
+
+        var body: some View {
+            List {
+                Section(header: Text("Models")) {
+                    ForEach(ModelsView.models) { model in
+                        DownloadButton(model: model)
+                            .onLoad(perform: loadModel)
+                    }
+                }
+            }
+            .listStyle(GroupedListStyle())
+            .navigationBarTitle("Models", displayMode: .inline).toolbar {}
+        }
     }
 }
+
+//struct ContentView_Previews: PreviewProvider {
+//    static var previews: some View {
+//        ContentView()
+//    }
+//}

examples/whisper.swiftui/whisper.swiftui.demo/UI/DownloadButton.swift

@@ -0,0 +1,102 @@
+import SwiftUI
+
+struct DownloadButton: View {
+    private var model: Model
+
+    @State private var status: String
+
+    @State private var downloadTask: URLSessionDownloadTask?
+    @State private var progress = 0.0
+    @State private var observation: NSKeyValueObservation?
+
+    private var onLoad: ((_ model: Model) -> Void)?
+
+    init(model: Model) {
+        self.model = model
+        status = model.fileExists() ? "downloaded" : "download"
+    }
+
+    func onLoad(perform action: @escaping (_ model: Model) -> Void) -> DownloadButton {
+        var button = self
+        button.onLoad = action
+        return button
+    }
+
+    private func download() {
+        status = "downloading"
+        print("Downloading model \(model.name) from \(model.url)")
+        guard let url = URL(string: model.url) else { return }
+
+        downloadTask = URLSession.shared.downloadTask(with: url) { temporaryURL, response, error in
+            if let error = error {
+                print("Error: \(error.localizedDescription)")
+                return
+            }
+
+            guard let response = response as? HTTPURLResponse, (200...299).contains(response.statusCode) else {
+                print("Server error!")
+                return
+            }
+
+            do {
+                if let temporaryURL = temporaryURL {
+                    try FileManager.default.copyItem(at: temporaryURL, to: model.fileURL)
+                    print("Writing to \(model.filename) completed")
+                    status = "downloaded"
+                }
+            } catch let err {
+                print("Error: \(err.localizedDescription)")
+            }
+        }
+
+        observation = downloadTask?.progress.observe(\.fractionCompleted) { progress, _ in
+            self.progress = progress.fractionCompleted
+        }
+
+        downloadTask?.resume()
+    }
+
+    var body: some View {
+        VStack {
+            Button(action: {
+                if (status == "download") {
+                    download()
+                } else if (status == "downloading") {
+                    downloadTask?.cancel()
+                    status = "download"
+                } else if (status == "downloaded") {
+                    if !model.fileExists() {
+                        download()
+                    }
+                    onLoad?(model)
+                }
+            }) {
+                let title = "\(model.name) \(model.info)"
+                if (status == "download") {
+                    Text("Download \(title)")
+                } else if (status == "downloading") {
+                    Text("\(title) (Downloading \(Int(progress * 100))%)")
+                } else if (status == "downloaded") {
+                    Text("Load \(title)")
+                } else {
+                    Text("Unknown status")
+                }
+            }.swipeActions {
+                if (status == "downloaded") {
+                    Button("Delete") {
+                        do {
+                            try FileManager.default.removeItem(at: model.fileURL)
+                        } catch {
+                            print("Error deleting file: \(error)")
+                        }
+                        status = "download"
+                    }
+                    .tint(.red)
+                }
+            }
+        }
+        .onDisappear() {
+            downloadTask?.cancel()
+        }
+    }
+}

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -17,6 +17,8 @@
 		0AAC5D9F29539CD0003032C3 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 0AAC5D9E29539CD0003032C3 /* Assets.xcassets */; };
 		0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
 		0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
+		7F79E0EE2CE0A78000ACD7BF /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */; };
+		7F79E0F02CE0C6F700ACD7BF /* Model.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7F79E0EF2CE0C6F700ACD7BF /* Model.swift */; };
 		E3F92DC52AFA8E3800A6A9D4 /* whisper in Frameworks */ = {isa = PBXBuildFile; productRef = E3F92DC42AFA8E3800A6A9D4 /* whisper */; };
 /* End PBXBuildFile section */
 
@@ -33,6 +35,8 @@
 		0AAC5DA029539CD0003032C3 /* WhisperCppDemo.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = WhisperCppDemo.entitlements; sourceTree = "<group>"; };
 		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
 		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
+		7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
+		7F79E0EF2CE0C6F700ACD7BF /* Model.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Model.swift; sourceTree = "<group>"; };
 		E3F92DC22AFA8DD800A6A9D4 /* whisper.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = whisper.cpp; path = ../..; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
@@ -52,6 +56,7 @@
 			isa = PBXGroup;
 			children = (
 				0AAC5DCD2953A05C003032C3 /* WhisperState.swift */,
+				7F79E0EF2CE0C6F700ACD7BF /* Model.swift */,
 			);
 			path = Models;
 			sourceTree = "<group>";
@@ -119,6 +124,7 @@
 			isa = PBXGroup;
 			children = (
 				0AAC5D9C29539CCF003032C3 /* ContentView.swift */,
+				7F79E0ED2CE0A78000ACD7BF /* DownloadButton.swift */,
 			);
 			path = UI;
 			sourceTree = "<group>";
@@ -220,7 +226,9 @@
 				0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */,
 				0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */,
 				0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
+				7F79E0EE2CE0A78000ACD7BF /* DownloadButton.swift in Sources */,
 				0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
+				7F79E0F02CE0C6F700ACD7BF /* Model.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -370,7 +378,9 @@
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
-				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator";
+				SUPPORTS_MACCATALYST = NO;
+				SUPPORTS_MAC_DESIGNED_FOR_IPHONE_IPAD = YES;
 				SWIFT_EMIT_LOC_STRINGS = YES;
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
@@ -415,7 +425,9 @@
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;
-				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator macosx";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator";
+				SUPPORTS_MACCATALYST = NO;
+				SUPPORTS_MAC_DESIGNED_FOR_IPHONE_IPAD = YES;
 				SWIFT_EMIT_LOC_STRINGS = YES;
 				SWIFT_VERSION = 5.0;
 				TARGETED_DEVICE_FAMILY = "1,2";

ggml/CMakeLists.txt

@@ -32,7 +32,15 @@ else()
     endif()
 endif()
 
+# remove the lib prefix on win32 mingw
+if (WIN32)
+    set(CMAKE_STATIC_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_MODULE_PREFIX  "")
+endif()
+
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 
 #
 # option list
@@ -91,28 +99,38 @@ else()
     set(INS_ENB ON)
 endif()
 
-option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
-
-option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
-option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
-option(GGML_AVX512      "ggml: enable AVX512"           OFF)
-option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
-option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
-option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
-option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
+option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
+option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512           "ggml: enable AVX512F"          OFF)
+option(GGML_AVX512_VBMI      "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI      "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16      "ggml: enable AVX512-BF16"      OFF)
 if (NOT MSVC)
-    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+    # in MSVC F16C and FMA is implied with AVX2/AVX512
+    option(GGML_FMA          "ggml: enable FMA"              ${INS_ENB})
+    option(GGML_F16C         "ggml: enable F16C"             ${INS_ENB})
+    # MSVC does not seem to support AMX
+    option(GGML_AMX_TILE     "ggml: enable AMX-TILE"         OFF)
+    option(GGML_AMX_INT8     "ggml: enable AMX-INT8"         OFF)
+    option(GGML_AMX_BF16     "ggml: enable AMX-BF16"         OFF)
 endif()
-option(GGML_LASX        "ggml: enable lasx"             ON)
-option(GGML_LSX         "ggml: enable lsx"              ON)
-option(GGML_SVE         "ggml: enable SVE"              OFF)
+option(GGML_LASX             "ggml: enable lasx"             ON)
+option(GGML_LSX              "ggml: enable lsx"              ON)
+option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_SVE              "ggml: enable SVE"              OFF)
+option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
+
 
 if (WIN32)
-    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+    set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
@@ -123,14 +141,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -138,7 +151,7 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
-option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
@@ -150,6 +163,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
 option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
 option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
@@ -162,6 +176,13 @@ option(GGML_SYCL                            "ggml: use SYCL"
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
+set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
+                                            "ggml: sycl device architecture")
+
+option(GGML_OPENCL                          "ggml: use OpenCL"                                OFF)
+option(GGML_OPENCL_PROFILING                "ggml: use OpenCL profiling (increases overhead)" OFF)
+option(GGML_OPENCL_EMBED_KERNELS            "ggml: embed kernels"                             ON)
+option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adreno"          ON)
 
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
@@ -174,11 +195,7 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 
-if (GGML_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
+set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 
 set(THREADS_PREFER_PTHREAD_FLAG ON)
@@ -214,13 +231,14 @@ include(CMakePackageConfigHelpers)
 # all public headers
 set(GGML_PUBLIC_HEADERS
     include/ggml.h
+    include/ggml-cpu.h
     include/ggml-alloc.h
     include/ggml-backend.h
     include/ggml-blas.h
     include/ggml-cann.h
     include/ggml-cuda.h
-    include/ggml.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -230,15 +248,14 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml PUBLIC_HEADER)
-
-if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml LIBRARY)
-endif()
+install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml-base LIBRARY)
 
+# FIXME: this should be done in the backend cmake files
 if (GGML_METAL)
+    # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
     install(
-        FILES src/ggml-metal.metal
+        FILES src/ggml-metal/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE

ggml/ggml_vk_generate_shaders.py

@@ -1,220 +0,0 @@
-#!/usr/bin/env python
-
-import logging
-import argparse
-import asyncio
-import os
-from tempfile import gettempdir
-
-logger = logging.getLogger("ggml-vk-generate-shaders")
-
-GLSLC = "glslc"
-
-type_names = [
-    "f32",
-    "f16",
-    "q4_0",
-    "q4_1",
-    "q5_0",
-    "q5_1",
-    "q8_0",
-    "q2_k",
-    "q3_k",
-    "q4_k",
-    "q5_k",
-    "q6_k",
-]
-
-ASYNCIO_CONCURRENCY = 64
-
-input_dir = "vulkan-shaders"
-output_dir = gettempdir()
-
-lock = asyncio.Lock()
-shader_fnames = []
-
-
-async def string_to_spv(name, in_fname, defines, fp16=True):
-    name = f"{name}{'_fp32' if not fp16 else ''}"
-    out_fname = os.path.join(output_dir, f"{name}.spv")
-
-    in_path = os.path.join(input_dir, in_fname)
-
-    cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname]
-
-    cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
-
-    proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
-
-    stdout, stderr = await proc.communicate()
-
-    stdout = stdout.decode()
-    error = stderr.decode()
-
-    if proc.returncode:
-        cmd = " ".join(cmd)
-        logger.error(f"cannot compile {name}\n\n{cmd}\n\n{error}")
-        return
-
-    async with lock:
-        shader_fnames.append((name, out_fname))
-
-
-def matmul_shaders(tasks, fp16, matmul_id):
-    if fp16:
-        load_vec = "8"
-        aligned_b_type_f32 = "mat2x4"
-        aligned_b_type_f16 = "f16mat2x4"
-    else:
-        load_vec = "4"
-        aligned_b_type_f32 = "vec4"
-        aligned_b_type_f16 = "f16vec4"
-
-    base_dict = {"FLOAT_TYPE": "float" if not fp16 else "float16_t"}
-    shader_name = "matmul"
-
-    if matmul_id:
-        base_dict["MUL_MAT_ID"] = "1"
-        shader_name = "matmul_id"
-
-    if fp16:
-        base_dict["FLOAT16"] = "1"
-
-    # Shaders with f16 B_TYPE
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    tasks.append(string_to_spv(f"{shader_name}_f16", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    for tname in type_names:
-        data_a_key = f"DATA_A_{tname.upper()}"
-        load_vec_a = load_vec if tname in ("f32", "f16") else "2"
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32", "mul_mm.comp", base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32_aligned", "mul_mm.comp", base_dict | {data_a_key: "2", "LOAD_VEC_A": load_vec_a, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f32, "D_TYPE": "float"}, fp16))
-
-
-async def main():
-    logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
-
-    tasks = []
-
-    for fp16 in (False, True):
-        # MUL_MAT
-        matmul_shaders(tasks, fp16, False)
-        # MUL_MAT_ID
-        matmul_shaders(tasks, fp16, True)
-
-    for tname in type_names:
-        base_dict = {"FLOAT_TYPE": "float"}
-
-        # mul mat vec
-        data_a_key = f"DATA_A_{tname.upper()}"
-        shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp"
-
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f32_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f16_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-        tasks.append(string_to_spv(f"mul_mat_vec_id_{tname}_f32", shader, base_dict | {"MUL_MAT_ID": "1", data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-
-        # Dequant shaders
-        if tname != "f16":
-            tasks.append(string_to_spv(f"dequant_{tname}", f"dequant_{tname}.comp", base_dict | {data_a_key: "1", "D_TYPE": "float16_t"}))
-
-        # get_rows
-        if not tname.endswith("_k"):
-            shader = "get_rows.comp" if tname in ("f32", "f16") else "get_rows_quant.comp"
-
-            if tname == "f16":
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-            else:
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t"}))
-            tasks.append(string_to_spv(f"get_rows_{tname}_f32", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-
-    # Norms
-    tasks.append(string_to_spv("norm_f32", "norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rms_norm_f32", "rms_norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("cpy_f32_f32", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("cpy_f32_f16", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
-    tasks.append(string_to_spv("cpy_f16_f16", "copy.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-
-    tasks.append(string_to_spv("add_f32", "add.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}))
-
-    tasks.append(string_to_spv("mul_f32", "mul.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("div_f32", "div.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("scale_f32", "scale.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sqr_f32", "square.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("clamp_f32", "clamp.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("gelu_f32", "gelu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("silu_f32", "silu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("relu_f32", "relu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("soft_max_f32", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("soft_max_f32_f16", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("rope_norm_f32", "rope_norm.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_norm_f16", "rope_norm.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("rope_neox_f32", "rope_neox.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_neox_f16", "rope_neox.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("argsort_f32", "argsort.comp", {"A_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sum_rows_f32", "sum_rows.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    # Helper to decorate tasks with semaphore acquisition.
-    async def withSemaphore(sem, task):
-        async with sem:
-            return await task
-
-    # Run tasks concurrently guarded by a concurrency limit.
-    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
-    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
-
-    with open("ggml-vulkan-shaders.hpp", "w") as f:
-        f.write("#include <cstdint>\n\n")
-        for name, path in sorted(shader_fnames):
-
-            with open(path, "rb") as spv:
-                counter = 0
-                newline_counter = 0
-                f.write(f"unsigned char {name}_data[] = {{\n")
-                for val in spv.read():
-                    f.write(f"0x{val:02x},")
-                    newline_counter += 1
-                    counter += 1
-                    if newline_counter >= 12:
-                        newline_counter = 0
-                        f.write("\n")
-            f.write("\n};\n")
-            f.write(f"const uint64_t {name}_len = {counter};\n\n")
-            os.remove(path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
-
-    parser.add_argument("--glslc", help="Path to glslc")
-    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-
-    args = parser.parse_args()
-
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-
-    if args.glslc:
-        GLSLC = args.glslc
-
-    asyncio.run(main())

ggml/include/ggml-backend.h

@@ -3,6 +3,20 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
 
+#ifdef GGML_BACKEND_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef GGML_BACKEND_BUILD
+#            define GGML_BACKEND_API __declspec(dllexport) extern
+#        else
+#            define GGML_BACKEND_API __declspec(dllimport) extern
+#        endif
+#    else
+#        define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
+#    endif
+#else
+#    define GGML_BACKEND_API extern
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -72,7 +86,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    // "offset" refers to the offset of the tensor data for setting/getting data
+    // "offset" refers to the offset in tensor->data for setting/getting data
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
@@ -114,11 +128,12 @@ extern "C" {
     //
 
     enum ggml_backend_dev_type {
+        // CPU device using system memory
         GGML_BACKEND_DEVICE_TYPE_CPU,
+        // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
-        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
-        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
-        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
+        // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
+        GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
 
     // functionality supported by the device
@@ -127,6 +142,8 @@ extern "C" {
         bool async;
         // pinned host buffer
         bool host_buffer;
+        // creating buffers from host ptr
+        bool buffer_from_host_ptr;
         // event synchronization
         bool events;
     };
@@ -165,9 +182,22 @@ extern "C" {
     GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
     GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
 
+    // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // Functions that may be obtained using ggml_backend_reg_get_proc_address
-    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
+    // Split buffer type for tensor parallelism
+    typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
+    // Set the number of threads for the backend
+    typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
+    // Get additional buffer types provided by the device (returns a NULL-terminated array)
+    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    // Set the abort callback for the backend
+    typedef void                         (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data);
+    // Get a list of feature flags supported by the backend (returns a NULL-terminated array)
+    struct ggml_backend_feature {
+        const char * name;
+        const char * value;
+    };
+    typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg);
 
     //
     // Backend registry
@@ -189,9 +219,17 @@ extern "C" {
     GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
     GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
-    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
+    // Load a backend from a dynamic library and register it
+    GGML_API ggml_backend_reg_t ggml_backend_load(const char * path);
+    // Unload a backend if loaded dynamically and unregister it
+    GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
+    // Load all known backends from dynamic libraries
+    GGML_API void               ggml_backend_load_all(void);
+    GGML_API void               ggml_backend_load_all_from_path(const char * dir_path);
+
     //
     // Backend scheduler
     //
@@ -220,14 +258,20 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
+        for (int i = 0; i < 10; ++i) {
+            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
+        }
 
         // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
+        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
+        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
+        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph); // execute the graph
+
+        // as an alternative to the above it is also possible to assign the inputs to a dedicated context and
+        // allocate them statically via ggml_backend_alloc_ctx_tensors
     }
     */
 
@@ -242,7 +286,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler
+    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -267,7 +311,9 @@ extern "C" {
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
-    // Reset all assignments and allocators - must be called before changing the node backends
+    // Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
+    // This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
+    // The correct way to use this API is to discard the deallocated tensors and create new ones.
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
@@ -297,27 +343,10 @@ extern "C" {
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
+    // CPU buffer types are always available
     GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
-    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-blas.h

@@ -9,13 +9,15 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_blas_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
 
 // number of threads used for conversion to float
 // for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-cann.h

@@ -34,6 +34,8 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+
 /**
  * @brief Initializes the CANN backend for a specified device.
  *
@@ -44,7 +46,7 @@ extern "C" {
  * @param device The index of the device to initialize.
  * @return A pointer to the initialized backend instance, or nullptr on failure.
  */
-GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
 
 /**
  * @brief Checks if a given backend is a CANN backend.
@@ -55,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
  * @param backend The backend instance to check.
  * @return True if the backend is a CANN backend, false otherwise.
  */
-GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
 
 /**
  * @brief Retrieves the CANN buffer type for a specified device.
@@ -67,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_API ggml_backend_buffer_type_t
+GGML_BACKEND_API ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device);
 
 /**
@@ -78,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
  *
  * @return The number of CANN devices available.
  */
-GGML_API int32_t ggml_backend_cann_get_device_count(void);
+GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
 
 /**
  * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
  *
  * @return A pointer to the host buffer type interface.
  */
-GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
 
 /**
  * @brief Retrieves the description of a specific CANN device.
@@ -97,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_API void ggml_backend_cann_get_device_description(
+GGML_BACKEND_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -112,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
                                                   size_t* free,
                                                   size_t* total);
 

ggml/include/ggml-cpp.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include <memory>
+
+// Smart pointers for ggml types
+
+// ggml
+
+struct ggml_context_deleter { void operator()(ggml_context * ctx) { ggml_free(ctx); } };
+struct gguf_context_deleter { void operator()(gguf_context * ctx) { gguf_free(ctx); } };
+
+typedef std::unique_ptr<ggml_context, ggml_context_deleter> ggml_context_ptr;
+typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
+
+// ggml-alloc
+
+struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
+
+typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+
+// ggml-backend
+
+struct ggml_backend_deleter        { void operator()(ggml_backend_t backend)       { ggml_backend_free(backend); } };
+struct ggml_backend_buffer_deleter { void operator()(ggml_backend_buffer_t buffer) { ggml_backend_buffer_free(buffer); } };
+struct ggml_backend_event_deleter  { void operator()(ggml_backend_event_t event)   { ggml_backend_event_free(event); } };
+struct ggml_backend_sched_deleter  { void operator()(ggml_backend_sched_t sched)   { ggml_backend_sched_free(sched); } };
+
+typedef std::unique_ptr<ggml_backend,        ggml_backend_deleter>        ggml_backend_ptr;
+typedef std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> ggml_backend_buffer_ptr;
+typedef std::unique_ptr<ggml_backend_event,  ggml_backend_event_deleter>  ggml_backend_event_ptr;
+typedef std::unique_ptr<ggml_backend_sched,  ggml_backend_sched_deleter>  ggml_backend_sched_ptr;

ggml/include/ggml-cpu.h

@@ -0,0 +1,135 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    // the compute plan that needs to be prepared for ggml_graph_compute()
+    // since https://github.com/ggerganov/ggml/issues/287
+    struct ggml_cplan {
+        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
+        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
+
+        int n_threads;
+        struct ggml_threadpool * threadpool;
+
+        // abort ggml_graph_compute when true
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
+    };
+
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
+    GGML_BACKEND_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_BACKEND_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+    GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+
+    GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+    GGML_BACKEND_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+    GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new           (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free          (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume        (struct ggml_threadpool * threadpool);
+
+    // ggml_graph_plan() has to be called before ggml_graph_compute()
+    // when plan.work_size > 0, caller must allocate memory for plan.work_data
+    GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
+    // same as ggml_graph_compute() but the work data is allocated as a part of the context
+    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+
+    //
+    // system info
+    //
+
+    // x86
+    GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
+    GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
+    GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
+    GGML_BACKEND_API int ggml_cpu_has_fma        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_BACKEND_API int ggml_cpu_has_amx_int8   (void);
+    // ARM
+    GGML_BACKEND_API int ggml_cpu_has_neon       (void);
+    GGML_BACKEND_API int ggml_cpu_has_arm_fma    (void);
+    GGML_BACKEND_API int ggml_cpu_has_fp16_va    (void);
+    GGML_BACKEND_API int ggml_cpu_has_dotprod    (void);
+    GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
+    GGML_BACKEND_API int ggml_cpu_has_sve        (void);
+    GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
+    // other
+    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
+    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
+    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
+
+    // Internal types and functions exposed for tests and benchmarks
+
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+
+    struct ggml_type_traits_cpu {
+        ggml_from_float_t        from_float;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+    };
+
+    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+
+    GGML_BACKEND_API void ggml_cpu_init(void);
+
+    //
+    // CPU backend
+    //
+
+    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_BACKEND_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml-cuda.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#ifdef GGML_USE_HIPBLAS
+#ifdef GGML_USE_HIP
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
 #elif defined(GGML_USE_MUSA)
@@ -20,27 +20,27 @@ extern "C" {
 #define GGML_CUDA_MAX_DEVICES       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_API int  ggml_backend_cuda_get_device_count(void);
-GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-kompute.h

@@ -11,6 +11,8 @@
 extern "C" {
 #endif
 
+#define GGML_KOMPUTE_MAX_DEVICES 16
+
 struct ggml_vk_device {
     int index;
     int type; // same as VkPhysicalDeviceType
@@ -35,11 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
 // forward declaration
 typedef struct ggml_backend * ggml_backend_t;
 
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
 
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-metal.h

@@ -39,23 +39,27 @@ extern "C" {
 // user-code should use only these functions
 //
 
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+GGML_DEPRECATED(
+        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
 
-GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
 #ifdef __cplusplus
 }

ggml/include/ggml-opencl.h

@@ -0,0 +1,26 @@
+#ifndef GGML_OPENCL_H
+#define GGML_OPENCL_H
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+//
+GGML_BACKEND_API ggml_backend_t ggml_backend_opencl_init(void);
+GGML_BACKEND_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_opencl_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif
+
+#endif // GGML_OPENCL_H

ggml/include/ggml-opt.h

@@ -0,0 +1,216 @@
+// This file contains functionality for training models using GGML.
+// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
+// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
+//
+// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdint.h>
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    struct ggml_opt_dataset;
+    struct ggml_opt_context;
+    struct ggml_opt_result;
+
+    typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
+    typedef struct ggml_opt_context * ggml_opt_context_t;
+    typedef struct ggml_opt_result  * ggml_opt_result_t;
+
+    // ====== Loss ======
+
+    // built-in loss types, i.e. the built-in quantities minimized by the optimizer
+    // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
+    enum ggml_opt_loss_type {
+        GGML_OPT_LOSS_TYPE_MEAN,
+        GGML_OPT_LOSS_TYPE_SUM,
+        GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
+        GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+    };
+
+    // ====== Dataset ======
+
+    GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
+            int64_t ne_datapoint, // number of elements per datapoint
+            int64_t ne_label,     // number of elements per label
+            int64_t ndata,        // total number of datapoints/labels
+            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+    GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
+
+    // get underlying tensors that store the data
+    GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
+    GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
+
+    // shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
+    GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
+
+    // get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
+    GGML_API void ggml_opt_dataset_get_batch(
+            ggml_opt_dataset_t   dataset,
+            struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
+            struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
+            int64_t              ibatch);
+
+    // ====== Model / Context ======
+
+    enum ggml_opt_build_type {
+        GGML_OPT_BUILD_TYPE_FORWARD,
+        GGML_OPT_BUILD_TYPE_GRAD,
+        GGML_OPT_BUILD_TYPE_OPT,
+    };
+
+    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
+    struct ggml_opt_optimizer_params {
+        // AdamW optimizer parameters
+        struct {
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float wd;    // weight decay for AdamW, use 0.0f to disable
+        } adamw;
+    };
+
+    // callback to calculate optimizer parameters prior to a backward pass
+    // userdata can be used to pass arbitrary data
+    typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
+
+    // returns the default optimizer params (constant)
+    // userdata is not used
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
+
+    // parameters for initializing a new optimization context
+    struct ggml_opt_params {
+        ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
+
+        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
+
+        // the forward graph is defined by inputs and outputs
+        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
+        struct ggml_tensor * inputs;
+        struct ggml_tensor * outputs;
+
+        enum ggml_opt_loss_type  loss_type;
+        enum ggml_opt_build_type build_type;
+
+        int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    // get parameters for an optimization context with defaults set where possible
+    // parameters for which no sensible defaults exist are supplied as arguments to this function
+    GGML_API ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t      backend_sched,
+            struct ggml_context     * ctx_compute,
+            struct ggml_tensor      * inputs,
+            struct ggml_tensor      * outputs,
+            enum ggml_opt_loss_type   loss_type);
+
+    GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
+    GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
+
+    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
+
+    // get underlying tensors that store data
+    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
+    GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
+    GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
+    GGML_API struct ggml_tensor * ggml_opt_loss(    ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
+    GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
+    GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
+
+    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
+
+    // ====== Optimization Result ======
+
+    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
+    GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
+
+    // get data from result, uncertainties are optional and can be ignored by passing NULL
+    GGML_API void ggml_opt_result_ndata(   ggml_opt_result_t result, int64_t * ndata);                  // writes 1 value, number of datapoints
+    GGML_API void ggml_opt_result_loss(    ggml_opt_result_t result, double  * loss,     double * unc); // writes 1 value
+    GGML_API void ggml_opt_result_pred(    ggml_opt_result_t result, int32_t * pred);                   // writes ndata values
+    GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double  * accuracy, double * unc); // writes 1 value
+
+    // ====== Computation ======
+
+    // do forward pass, increment result if not NULL
+    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // do forward pass, increment result if not NULL, do backward pass
+    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // ############################################################################
+    // ## The high-level functions start here. They do not depend on any private ##
+    // ## functions or structs and can be copied to and adapted for user code.   ##
+    // ############################################################################
+
+    // ====== Intended Usage ======
+    //
+    // 1. Select the appropriate loss for your problem.
+    // 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
+    //    Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
+    // 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
+    //    The first context should contain the model parameters and inputs and be allocated statically in user code.
+    //    The second context should contain all other tensors and will be (re)allocated automatically.
+    //    Due to this automated allocation the data of the second context is not defined when accessed in user code.
+    //    Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
+    // 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
+
+    // signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
+    typedef void (*ggml_opt_epoch_callback)(
+            bool               train,       // true after training evaluation, false after validation evaluation
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,      // result associated with the dataset subsection
+            int64_t            ibatch,      // number of batches that have been evaluated so far
+            int64_t            ibatch_max,  // total number of batches in this dataset subsection
+            int64_t            t_start_us); // time at which the evaluation on the dataset subsection was started
+
+    // do training on front of dataset, do evaluation only on back of dataset
+    GGML_API void ggml_opt_epoch(
+            ggml_opt_context_t      opt_ctx,
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,   // result to increment during training, ignored if NULL
+            ggml_opt_result_t       result_eval,    // result to increment during evaluation, ignored if NULL
+            int64_t                 idata_split,    // data index at which to split training and evaluation
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    // callback that prints a progress bar on stderr
+    GGML_API void ggml_opt_epoch_callback_progress_bar(
+            bool               train,
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,
+            int64_t            ibatch,
+            int64_t            ibatch_max,
+            int64_t            t_start_us);
+
+    // fit model defined by inputs and outputs to dataset
+    GGML_API void ggml_opt_fit(
+            ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
+            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
+            enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
+            int64_t                         nepoch,         // how many times the dataset should be iterated over
+            int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
+            float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
+            bool                            silent);        // whether or not info prints to stderr should be suppressed
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-rpc.h

@@ -10,14 +10,18 @@ extern "C" {
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml-sycl.h

@@ -17,26 +17,33 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 
 // devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_API void ggml_backend_sycl_print_sycl_devices(void);
-GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API int  ggml_backend_sycl_get_device_count();
-GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
+                                                       char *description,
+                                                       size_t description_size);
+GGML_BACKEND_API int  ggml_backend_sycl_get_device_count();
+GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
-// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-vulkan.h

@@ -10,19 +10,21 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
-GGML_API void ggml_vk_instance_init(void);
+GGML_BACKEND_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API int  ggml_backend_vk_get_device_count(void);
-GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_BACKEND_API int  ggml_backend_vk_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
 
 #ifdef  __cplusplus
 }

ggml/include/ggml.h

@@ -176,15 +176,15 @@
 #ifdef GGML_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef GGML_BUILD
-#            define GGML_API __declspec(dllexport)
+#            define GGML_API __declspec(dllexport) extern
 #        else
-#            define GGML_API __declspec(dllimport)
+#            define GGML_API __declspec(dllimport) extern
 #        endif
 #    else
-#        define GGML_API __attribute__ ((visibility ("default")))
+#        define GGML_API __attribute__ ((visibility ("default"))) extern
 #    endif
 #else
-#    define GGML_API
+#    define GGML_API extern
 #endif
 
 // TODO: support for clang
@@ -237,7 +237,9 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#define GGML_ROPE_TYPE_NEOX 2
+#define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
 
 #define GGUF_MAGIC "GGUF"
 
@@ -384,12 +386,15 @@ extern "C" {
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
         GGML_TYPE_BF16    = 30,
-        GGML_TYPE_Q4_0_4_4 = 31,
-        GGML_TYPE_Q4_0_4_8 = 32,
-        GGML_TYPE_Q4_0_8_8 = 33,
+        // GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
+        // GGML_TYPE_Q4_0_4_8 = 32,
+        // GGML_TYPE_Q4_0_8_8 = 33,
         GGML_TYPE_TQ1_0   = 34,
         GGML_TYPE_TQ2_0   = 35,
-        GGML_TYPE_COUNT,
+        // GGML_TYPE_IQ4_NL_4_4 = 36,
+        // GGML_TYPE_IQ4_NL_4_8 = 37,
+        // GGML_TYPE_IQ4_NL_8_8 = 38,
+        GGML_TYPE_COUNT   = 39,
     };
 
     // precision
@@ -430,9 +435,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
     };
 
     // available tensor operations:
@@ -496,6 +498,7 @@ extern "C" {
         GGML_OP_POOL_2D_BACK,
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
+        GGML_OP_PAD_REFLECT_1D,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -509,7 +512,7 @@ extern "C" {
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
-        GGML_OP_RWKV_WKV,
+        GGML_OP_RWKV_WKV6,
 
         GGML_OP_UNARY,
 
@@ -558,10 +561,10 @@ extern "C" {
 
     enum ggml_log_level {
         GGML_LOG_LEVEL_NONE  = 0,
-        GGML_LOG_LEVEL_INFO  = 1,
-        GGML_LOG_LEVEL_WARN  = 2,
-        GGML_LOG_LEVEL_ERROR = 3,
-        GGML_LOG_LEVEL_DEBUG = 4,
+        GGML_LOG_LEVEL_DEBUG = 1,
+        GGML_LOG_LEVEL_INFO  = 2,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_ERROR = 4,
         GGML_LOG_LEVEL_CONT  = 5, // continue previous log
     };
 
@@ -573,6 +576,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    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
+    };
+
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type type;
@@ -595,7 +605,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -608,7 +617,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        // char padding[4];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -618,67 +627,6 @@ extern "C" {
     // If it returns true, the computation is aborted
     typedef bool (*ggml_abort_callback)(void * data);
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
-    // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
-    struct ggml_cplan {
-        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
-        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
-
-        int n_threads;
-        struct ggml_threadpool * threadpool;
-
-        // abort ggml_graph_compute when true
-        ggml_abort_callback abort_callback;
-        void *              abort_callback_data;
-    };
-
-    // scratch buffer
-    // TODO: deprecate and remove
-    struct ggml_scratch {
-        size_t offs;
-        size_t size;
-        void * data;
-    };
-
-    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
-    };
-
-    // numa strategies
-    enum ggml_numa_strategy {
-        GGML_NUMA_STRATEGY_DISABLED   = 0,
-        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
-        GGML_NUMA_STRATEGY_ISOLATE    = 2,
-        GGML_NUMA_STRATEGY_NUMACTL    = 3,
-        GGML_NUMA_STRATEGY_MIRROR     = 4,
-        GGML_NUMA_STRATEGY_COUNT
-    };
 
     //
     // GUID
@@ -701,9 +649,6 @@ extern "C" {
     // accepts a UTF-8 path, even on Windows
     GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
-    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
-    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
-
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
@@ -766,7 +711,6 @@ extern "C" {
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
     GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
@@ -806,8 +750,7 @@ extern "C" {
             int64_t ne2,
             int64_t ne3);
 
-    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
-    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+    GGML_API void * ggml_new_buffer(struct ggml_context * ctx, size_t nbytes);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
@@ -817,35 +760,25 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
-    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
-    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
-
     // Converts a flat index into coordinates
-    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+    GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
-    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
-
-    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
-
-    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
-
-    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
-
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
     GGML_ATTRIBUTE_FORMAT(2, 3)
     GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
+    // Tensor flags
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
+
     //
     // operations on tensors with backpropagation
     //
@@ -1512,6 +1445,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,
@@ -1559,7 +1508,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    void ggml_rope_yarn_corr_dims(
+    GGML_API void ggml_rope_yarn_corr_dims(
         int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
@@ -1762,6 +1711,13 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
+    GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   p0,
+            int                   p1);
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]
@@ -1815,6 +1771,9 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
+            const struct ggml_tensor * a);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
@@ -1888,7 +1847,7 @@ extern "C" {
             struct ggml_tensor  * pw,
             struct ggml_tensor  * ph);
 
-    GGML_API struct ggml_tensor * ggml_rwkv_wkv(
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
             struct ggml_context * ctx,
             struct ggml_tensor  * k,
             struct ggml_tensor  * v,
@@ -2051,31 +2010,20 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+            struct ggml_tensor  * m,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
-    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
-
-    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
-
-    GGML_API void ggml_build_opt_adamw(
-            struct ggml_context * ctx,
-            struct ggml_cgraph  * gf,
-            struct ggml_cgraph  * gb,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
+        struct ggml_context * ctx_compute, // context for gradient computation
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2095,28 +2043,9 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
-
-    // ggml_graph_plan() has to be called before ggml_graph_compute()
-    // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan(
-                  const struct ggml_cgraph * cgraph,
-                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
-                    struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
-
-    // same as ggml_graph_compute() but the work data is allocated as a part of the context
-    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
-
-    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
     GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@@ -2127,201 +2056,14 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    GGML_API void ggml_build_backward_gradient_checkpointing(
-            struct ggml_context   * ctx,
-            struct ggml_cgraph    * gf,
-            struct ggml_cgraph    * gb,
-            struct ggml_cgraph    * gb_tmp,
-            struct ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
-    };
-
-    // linesearch methods
-    enum ggml_linesearch {
-        GGML_LINESEARCH_DEFAULT = 1,
-
-        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
-
-        GGML_LINESEARCH_FAIL = -128,
-        GGML_LINESEARCH_MINIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
     GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (ggml_opt_default_params) for default values
-    //
-    struct ggml_opt_params {
-        enum ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct ggml_opt_context {
-        struct ggml_context * ctx;
-        struct ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct ggml_tensor * g;  // current gradient
-            struct ggml_tensor * m;  // first moment
-            struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct ggml_tensor * x;    // current parameters
-            struct ggml_tensor * xp;   // previous parameters
-            struct ggml_tensor * g;    // current gradient
-            struct ggml_tensor * gp;   // previous gradient
-            struct ggml_tensor * d;    // search direction
-            struct ggml_tensor * pf;   // past function values
-            struct ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct ggml_tensor * lmys; // the L-BFGS memory ys
-            struct ggml_tensor * lms;  // the L-BFGS memory s
-            struct ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
-    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
-
-    // optimize the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt(
-            struct ggml_context * ctx,
-            struct ggml_opt_params params,
-            struct ggml_tensor * f);
-
-    // initialize optimizer context
-    GGML_API void ggml_opt_init(
-            struct ggml_context     * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume_g(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f,
-            struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb,
-            ggml_opt_callback callback,
-            void * callback_data);
-
-    //
-    // tensor flags
-    //
-    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
     //
     // quantization
@@ -2478,83 +2220,65 @@ extern "C" {
     GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
-    //
-    // system info
-    //
-
-    GGML_API int ggml_cpu_has_avx        (void);
-    GGML_API int ggml_cpu_has_avx_vnni   (void);
-    GGML_API int ggml_cpu_has_avx2       (void);
-    GGML_API int ggml_cpu_has_avx512     (void);
-    GGML_API int ggml_cpu_has_avx512_vbmi(void);
-    GGML_API int ggml_cpu_has_avx512_vnni(void);
-    GGML_API int ggml_cpu_has_avx512_bf16(void);
-    GGML_API int ggml_cpu_has_fma        (void);
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
-    GGML_API int ggml_cpu_has_arm_fma    (void);
-    GGML_API int ggml_cpu_has_metal      (void);
-    GGML_API int ggml_cpu_has_f16c       (void);
-    GGML_API int ggml_cpu_has_fp16_va    (void);
-    GGML_API int ggml_cpu_has_wasm_simd  (void);
-    GGML_API int ggml_cpu_has_blas       (void);
-    GGML_API int ggml_cpu_has_cuda       (void);
-    GGML_API int ggml_cpu_has_vulkan     (void);
-    GGML_API int ggml_cpu_has_kompute    (void);
-    GGML_API int ggml_cpu_has_gpublas    (void);
-    GGML_API int ggml_cpu_has_sse3       (void);
-    GGML_API int ggml_cpu_has_ssse3      (void);
-    GGML_API int ggml_cpu_has_riscv_v    (void);
-    GGML_API int ggml_cpu_has_sycl       (void);
-    GGML_API int ggml_cpu_has_rpc        (void);
-    GGML_API int ggml_cpu_has_vsx        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
-    GGML_API int ggml_cpu_has_cann       (void);
-    GGML_API int ggml_cpu_has_llamafile  (void);
-
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
-
-    //
-    // Internal types and functions exposed for tests and benchmarks
-    //
-
-#ifdef  __cplusplus
-// restrict not standard in C++
-#define GGML_RESTRICT
+#ifdef __cplusplus
+    // restrict not standard in C++
+#    if defined(__GNUC__)
+#        define GGML_RESTRICT __restrict__
+#    elif defined(__clang__)
+#        define GGML_RESTRICT __restrict
+#    elif defined(_MSC_VER)
+#        define GGML_RESTRICT __restrict
+#    else
+#        define GGML_RESTRICT
+#    endif
 #else
-#define GGML_RESTRICT restrict
+#    define GGML_RESTRICT restrict
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
-    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
-                                       const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
 
-    typedef struct {
+    struct ggml_type_traits {
         const char             * type_name;
         int64_t                  blck_size;
         int64_t                  blck_size_interleave; // interleave elements in blocks
         size_t                   type_size;
         bool                     is_quantized;
         ggml_to_float_t          to_float;
-        ggml_from_float_t        from_float;
         ggml_from_float_t        from_float_ref;
-        ggml_from_float_to_mat_t from_float_to_mat;
-        ggml_vec_dot_t           vec_dot;
-        enum ggml_type           vec_dot_type;
-        int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
-        ggml_gemv_t              gemv;
-        ggml_gemm_t              gemm;
-    } ggml_type_traits_t;
+    };
 
-    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
+
+    // ggml threadpool
+    // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend
+    // the goal should be to create an API that other backends can use move everything to the ggml base
+
+    // scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-aarch64.h

@@ -1,39 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-#ifdef __cplusplus
-}
-#endif
-

ggml/src/ggml-alloc.c

@@ -14,7 +14,7 @@
 
 //#define GGML_ALLOCATOR_DEBUG
 
-//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+//#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #define AT_PRINTF(...)
 
 
@@ -89,7 +89,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
     size = GGML_PAD(size, talloc->alignment);
 
     if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
-        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+        GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
                 __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
         GGML_ABORT("not enough space in the buffer");
     }
@@ -172,7 +172,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
             best_fit_block = alloc->n_free_blocks - 1;
         } else {
             // this should never happen
-            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+            GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
                     __func__, size, max_avail);
             GGML_ABORT("not enough space in the buffer");
         }
@@ -209,16 +209,16 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
                 }
             }
         }
-        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
         for (int i = 0; i < 1024; i++) {
             if (alloc->allocated_tensors[i].tensor) {
-                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
                     alloc->allocated_tensors[i].offset,
                     alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
                     ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
             }
         }
-        fprintf(stderr, "\n");
+        GGML_LOG_DEBUG("\n");
     }
 #endif
 
@@ -348,7 +348,6 @@ struct tensor_alloc {
 };
 
 struct leaf_alloc {
-    int buffer_id;
     struct tensor_alloc leaf;
 };
 
@@ -467,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
     if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@@ -541,7 +534,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
         size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
         hn->buffer_id = buffer_id;
         hn->offset = offset;
-        return;
     }
 }
 
@@ -740,7 +732,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
         if (leaf->view_src || leaf->data) {
             galloc->leaf_allocs[i].leaf.buffer_id = -1;
             galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
@@ -768,13 +759,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
         if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
 
             ggml_backend_buffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
             if (galloc->buffers[i] == NULL) {
-                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
                 return false;
             }
             ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
@@ -818,21 +809,25 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
     return talloc->size_max >= node_size;
 }
 
 static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
     if (galloc->n_nodes != graph->n_nodes) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
+        GGML_LOG_DEBUG("%s: graph has different number of nodes\n", __func__);
 #endif
         return true;
     }
 
     if (galloc->n_leafs != graph->n_leafs) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
+        GGML_LOG_DEBUG("%s: graph has different number of leafs\n", __func__);
 #endif
         return true;
     }
@@ -843,7 +838,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
 
         if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
+            GGML_LOG_DEBUG("%s: node %s is not valid\n", __func__, node->name);
 #endif
             return true;
         }
@@ -855,7 +850,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
             }
             if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
 #ifndef NDEBUG
-                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+                GGML_LOG_DEBUG("%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
 #endif
                 return true;
             }
@@ -869,14 +864,14 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     if (ggml_gallocr_needs_realloc(galloc, graph)) {
         if (galloc->n_buffers == 1) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
+            GGML_LOG_DEBUG("%s: reallocating buffers automatically\n", __func__);
 #endif
             if (!ggml_gallocr_reserve(galloc, graph)) {
                 return false;
             }
         } else {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+            GGML_LOG_DEBUG("%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
 #endif
             return false;
         }
@@ -940,7 +935,7 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
     ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     if (buffer == NULL) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
+        GGML_LOG_DEBUG("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
         for (size_t i = 0; i < *n_buffers; i++) {
             ggml_backend_buffer_free((*buffers)[i]);
@@ -990,7 +985,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
         }
 
         if (this_size > max_size) {
-            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+            GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
                     __func__, t->name,
                     ggml_backend_buft_name(buft),
                     this_size, max_size);
@@ -1022,7 +1017,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     if (n_buffers == 0) {
 #ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+        GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
 #endif
         return NULL;
     }

ggml/src/ggml-amx/CMakeLists.txt

@@ -0,0 +1,107 @@
+if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
+        CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
+    message(STATUS "Using AMX")
+
+    file(GLOB   GGML_HEADERS_AMX "*.h")
+    list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
+
+    file(GLOB   GGML_SOURCES_AMX "*.cpp")
+
+    add_library(ggml-amx
+                ${GGML_HEADERS_AMX}
+                ${GGML_SOURCES_AMX})
+
+    target_link_libraries(ggml-amx PRIVATE ggml-base)
+    target_include_directories(ggml-amx PRIVATE . ..)
+
+    # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
+    # TODO: integrate AMX backend into the CPU backend
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../ggml-cpu/cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+
+    target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
+else()
+    set(GGML_AMX OFF PARENT_SCOPE)
+    message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
+endif()

ggml/src/ggml-amx/common.h

@@ -0,0 +1,94 @@
+#pragma once
+
+#include "ggml.h"
+// hack until AMX is moved into the CPU backend
+#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
+
+#include <algorithm>
+#include <memory>
+#include <type_traits>
+
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+#define VNNI_BLK 4
+
+#define AMX_BLK_SIZE 32
+
+#define TMM0 0
+#define TMM1 1
+#define TMM2 2
+#define TMM3 3
+#define TMM4 4
+#define TMM5 5
+#define TMM6 6
+#define TMM7 7
+
+// parallel routines
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int nth, int n, const func_t& f) {
+#if defined(_OPENMP)
+#pragma omp parallel num_threads(nth)
+{
+    //int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+
+    GGML_UNUSED(nth);
+#endif
+}
+
+// quantized types that have AMX support
+inline bool qtype_has_amx_kernels(const enum ggml_type type) {
+    // TODO: fix padding for vnni format
+    return (type == GGML_TYPE_Q4_0) ||
+        (type == GGML_TYPE_Q4_1);
+        //(type == GGML_TYPE_Q8_0) ||
+        //(type == GGML_TYPE_Q4_K) ||
+        //(type == GGML_TYPE_Q5_K) ||
+        //(type == GGML_TYPE_Q6_K) ||
+        //(type == GGML_TYPE_IQ4_XS);
+}
+
+// ggml backend context
+struct ggml_backend_amx_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+};

ggml/src/ggml-amx/ggml-amx.cpp

@@ -0,0 +1,446 @@
+#include "ggml-amx.h"
+#include "ggml-amx/common.h"
+#include "ggml-amx/mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__)
+
+// AMX buffer interface
+static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)(buffer->context);
+}
+
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *)tensor->data + offset, data, size);
+    }
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
+        } else {
+            memcpy(dst->data, src->data, ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_amx_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    return ggml_backend_amx_get_alloc_size(tensor);
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_amx;
+}
+
+// backend interface
+
+static const char * ggml_backend_amx_name(ggml_backend_t backend) {
+    return "AMX";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_amx_free(ggml_backend_t backend) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_backend_amx_mul_mat(ctx, node);
+            break;
+
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+
+        default:
+            fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+            GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i ggml_backend_amx_i = {
+    /* .get_name                = */ ggml_backend_amx_name,
+    /* .free                    = */ ggml_backend_amx_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_amx_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_amx_guid() {
+    static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
+    return &guid;
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+
+ggml_backend_t ggml_backend_amx_init() {
+
+    // invoke a Linux system call to request access to AMX features
+    ggml_amx_init();
+
+    // backend context
+    ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
+
+    // ggml amx backend
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_amx_guid(),
+        /* .interface = */ ggml_backend_amx_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_amx(backend_amx));
+
+    ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
+    ctx->n_threads = n_threads;
+}
+
+// device interface
+
+static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
+    return "AMX";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
+    return "Intel Advanced Matrix Extensions";
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_amx_device_get_name(dev);
+    props->description = ggml_backend_amx_device_get_description(dev);
+    props->type        = ggml_backend_amx_device_get_type(dev);
+    ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_amx_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_amx_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+
+    // handle only 2d gemm for now
+    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+    };
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT: {
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const enum ggml_type type = src0->type;
+            const int64_t ne0 = op->ne[0];
+
+            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
+            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
+            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
+
+            bool can_use_amx =
+                is_contiguous_2d(src0) &&       // src0 must be contiguous
+                is_contiguous_2d(src1) &&       // src1 must be contiguous
+                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
+                has_amx_kernels &&              // with amx kernel impls
+                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
+
+            return can_use_amx;
+        }
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
+    /* .get_name             = */ ggml_backend_amx_device_get_name,
+    /* .get_description      = */ ggml_backend_amx_device_get_description,
+    /* .get_memory           = */ ggml_backend_amx_device_get_memory,
+    /* .get_type             = */ ggml_backend_amx_device_get_type,
+    /* .get_props            = */ ggml_backend_amx_device_get_props,
+    /* .init_backend         = */ ggml_backend_amx_device_init,
+    /* .get_buffer_type      = */ ggml_backend_amx_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_amx_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_amx_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
+    return "AMX";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_amx_device = {
+        /* .iface   = */ ggml_backend_amx_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_amx_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_amx_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
+    /* .get_name         = */ ggml_backend_amx_reg_get_name,
+    /* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_amx_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_amx_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    static struct ggml_backend_reg ggml_backend_amx_reg = {
+        /* .iface   = */ ggml_backend_amx_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_amx_reg;
+}
+
+#else // if defined(__AMX_INT8__)
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
+    return nullptr;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+    return false;
+}
+
+ggml_backend_t ggml_backend_amx_init(void) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+    return nullptr;
+}
+
+void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
+    fprintf(stderr, "GGML is not compiled with AMX support!\n");
+
+    GGML_UNUSED(backend_amx);
+    GGML_UNUSED(n_threads);
+}
+
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    return nullptr;
+}
+
+#endif

ggml/src/ggml-amx/mmq.h

@@ -0,0 +1,17 @@
+#pragma once
+#include "common.h"
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
+
+void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+
+void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst);
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-backend-impl.h

@@ -8,6 +8,8 @@
 extern "C" {
 #endif
 
+    #define GGML_BACKEND_API_VERSION 1
+
     //
     // Backend buffer type
     //
@@ -22,7 +24,7 @@ extern "C" {
         size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
         // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
         size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
-        // (optional) check if tensor data is in host memory (defaults to false)
+        // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
         bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
     };
 
@@ -37,7 +39,6 @@ extern "C" {
     //
 
     struct ggml_backend_buffer_i {
-        const char * (*get_name)     (ggml_backend_buffer_t buffer);
         // (optional) free the buffer
         void         (*free_buffer)  (ggml_backend_buffer_t buffer);
         // base address of the buffer
@@ -64,20 +65,20 @@ extern "C" {
         enum ggml_backend_buffer_usage usage;
     };
 
-    ggml_backend_buffer_t ggml_backend_buffer_init(
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
                    ggml_backend_buffer_type_t buft,
             struct ggml_backend_buffer_i      iface,
                    void *                     context,
                    size_t                     size);
 
     // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     // multi-buffer
     // buffer that contains a collection of buffers
-    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend (stream)
@@ -88,18 +89,16 @@ extern "C" {
 
         void (*free)(ggml_backend_t backend);
 
-        // buffer allocation
-        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
-
         // (optional) asynchronous tensor data access
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) complete all pending operations
+        // (optional) complete all pending operations (required if the backend supports async operations)
         void (*synchronize)(ggml_backend_t backend);
 
-        // (optional) compute graph with a plan (not used currently)
+        // (optional) graph plans (not used currently)
+        // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
@@ -110,21 +109,6 @@ extern "C" {
         // compute graph (always async if supported by the backend)
         enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
-        // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface
-        //            new backends should implement the device interface instead
-
-        // These functions are being moved to the device interface
-        // check if the backend can compute an operation
-        bool (*supports_op)  (ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend can use tensors allocated in a buffer type
-        bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*offload_op)   (ggml_backend_t backend, const struct ggml_tensor * op);
-
         // (optional) event synchronization
         // record an event on this stream
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
@@ -184,9 +168,8 @@ extern "C" {
         // check if the backend can use tensors allocated in a buffer type
         bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
 
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
+        // (optional) check if the backend wants to run an operation, even if the weights are allocated in an incompatible buffer
+        // these should be expensive operations that may benefit from running on this backend instead of the CPU backend
         bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
 
         // (optional) event synchronization
@@ -218,17 +201,55 @@ extern "C" {
     };
 
     struct ggml_backend_reg {
-        // int api_version; // TODO: for dynamic loading
+        int api_version; // initialize to GGML_BACKEND_API_VERSION
         struct ggml_backend_reg_i iface;
         void * context;
     };
 
-
     // Internal backend registry API
-    void ggml_backend_register(ggml_backend_reg_t reg);
-    void ggml_backend_device_register(ggml_backend_dev_t device);
-    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
-    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
+    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
+
+    // Add backend dynamic loading support to the backend
+
+    // Initialize the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+    // Optional: obtain a score for the backend based on the system configuration
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
+    typedef int                (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+#    ifdef __cplusplus
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            extern "C" {                                                 \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+            }                                                            \
+            ggml_backend_reg_t ggml_backend_init(void) {                 \
+                return reg_fn();                                         \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
+            extern "C" {                                   \
+            GGML_BACKEND_API int ggml_backend_score(void); \
+            }                                              \
+            int ggml_backend_score(void) {                 \
+                return score_fn();                         \
+            }
+#    else
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
+            ggml_backend_reg_t                  ggml_backend_init(void) { \
+                return reg_fn();                                          \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
+            GGML_BACKEND_API int ggml_backend_score(void);  \
+            int                  ggml_backend_score(void) { \
+                return score_fn();                          \
+            }
+#    endif
+#else
+#    define GGML_BACKEND_DL_IMPL(reg_fn)
+#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend-reg.cpp

@@ -0,0 +1,552 @@
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include <algorithm>
+#include <codecvt>
+#include <cstring>
+#include <filesystem>
+#include <locale>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <vector>
+
+#ifdef _WIN32
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#elif defined(__APPLE__)
+#    include <mach-o/dyld.h>
+#    include <dlfcn.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+
+// Backend registry
+#ifdef GGML_USE_CPU
+#include "ggml-cpu.h"
+#endif
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef GGML_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
+#ifdef GGML_USE_BLAS
+#include "ggml-blas.h"
+#endif
+
+#ifdef GGML_USE_RPC
+#include "ggml-rpc.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+#include "ggml-kompute.h"
+#endif
+
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static dl_handle * dl_load_library(const std::wstring & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static dl_handle * dl_load_library(const std::string & path) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return dl_load_library(converter.from_bytes(path));
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static void * dl_load_library(const std::string & path) {
+    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
+struct ggml_backend_reg_entry {
+    ggml_backend_reg_t reg;
+    dl_handle_ptr handle;
+};
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_entry> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+#ifdef GGML_USE_METAL
+        register_backend(ggml_backend_metal_reg());
+#endif
+#ifdef GGML_USE_SYCL
+        register_backend(ggml_backend_sycl_reg());
+#endif
+#ifdef GGML_USE_VULKAN
+        register_backend(ggml_backend_vk_reg());
+#endif
+#ifdef GGML_USE_OPENCL
+        register_backend(ggml_backend_opencl_reg());
+#endif
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
+#endif
+#ifdef GGML_USE_BLAS
+        register_backend(ggml_backend_blas_reg());
+#endif
+#ifdef GGML_USE_RPC
+        register_backend(ggml_backend_rpc_reg());
+#endif
+#ifdef GGML_USE_KOMPUTE
+        register_backend(ggml_backend_kompute_reg());
+#endif
+#ifdef GGML_USE_CPU
+        register_backend(ggml_backend_cpu_reg());
+#endif
+    }
+
+    ~ggml_backend_registry() {
+        // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
+        // since backend threads may still be running and accessing resources from the dynamic library
+        for (auto & entry : backends) {
+            if (entry.handle) {
+                entry.handle.release(); // NOLINT
+            }
+        }
+    }
+
+    void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
+        if (!reg) {
+            return;
+        }
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back({ reg, std::move(handle) });
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+
+    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+        dl_handle_ptr handle { dl_load_library(path) };
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+        if (score_fn && score_fn() == 0) {
+            if (!silent) {
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
+        if (!backend_init_fn) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        ggml_backend_reg_t reg = backend_init_fn();
+        if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
+            if (!silent) {
+                if (!reg) {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                } else {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
+                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                }
+            }
+            return nullptr;
+        }
+
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+
+        register_backend(reg, std::move(handle));
+
+        return reg;
+    }
+
+    void unload_backend(ggml_backend_reg_t reg, bool silent) {
+        auto it = std::find_if(backends.begin(), backends.end(),
+                               [reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
+
+        if (it == backends.end()) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: backend not found\n", __func__);
+            }
+            return;
+        }
+
+        if (!silent) {
+            GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
+        }
+
+        // remove devices
+        devices.erase(
+            std::remove_if(devices.begin(), devices.end(),
+                            [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
+            devices.end());
+
+        // remove backend
+        backends.erase(it);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index].reg;
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (striequals(ggml_backend_reg_name(reg), name)) {
+            return reg;
+        }
+    }
+    return nullptr;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (striequals(ggml_backend_dev_name(dev), name)) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    }
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, nullptr);
+}
+
+// Dynamic loading
+ggml_backend_reg_t ggml_backend_load(const char * path) {
+    return get_reg().load_backend(path, false);
+}
+
+void ggml_backend_unload(ggml_backend_reg_t reg) {
+    get_reg().unload_backend(reg, true);
+}
+
+static std::string get_executable_path() {
+#if defined(__APPLE__)
+    // get executable path
+    std::vector<char> path;
+    uint32_t size;
+    while (true) {
+        size = path.size();
+        if (_NSGetExecutablePath(path.data(), &size) == 0) {
+            break;
+        }
+        path.resize(size);
+    }
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "/";
+#elif defined(__linux__)
+    std::string base_path = ".";
+    std::vector<char> path(1024);
+    while (true) {
+        // get executable path
+        ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+        if (len == -1) {
+            break;
+        }
+        if (len < (ssize_t) path.size()) {
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
+        }
+        path.resize(path.size() * 2);
+    }
+
+    return base_path + "/";
+#elif defined(_WIN32)
+    std::vector<char> path(MAX_PATH);
+    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    if (len == 0) {
+        return "";
+    }
+    std::string base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "\\";
+#endif
+}
+
+static std::string backend_filename_prefix() {
+#ifdef _WIN32
+    return "ggml-";
+#else
+    return "libggml-";
+#endif
+}
+
+static std::string backend_filename_suffix() {
+#ifdef _WIN32
+    return ".dll";
+#else
+    return ".so";
+#endif
+}
+
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
+    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
+     // TODO: search system paths
+    std::string file_prefix = backend_filename_prefix() + name + "-";
+    std::vector<std::string> search_paths;
+    if (user_search_path == nullptr) {
+        search_paths.push_back("./");
+        search_paths.push_back(get_executable_path());
+    } else {
+#if defined(_WIN32)
+        search_paths.push_back(std::string(user_search_path) + "\\");
+#else
+        search_paths.push_back(std::string(user_search_path) + "/");
+#endif
+    }
+
+    int best_score = 0;
+    std::string best_path;
+
+    namespace fs = std::filesystem;
+    for (const auto & search_path : search_paths) {
+        if (!fs::exists(search_path)) {
+            continue;
+        }
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
+        for (const auto & entry : dir_it) {
+            if (entry.is_regular_file()) {
+                std::string filename = entry.path().filename().string();
+                std::string ext = entry.path().extension().string();
+                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
+                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    if (!handle && !silent) {
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                    }
+                    if (handle) {
+                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().string();
+                            }
+                        } else {
+                            if (!silent) {
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    if (best_score == 0) {
+        // try to load the base backend
+        for (const auto & search_path : search_paths) {
+            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
+            if (fs::exists(path)) {
+                return get_reg().load_backend(path.c_str(), silent);
+            }
+        }
+        return nullptr;
+    }
+
+    return get_reg().load_backend(best_path.c_str(), silent);
+}
+
+void ggml_backend_load_all() {
+    ggml_backend_load_all_from_path(nullptr);
+}
+
+void ggml_backend_load_all_from_path(const char * dir_path) {
+#ifdef NDEBUG
+    bool silent = true;
+#else
+    bool silent = false;
+#endif
+
+    ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("cann", silent, dir_path);
+    ggml_backend_load_best("cuda", silent, dir_path);
+    ggml_backend_load_best("hip", silent, dir_path);
+    ggml_backend_load_best("kompute", silent, dir_path);
+    ggml_backend_load_best("metal", silent, dir_path);
+    ggml_backend_load_best("rpc", silent, dir_path);
+    ggml_backend_load_best("sycl", silent, dir_path);
+    ggml_backend_load_best("vulkan", silent, dir_path);
+    ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("musa", silent, dir_path);
+    ggml_backend_load_best("cpu", silent, dir_path);
+}

ggml/src/ggml-blas/CMakeLists.txt

@@ -0,0 +1,87 @@
+if (GGML_STATIC)
+    set(BLA_STATIC ON)
+endif()
+#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+#    set(BLA_SIZEOF_INTEGER 8)
+#endif()
+
+set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+find_package(BLAS)
+
+if (BLAS_FOUND)
+    message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+    ggml_add_backend_library(ggml-blas
+                             ggml-blas.cpp
+                            )
+
+    if (${GGML_BLAS_VENDOR} MATCHES "Apple")
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+        add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
+    elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+        # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+        # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+        find_package(PkgConfig REQUIRED)
+        if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+            pkg_check_modules(DepBLAS blas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+            # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+            pkg_check_modules(DepBLAS openblas64)
+            if (NOT DepBLAS_FOUND)
+                pkg_check_modules(DepBLAS openblas)
+            endif()
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+            add_compile_definitions(GGML_BLAS_USE_BLIS)
+            pkg_check_modules(DepBLAS blis)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+            pkg_check_modules(DepBLAS blas-atlas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+            pkg_check_modules(DepBLAS flexiblas_api)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+            # all Intel* libraries share the same include path
+            pkg_check_modules(DepBLAS mkl-sdl)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+            # this doesn't provide pkg-config
+            # suggest to assign BLAS_INCLUDE_DIRS on your own
+            if ("${NVHPC_VERSION}" STREQUAL "")
+                message(WARNING "Better to set NVHPC_VERSION")
+            else()
+                set(DepBLAS_FOUND ON)
+                set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+            endif()
+        endif()
+        if (DepBLAS_FOUND)
+            set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+        else()
+            message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+            " detected by pkgconfig, trying to find cblas.h from possible paths...")
+            find_path(BLAS_INCLUDE_DIRS
+                NAMES cblas.h
+                HINTS
+                    /usr/include
+                    /usr/local/include
+                    /usr/include/openblas
+                    /opt/homebrew/opt/openblas/include
+                    /usr/local/opt/openblas/include
+                    /usr/include/x86_64-linux-gnu/openblas/include
+            )
+        endif()
+    endif()
+
+    message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+    target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
+
+    if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+        add_compile_definitions(GGML_BLAS_USE_MKL)
+    endif()
+
+    target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
+    target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
+else()
+    message(ERROR "BLAS not found, please refer to "
+                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                  " to set correct GGML_BLAS_VENDOR")
+endif()

ggml/src/ggml-blas/ggml-blas.cpp

@@ -4,8 +4,9 @@
 
 #include <future>
 #include <vector>
+#include <cstring>
 
-#if defined(GGML_USE_ACCELERATE)
+#if defined(GGML_BLAS_USE_ACCELERATE)
 #   include <Accelerate/Accelerate.h>
 #elif defined(GGML_BLAS_USE_MKL)
 #   include <mkl.h>
@@ -26,30 +27,6 @@ struct ggml_backend_blas_context {
 #endif
 };
 
-// helper function to determine if it is better to use BLAS or not
-// for large matrices, BLAS is faster
-static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
-    const struct ggml_tensor * src0 = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-
-    const int64_t ne10 = src1->ne[0];
-
-    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) &&
-        ggml_is_contiguous(src1) &&
-        src1->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
-
-        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
-        return true;
-    }
-
-    return false;
-}
-
 static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
@@ -88,8 +65,8 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
 
     // convert src0 to float
     if (type != GGML_TYPE_F32) {
-        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
-        ggml_to_float_t const to_float = type_traits.to_float;
+        const auto * type_traits = ggml_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits->to_float;
 
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -235,7 +212,7 @@ static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct g
 
 // backend interface
 
-static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+static const char * ggml_backend_blas_get_name(ggml_backend_t backend) {
     return "BLAS";
 
     GGML_UNUSED(backend);
@@ -247,12 +224,6 @@ static void ggml_backend_blas_free(ggml_backend_t backend) {
     delete backend;
 }
 
-static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(backend);
-}
-
 static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
 
@@ -285,31 +256,9 @@ static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend,
     GGML_UNUSED(backend);
 }
 
-static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    const struct ggml_tensor * src0 = op->src[0];
-    const struct ggml_tensor * src1 = op->src[1];
-
-    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
-           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
-                                          op->src[1]->type == GGML_TYPE_F32 &&
-                                          ggml_is_matrix(src0) &&
-                                          ggml_is_matrix(src1) &&
-                                          ggml_is_contiguous(src0) &&
-                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
-
-    GGML_UNUSED(backend);
-}
-
-static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(backend);
-}
-
 static struct ggml_backend_i blas_backend_i = {
-    /* .get_name                = */ ggml_backend_blas_name,
+    /* .get_name                = */ ggml_backend_blas_get_name,
     /* .free                    = */ ggml_backend_blas_free,
-    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
@@ -319,9 +268,6 @@ static struct ggml_backend_i blas_backend_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_blas_graph_compute,
-    /* .supports_op             = */ ggml_backend_blas_supports_op,
-    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
-    /* .offload_op              = */ NULL,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
 };
@@ -337,18 +283,18 @@ ggml_backend_t ggml_backend_blas_init(void) {
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_blas_guid(),
         /* .interface = */ blas_backend_i,
-        /* .device    = */ nullptr,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
         /* .context   = */ ctx,
     };
 
-#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
     if (openblas_get_parallel() != OPENBLAS_OPENMP) {
-        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+        GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
     }
 #endif
 
-#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
-    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
 #endif
 
     return backend;
@@ -364,3 +310,208 @@ void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads)
     ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
     ctx->n_threads = n_threads;
 }
+
+// device interface
+
+static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
+    return "BLAS";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
+    #if defined(GGML_BLAS_USE_ACCELERATE)
+        return "Accelerate";
+    #elif defined(GGML_BLAS_USE_MKL)
+        return "MKL";
+    #elif defined(GGML_BLAS_USE_BLIS)
+        return "BLIS";
+    #elif defined(GGML_BLAS_USE_NVPL)
+        return "NVPL";
+    #elif defined(OPENBLAS_VERSION)
+        return "OpenBLAS";
+    #else
+        return "BLAS";
+    #endif
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_blas_device_get_name(dev);
+    props->description = ggml_backend_blas_device_get_description(dev);
+    props->type        = ggml_backend_blas_device_get_type(dev);
+    ggml_backend_blas_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_blas_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+        {
+            // BLAS usually is only faster for large matrices
+            const struct ggml_tensor * src0 = op->src[0];
+            const struct ggml_tensor * src1 = op->src[1];
+
+            const int64_t ne10 = src1->ne[0];
+
+            const int64_t ne0 = op->ne[0];
+            const int64_t ne1 = op->ne[1];
+
+            // TODO: find the optimal value
+            const int64_t min_batch = 32;
+
+            return ggml_is_contiguous(src0) &&
+                   ggml_is_contiguous(src1) &&
+                   src1->type == GGML_TYPE_F32 &&
+                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+        }
+
+        case GGML_OP_OUT_PROD:
+            return op->src[0]->type == GGML_TYPE_F32 &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   ggml_is_matrix(src0) &&
+                   ggml_is_matrix(src1) &&
+                   ggml_is_contiguous(src0) &&
+                   (ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+
+        default:
+            return false;
+
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_blas_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
+    /* .get_name             = */ ggml_backend_blas_device_get_name,
+    /* .get_description      = */ ggml_backend_blas_device_get_description,
+    /* .get_memory           = */ ggml_backend_blas_device_get_memory,
+    /* .get_type             = */ ggml_backend_blas_device_get_type,
+    /* .get_props            = */ ggml_backend_blas_device_get_props,
+    /* .init_backend         = */ ggml_backend_blas_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_blas_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_blas_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_blas_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// backend reg interface
+
+static const char * ggml_backend_blas_reg_get_name(ggml_backend_reg_t reg) {
+    return "BLAS";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_blas_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_blas_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_device ggml_backend_blas_device = {
+        /* .iface   = */ ggml_backend_blas_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ nullptr,
+    };
+
+    return &ggml_backend_blas_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static void * ggml_backend_blas_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        return (void *)ggml_backend_blas_set_n_threads;
+    }
+    return NULL;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
+    /* .get_name         = */ ggml_backend_blas_reg_get_name,
+    /* .get_device_count = */ ggml_backend_blas_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_blas_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_blas_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_blas_reg(void) {
+    static struct ggml_backend_reg ggml_backend_blas_reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_blas_reg_i,
+        /* .context     = */ NULL,
+    };
+
+    return &ggml_backend_blas_reg;
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)

ggml/src/ggml-cann/CMakeLists.txt

@@ -0,0 +1,76 @@
+if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+    set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+    message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+endif()
+
+# Auto-detech Soc type and Soc version, if detect failed, will abort build
+set(SOC_VERSION "")
+function(detect_ascend_soc_type SOC_VERSION)
+    execute_process(
+        COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
+        OUTPUT_VARIABLE npu_info
+        RESULT_VARIABLE npu_result
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    if("${npu_info}" STREQUAL "" OR ${npu_result})
+        message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
+    endif()
+    set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
+endfunction()
+
+if(NOT SOC_TYPE)
+    detect_ascend_soc_type(SOC_VERSION)
+    set(SOC_TYPE "${SOC_VERSION}")
+    message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
+endif()
+
+string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
+
+# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND_310P.
+string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
+set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
+string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
+
+if (CANN_INSTALL_DIR)
+    # Only Support Linux.
+    if (NOT UNIX)
+        message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
+    endif()
+
+    # Supported platforms: x86-64, arm64
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+    else()
+        message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    # Set header and libs
+    set(CANN_INCLUDE_DIRS
+        ${CANN_INSTALL_DIR}/include
+        ${CANN_INSTALL_DIR}/include/aclnn
+        ${CANN_INSTALL_DIR}/acllib/include
+    )
+
+    add_subdirectory(kernels)
+    list(APPEND CANN_LIBRARIES
+        ascendcl
+        nnopbase
+        opapi
+        acl_op_compiler
+        ascendc_kernels
+    )
+
+    file(GLOB GGML_SOURCES_CANN "*.cpp")
+
+    ggml_add_backend_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE ${CANN_INCLUDE_DIRS})
+    target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
+
+    target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
+
+    message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+    message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+else()
+    message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
+endif()

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -22,11 +22,14 @@
 
 #include "aclnn_ops.h"
 
+#include <aclnnop/aclnn_addcdiv.h>
 #include <aclnnop/aclnn_avgpool2d.h>
+#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_cast.h>
 #include <aclnnop/aclnn_constant_pad_nd.h>
 #include <aclnnop/aclnn_copy.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_div.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_group_norm.h>
@@ -34,6 +37,7 @@
 #include <aclnnop/aclnn_layer_norm.h>
 #include <aclnnop/aclnn_matmul.h>
 #include <aclnnop/aclnn_max_pool.h>
+#include <aclnnop/aclnn_mm.h>
 #include <aclnnop/aclnn_permute.h>
 #include <aclnnop/aclnn_pow_tensor_tensor.h>
 #include <aclnnop/aclnn_reduce_sum.h>
@@ -53,6 +57,7 @@
 #include <exception>
 #include <vector>
 
+#include "ggml-impl.h"
 #include "kernels/ascendc_kernels.h"
 
 #define GGML_COMMON_DECL_C
@@ -241,10 +246,14 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    int64_t concat_dim = 1;
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+    int32_t acl_dim = 3 - dim;
+
     aclTensor* tensors[] = {acl_src0, acl_src1};
     aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
-    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+    aclnn_concat(ctx, tensorList, acl_dst, acl_dim);
 
     ACL_CHECK(aclDestroyTensorList(tensorList));
     ACL_CHECK(aclDestroyTensor(acl_dst));
@@ -1096,9 +1105,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
 }
 
 /**
- * @brief Creates an ACL tensor initialized with ones using a provided buffer.
+ * @brief Creates an ACL tensor initialized with value using a provided buffer.
  *
- * This function initializes a tensor with ones using the specified buffer and
+ * This function initializes a tensor with value using the specified buffer and
  * tensor parameters.
  *
  * @param ctx The context for the CANN backend operations.
@@ -1111,12 +1120,12 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
  * @param type_size The size of each element in the tensor data type.
  * @param value The value to be used for initializing the tensor (default
  * is 1.0).
- * @return An ACL tensor initialized with ones.
+ * @return An ACL tensor initialized with value.
  */
-static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer,
-                             size_t n_bytes, int64_t* ne, int64_t dims,
-                             aclDataType type, size_t type_size,
-                             float value = 1.0f) {
+static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
+                               size_t n_bytes, int64_t* ne, int64_t dims,
+                               aclDataType type, size_t type_size,
+                               float value = 1.0f) {
     aclTensor* acl_tensor =
         aclnn_zero(ctx, buffer, n_bytes, ne, dims, type, type_size);
     float alpha_host = 1.0f;
@@ -1158,7 +1167,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
-    aclTensor* acl_gamma = aclnn_ones(
+    aclTensor* acl_gamma = aclnn_values(
         ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
         ggml_cann_type_mapping(src->type), ggml_element_size(src));
 
@@ -1202,9 +1211,9 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
     aclTensor* mask_tensor =
-        aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne,
-                   GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src), value);
+        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
+                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
+                     ggml_element_size(src), value);
 
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -1437,10 +1446,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
@@ -1462,9 +1467,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
-
     // memory allocated increased to 3x when is_2D == false
     const int64_t n_bytes_factor = is_2D ? 1 : 3;
 
@@ -1768,6 +1770,92 @@ static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
     ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs element-wise division of tensor1 by tensor2 , multiplies the
+ result by the scalar value and adds it to self .
+ *
+ * Performs element-wise division of tensor1 by tensor2,
+ * multiplies the result by the scalar value and adds it to self .
+ * The operation is defined as:
+ * \f[
+ *     \text{out}_i = \text{selft}_i + \text{value} \times
+ \frac{\text{tensor1}_i}{\text{tensor2}_i}
+ * \f]
+
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_self The source tensor on which the addcdiv function will be
+ applied.
+ * @param tensor1 Numerator tensor.
+ * @param tensor2 Denominator tensor.
+ * @param value The value to be used for coefficient.
+ */
+static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx,
+                                  aclTensor* acl_self, aclTensor* tensor1,
+                                  aclTensor* tensor2, float value) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+    aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize(
+        acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+}
+
+/**
+ * @brief Matrix division, optionally in-place.
+ *
+ * This function division each element of the source tensor `acl_src` by the
+ * tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * If `inplace` is true, `acl_dst` will not be used and the operation is
+ * performed in-place on `acl_src`. The operation is defined as: \f[
+ *     \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src Numerator tensor..
+ * @param acl_other Denominator tensor.
+ * @param acl_dst The destination tensor where the result will be stored if
+ * `inplace` is false.
+ * @param inplace Flag indicating whether to perform the operation in-place on
+ * `acl_src`.
+ */
+static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                             aclTensor* acl_other, aclTensor* acl_dst,
+                             bool inplace) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    if (inplace) {
+        ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other,
+                                                  &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+    } else {
+        ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst,
+                                           &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(
+            aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    }
+}
+
 void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
                                   ggml_tensor* dst) {
     const ggml_tensor* src = dst->src[0];
@@ -2311,7 +2399,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                ctx.stream()));
 
     switch (src0->type) {
-        case GGML_TYPE_F32:
+        case GGML_TYPE_F32: {
+#ifdef ASCEND_310P
+            // Special operation for get_row_f32 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 8) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
+                                 src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f32(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2320,7 +2417,19 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
-        case GGML_TYPE_F16:
+        }
+        case GGML_TYPE_F16: {
+#ifdef ASCEND_310P
+            // Special operation for get_row_f16 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 16) != 0) {
+                size_t dst_len =
+                    src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
+                    ggml_type_size(
+                        GGML_TYPE_F32);  // out is also f32, even input is f16
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f16(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2329,6 +2438,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_Q4_0:
             aclrtlaunch_ascendc_get_row_q4_0(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
@@ -2407,7 +2517,6 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
                           aclTensor* acl_weight, aclTensor* acl_dst) {
     int8_t cube_math_type = 1;  // ALLOW_FP32_DOWN_PRECISION, when input is
                                 // fp32, atlas a2 will transpose it to HFLOAT32.
-
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -2425,6 +2534,81 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
         aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs matrix multiplication of two 2D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                      cube_math_type, &workspaceSize,
+                                      &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication of two 3D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                               cube_math_type, &workspaceSize,
+                                               &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
 /**
  * @brief Performs matrix multiplication with floating-point precision on
  * tensors using the CANN backend.
@@ -2446,20 +2630,39 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     // broadcast, when weight ne2 or ne3 is not 1, weight need repeat.
     BCAST_MUL_MAT_SHAPE(input, weight, dst);
 
-    // transpose weight: [1,2,3,4] -> [1,2,4,3]
+    int64_t n_dims = bcast_dims;
+    if (bcast_input_ne[3] == bcast_weight_ne[3] && bcast_input_ne[3] == 1) {
+        if (bcast_input_ne[2] == 1 && bcast_weight_ne[2] == 1) {
+            n_dims = 2;
+        } else if (bcast_input_ne[2] == 1) {
+            n_dims = 3;
+        }
+    }
+
+    aclTensor* acl_input_tensor =
+        ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
     int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
                               bcast_weight_ne[2], bcast_weight_ne[3],
                               bcast_weight_ne[4], bcast_weight_ne[5]};
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-
     aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims);
-    aclTensor* acl_input_tensor =
-        ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input));
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst));
-    aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
+
+    switch (n_dims) {
+        case 2:
+            aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        case 3:
+            aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        default:
+            aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+    }
 
     ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
     ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@@ -2480,51 +2683,47 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * multiplication will be stored.
  */
 static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst,
-                                   const enum ggml_type type) {
+                                    ggml_tensor* dst,
+                                    const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
-    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
-    // is regarded as batch. weight need transpose.
-    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
+    // The shape of the weight is NCHW.
+    // Matrix multiplication uses HW dims.
+    // HC is regarded as batch.
+    // weight need transpose.
     float weight_elem_size;
     if (type == GGML_TYPE_Q4_0) {
         weight_elem_size = float(sizeof(uint8_t)) / 2;
-    }
-    else if (type == GGML_TYPE_Q8_0) {
+    } else if (type == GGML_TYPE_Q8_0) {
         weight_elem_size = float(sizeof(uint8_t));
-    }
-    else {
+    } else {
         GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
     }
-    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
-
-    // size of one matrix is element_size * height * width.
-    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
+    float weight_nb[] = {src0->ne[0] * weight_elem_size, weight_elem_size};
+    size_t weight_stride = src0->ne[1] * src0->ne[0] * weight_elem_size;
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
-    GGML_ASSERT(QK4_0 == QK8_0);
-    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
                          scale_elem_size};
-    size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0;
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
     char* scale_offset = (char*)src0->data + weight_size;
 
     // input
-    void* input_buffer;
     size_t input_elem_size = sizeof(uint16_t);
     int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
-    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
-    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
-
+    size_t input_nb[] = {input_elem_size, input_ne[0] * input_elem_size};
+    size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size;
     ggml_cann_pool_alloc input_alloctor(ctx.pool());
+    void* input_buffer = src1->data;
+
+    // case in
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
-        input_buffer = input_alloctor.get();
+        input_buffer =
+            input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
 
         int64_t* input_cast_ne = src1->ne;
         size_t input_cast_nb[GGML_MAX_DIMS];
@@ -2537,85 +2736,136 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
             input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
             input_cast_nb, GGML_MAX_DIMS);
         aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
+
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         ACL_CHECK(aclDestroyTensor(acl_src1_tensor));
-    } else {
-        input_buffer = src1->data;
     }
 
     // output
     size_t output_elem_size = sizeof(uint16_t);
-    int64_t output_ne[] = {dst->ne[0], dst->ne[1]};
-    size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]};
-    ggml_cann_pool_alloc output_alloctor(
-        ctx.pool(), ggml_nelements(dst) * output_elem_size);
-    void* output_buffer = output_alloctor.get();
-    size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1];
+    size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
+    ggml_cann_pool_alloc output_allocator(ctx.pool());
+    void* output_buffer =
+        output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
+    size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
 
     // aclnn
+    int64_t max_elem_size = 65535;
+    int64_t split_size = (src0->ne[1] / max_elem_size) + 1;
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
+    aclOpExecutor* executor = nullptr;
     uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
-
     for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) {
         for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) {
             int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]);
             int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]);
 
-            int64_t batch1 = n1 * src1->ne[2] + c1;
-            int64_t batch0 = n0 * src0->ne[2] + c0;
+            int64_t batch1 = (n1 * src1->ne[2]) + c1;
+            int64_t batch0 = (n0 * src0->ne[2]) + c0;
 
             aclTensor* acl_input_tensor = ggml_cann_create_tensor(
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
+
+            // first split
+            int64_t weight_ne_offset = 0;
+            int64_t weight_ne[2] = {
+                max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size,
+                src0->ne[0]};
+            int64_t scale_ne_offset = 0;
+            int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
+            int64_t output_ne_offset = 0;
+            int64_t output_ne[2] = {weight_ne[0], dst->ne[1]};
+
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 (char*)src0->data + batch0 * weight_stride,
                 ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
-                weight_nb, 2);
+                weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
-                scale_elem_size, scale_ne, scale_nb, 2);
+                scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                scale_ne_offset);
             aclTensor* acl_output_tensor = ggml_cann_create_tensor(
                 (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
-                output_elem_size, output_ne, output_nb, 2);
+                output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                output_ne_offset);
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                 acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
                 nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
                 &workspaceSize, &executor));
-
-            if (workspaceSize > 0 && workspaceAddr == nullptr) {
-                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
-                                                         workspaceSize);
-                workspaceAddr = workspace_allocator.get();
+            if (workspaceAddr == nullptr) {
+                workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
-
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
                 workspaceAddr, workspaceSize, executor, ctx.stream()));
 
-            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
             ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
             ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
             ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+
+            // other splits
+            for (int64_t split = 1; split < split_size; split++) {
+                weight_ne_offset +=
+                    weight_elem_size * weight_ne[0] * weight_ne[1];
+                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1]
+                                   ? src0->ne[1] - (max_elem_size * split)
+                                   : max_elem_size;
+                scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
+                scale_ne[0] = weight_ne[0];
+                output_ne_offset +=
+                    output_elem_size * output_ne[0] * output_ne[1];
+                output_ne[0] = weight_ne[0];
+
+                acl_weight_tensor = ggml_cann_create_tensor(
+                    (char*)src0->data + batch0 * weight_stride,
+                    ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                    weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
+                acl_scale_tensor = ggml_cann_create_tensor(
+                    scale_offset + batch0 * scale_stride, ACL_FLOAT16,
+                    scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                    scale_ne_offset);
+                acl_output_tensor = ggml_cann_create_tensor(
+                    (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
+                    output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                    output_ne_offset);
+
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
+                    acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                    nullptr, nullptr, nullptr, nullptr, QK8_0,
+                    acl_output_tensor, &workspaceSize, &executor));
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
+                    workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+                ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+            }
+
+            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         }
     }
 
     // cast out
-    int64_t* output_cast_ne = dst->ne;
-    size_t output_cast_nb[GGML_MAX_DIMS];
-    output_cast_nb[0] = sizeof(uint16_t);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t* output_cast_ne = dst->ne;
+        size_t output_cast_nb[GGML_MAX_DIMS];
+        output_cast_nb[0] = sizeof(uint16_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+        }
+
+        aclTensor* acl_output_tensor = ggml_cann_create_tensor(
+            output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne,
+            output_cast_nb, GGML_MAX_DIMS);
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor,
+                   ggml_cann_type_mapping(dst->type));
+
+        ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
     }
-
-    aclTensor* acl_output_tensor =
-        ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size,
-                                output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
-    aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-    aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT);
-
-    ACL_CHECK(aclDestroyTensor(acl_output_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
 }
 
 void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
@@ -2714,12 +2964,14 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                              aclTensor* acl_cos_repeat_tensor,
                              aclTensor* acl_sin_repeat_tensor,
-                             float theta_scale, bool is_neox) {
+                             float theta_scale, float freq_scale,
+                             float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
+    ggml_tensor* src2 = dst->src[2];  // freq_factors
 
     // arange, [0,1,...,ne0/2]
     int64_t arange_length = src0->ne[0] / 2;
@@ -2748,11 +3000,26 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
                                                arange_length * sizeof(float_t));
     void* theta_scale_buffer = theta_scale_allocator.get();
-    aclTensor* acl_theta_scale_tensor = aclnn_ones(
+    aclTensor* acl_theta_scale_tensor = aclnn_values(
         ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne,
         GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale);
     aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor);
 
+    // freq_scale
+    if (freq_scale != 1) {
+        aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
+    }
+
+    // freq_factors
+    if (src2) {
+        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+            src2->data, ggml_cann_type_mapping(src2->type),
+            ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS);
+        aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor,
+                         nullptr, true);
+        ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor));
+    }
+
     // position
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
@@ -2816,6 +3083,12 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor);
 
+    // attn_factor
+    if (attn_factor != 1) {
+        aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
+        aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
+    }
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -2841,15 +3114,27 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ACL_CHECK(aclDestroyTensor(acl_cos_tensor));
 }
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+    const aclTensor* x, const aclTensor* cos, const aclTensor* sin,
+    int64_t mode, const aclTensor* yOut, uint64_t* workspaceSize,
+    aclOpExecutor** executor);
+aclnnStatus aclnnRotaryPositionEmbedding(void* workspace,
+                                         uint64_t workspaceSize,
+                                         aclOpExecutor* executor,
+                                         aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
-    // TODO: with freq_factors
-    GGML_ASSERT(src2 == NULL);
-
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
     // const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -2867,13 +3152,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    GGML_ASSERT(n_dims <= ne0);
+    // TODO: n_dims <= ne0
+    GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
-
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
-    // TODO: freq_scale != 1
-    GGML_ASSERT(freq_scale == 1);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -2904,7 +3187,13 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                     theta_scale, is_neox);
+                     theta_scale, freq_scale, attn_factor, is_neox);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+#ifdef ASCEND_310P
+    // Special ROPE operation for 310P
 
     // roll input
     void* input_roll_buffer;
@@ -2947,7 +3236,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
-        acl_minus_one_tensor = aclnn_ones(
+        acl_minus_one_tensor = aclnn_values(
             ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
             minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
         int64_t dim = 3;
@@ -2974,17 +3263,15 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
         ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
-
         // init [-1, -1, -1, 1, 1，1，...]
         minus_one_scale_buffer = minus_one_scale_allocator.get();
-
         int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
         size_t minus_one_nb[GGML_MAX_DIMS];
         minus_one_nb[0] = sizeof(float_t);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
-        acl_minus_one_tensor = aclnn_ones(
+        acl_minus_one_tensor = aclnn_values(
             ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
             minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
         // -1 * first half
@@ -3026,14 +3313,12 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
               acl_input_roll_mul_scale_tensor);
 
     // output
-    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     void* output_fp32_buffer;
     if (src0->type == GGML_TYPE_F32) {
-        aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor);
+        aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor);
         aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
                           acl_sin_reshape_tensor);
-        aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst);
+        aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst);
         // TODO: ne0 != n_dims in mode2
     } else if (src0->type == GGML_TYPE_F16) {
         size_t input_fp32_nb[GGML_MAX_DIMS];
@@ -3060,7 +3345,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
             output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
             input_fp32_nb, GGML_MAX_DIMS);
-        aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1);
+        aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1);
         aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
                   input_fp32_tensor2);
         aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
@@ -3070,13 +3355,73 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
         ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
         ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_src));
+    }
+    return;
+#endif
+
+    // src0 == GGML_TYPE_F16
+    // TODO: optimization this `if` code
+    if (src0->type == GGML_TYPE_F16) {
+        ggml_cann_pool_alloc sin_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        ggml_cann_pool_alloc cos_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        void* sin_final_buffer = sin_final_allocator.get();
+        void* cos_final_buffer = cos_final_allocator.get();
+
+        int64_t sin_final_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
+        size_t sin_final_nb[GGML_MAX_DIMS];
+        sin_final_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            sin_final_nb[i] = sin_final_nb[i - 1] * sin_final_ne[i - 1];
+        }
+        aclTensor* acl_sin_final_tensor = ggml_cann_create_tensor(
+            sin_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_cos_final_tensor = ggml_cann_create_tensor(
+            cos_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+
+        aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        acl_sin_reshape_tensor = acl_sin_final_tensor;
+        acl_cos_reshape_tensor = acl_cos_final_tensor;
     }
 
-    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+
+    void* workspaceAddr = nullptr;
+
+    int acl_mode = mode;
+    if (mode == 0) {
+        acl_mode = 1;
+    }
+
+    ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+        acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
+        acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
+                                           executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));
 }

ggml/src/ggml-cann/common.h

@@ -211,17 +211,20 @@ struct ggml_cann_pool_alloc {
 struct ggml_backend_cann_context {
     int32_t device;                  /**< Device ID. */
     std::string name;                /**< Name of the device. */
+    std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {
-        {nullptr}}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {}
+        : device(device), name("CANN" + std::to_string(device)) {
+        ggml_cann_set_device(device);
+        description = aclrtGetSocName();
+    }
 
     /**
      * @brief Destructor for cleaning up resources.

ggml/src/ggml-cann/ggml-cann.cpp

@@ -39,6 +39,8 @@
 
 #include "ggml-common.h"
 
+#define GGML_CANN_NAME "CANN"
+
 /**
  * @brief Handles CANN errors by printing an error message and aborting.
  *
@@ -120,6 +122,10 @@ static ggml_cann_device_info ggml_cann_init() {
         ACL_CHECK(aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
             &info.devices[id].vmm_granularity));
+
+        size_t free, total;
+        ggml_backend_cann_get_device_memory(id, &free, &total);
+        info.devices[id].total_vram = free;
     }
 
     // TODO: add more device info later.
@@ -206,6 +212,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
+        const size_t alignment = 128;
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 #ifdef DEBUG_CANN_MALLOC
         int nnz = 0;
         size_t max_size = 0;
@@ -244,13 +255,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
             return ptr;
         }
         void* ptr;
-        size_t look_ahead_size = (size_t)(1.05 * size);
-        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
         ggml_cann_set_device(device);
         ACL_CHECK(
-            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = look_ahead_size;
-        pool_size += look_ahead_size;
+            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
 #ifdef DEBUG_CANN_MALLOC
         GGML_LOG_INFO(
             "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
@@ -294,7 +303,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
     /**
      * @brief The maximum size of the virtual memory pool (32 GB).
      */
-    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+    size_t max_size;
 
     /**
      * @brief The device ID associated with this buffer pool.
@@ -339,7 +348,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      */
     explicit ggml_cann_pool_vmm(int device)
         : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+        auto dev = ggml_cann_info().devices[device];
+        granularity = dev.vmm_granularity;
+        max_size = dev.total_vram;
+    }
 
     /**
      * @brief Destructor to free all buffers in the virtual memory pool.
@@ -368,17 +381,19 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
         // round up the allocation size to the alignment to ensure that all
         // allocations are aligned for all data types
         const size_t alignment = 128;
-        size = alignment * ((size + alignment - 1) / alignment);
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 
         size_t avail = pool_size - pool_used;
 
         if (size > avail) {
             // round up to the next multiple of the granularity
             size_t reserve_size = size - avail;
-            reserve_size =
-                granularity * ((reserve_size + granularity - 1) / granularity);
+            reserve_size = GGML_PAD(reserve_size, granularity);
 
-            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+            GGML_ASSERT(pool_size + reserve_size <= max_size);
 
             // allocate more physical memory
             aclrtPhysicalMemProp prop = {};
@@ -394,7 +409,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // reserve virtual address space (if not already reserved)
             if (pool_addr == 0) {
                 ACL_CHECK(aclrtReserveMemAddress(
-                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+                    &pool_addr, max_size, 0, NULL, 1));
             }
 
             // map at the end of the pool
@@ -407,10 +422,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // add to the pool
             pool_size += reserve_size;
 
-            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
-            // reserved %llu MB)\n",
-            //       device, (unsigned long long) (pool_size/1024/1024),
-            //       (unsigned long long) (reserve_size/1024/1024));
+#ifdef DEBUG_CANN_MALLOC
+             GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+                   device, (unsigned long long) (pool_size/1024/1024),
+                   (unsigned long long) (reserve_size/1024/1024));
+#endif
         }
 
         GGML_ASSERT(pool_addr != 0);
@@ -455,7 +471,6 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
 }
 
@@ -487,23 +502,6 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
-/**
- * @brief Retrieve the name associated with a CANN buffer.
- *
- * This function returns the name of a CANN buffer, which is stored in the
- * context of the buffer.
- *
- * @param buffer The CANN buffer whose name is to be retrieved.
- * @return A pointer to a C-string containing the name of the buffer.
- */
-
-static const char* ggml_backend_cann_buffer_get_name(
-    ggml_backend_buffer_t buffer) {
-    return "CANN";
-
-    GGML_UNUSED(buffer);
-}
-
 /**
  * @brief Check if a buffer is a CANN buffer.
  *
@@ -513,9 +511,10 @@ static const char* ggml_backend_cann_buffer_get_name(
  * @param buffer The buffer to check.
  * @return true if the buffer is a CANN buffer, false otherwise.
  */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
 static bool ggml_backend_buffer_is_cann(
     ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
+    return ggml_backend_buft_is_cann(buffer->buft);
 }
 
 /**
@@ -851,13 +850,6 @@ static void ggml_backend_cann_buffer_set_tensor(
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
                               transform_buffer, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
@@ -969,8 +961,7 @@ static void ggml_backend_cann_buffer_clear(
  * This structure defines function pointers to operations that can be performed
  * on a CANN buffer within the backend.
  */
-static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
-    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
+static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
@@ -1004,9 +995,10 @@ struct ggml_backend_cann_buffer_type_context {
  */
 static const char* ggml_backend_cann_buffer_type_name(
     ggml_backend_buffer_type_t buft) {
-    return "CANN";
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
 
-    GGML_UNUSED(buft);
+    return buft_ctx->name.c_str();
 }
 
 /**
@@ -1105,19 +1097,25 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
 /**
  * @brief Interface for managing CANN buffer types in the GGML backend.
  *
  * Provides function pointers for allocating, querying properties, and managing
  * memory for CANN buffer types in the GGML backend.
  */
-static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cann_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
-    /* .is_host          = */ NULL,
+    /* .is_host          = */ ggml_backend_cann_buffer_type_is_host,
 };
 
 /**
@@ -1145,9 +1143,10 @@ ggml_backend_cann_buffer_type(int32_t device) {
     static bool ggml_backend_cann_buffer_type_initialized = false;
 
     if (!ggml_backend_cann_buffer_type_initialized) {
-        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+        for (int32_t i = 0; i < ggml_cann_info().device_count; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), i),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1213,10 +1212,15 @@ static void * ggml_cann_host_malloc(size_t size) {
         return nullptr;
     }
 
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
+
     void * hostPtr = nullptr;
     aclError err = aclrtMallocHost((void **) &hostPtr, size);
     if (err != ACL_SUCCESS) {
-
         GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
                            size / 1024.0 / 1024.0, aclGetRecentErrMsg());
         return nullptr;
@@ -1241,7 +1245,6 @@ static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggm
 
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
     buffer->buft = buft;
-    buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
 
     return buffer;
@@ -1263,7 +1266,7 @@ ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
-        /* .device   = */ nullptr,
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
         /* .context  = */ nullptr,
     };
 
@@ -1463,24 +1466,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
-/**
- * @brief Retrieves the default buffer type associated with the CANN backend.
- *
- * This function returns the buffer type specific to the device associated
- * with the CANN backend. It is used to allocate buffers for computations
- * performed by the backend.
- *
- * @param backend Pointer to the CANN backend structure.
- * @return Pointer to the buffer type structure for the CANN backend.
- */
-static ggml_backend_buffer_type_t
-ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    return ggml_backend_cann_buffer_type(cann_ctx->device);
-}
-
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
@@ -1510,13 +1495,6 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);
 
-#ifndef NDEBUG
-        void *check_buffer = malloc(size);
-        ggml_backend_cann_transform_back(tensor, transform_buffer,
-                                         check_buffer);
-        GGML_ASSERT(memcmp(data, check_buffer, size));
-        free(check_buffer);
-#endif
         ACL_CHECK(aclrtMemcpyAsync(
             (char *)tensor->data + offset, size, transform_buffer, size,
             ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
@@ -1691,7 +1669,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(
  * @return bool Returns true if the operation is supported by the backend,
  *              otherwise false.
  */
-static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                                                     const ggml_tensor* op) {
     switch (op->op) {
         case GGML_OP_UNARY:
@@ -1709,12 +1687,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
+                case GGML_TYPE_Q8_0:
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
+                    if (op->src[0]->ne[0] <= QK8_0) {
+                        return false;
+                    }
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
-                case GGML_TYPE_Q8_0:
-                    // TODO: fix me
-                    // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
                 case GGML_TYPE_Q4_0:
                     return true;
                 default:
@@ -1746,9 +1726,50 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
                     return false;
             }
         }
+        case GGML_OP_CONT: {
+            // TODO: support GGML_TYPE_BF16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_ROPE: {
+            // TODO: with ops-test v == 1
+            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            // TODO: n_dims <= ne0
+            if (op->src[0]->ne[0] != op->op_params[1]) {
+                return false;
+            }
+            // TODO: ext_factor != 0
+            if (*ext_factor != 0) {
+                return false;
+            }
+
+            const int mode = ((const int32_t *) op->op_params)[2];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                return false;
+            }
+            if (mode & GGML_ROPE_TYPE_VISION) {
+                return false;
+            }
+
+            return true;
+        }
+        case GGML_OP_UPSCALE: {
+            // aclnnUpsampleNearest2dGetWorkspaceSize not support
+            // selfDimN[2]/outDimN[2] or selfDimC[3]/outDimC[3] not equal
+            if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
+                return false;
+            }
+            return true;
+        }
+        case GGML_OP_IM2COL:
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -1762,17 +1783,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -1782,7 +1799,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
             return false;
     }
 
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 }
 
 /**
@@ -1800,31 +1817,6 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
-/**
- * @brief Checks if the CANN backend supports a specific backend buffer type.
- *
- * This function determines whether the CANN backend supports the given backend
- * buffer type by comparing the device context of the backend and buffer type.
- * It returns true if the devices are same between the backend context and
- * buffer type context.
- *
- * @param backend Pointer to the CANN backend.
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the CANN backend supports the buffer type,
- *              otherwise false.
- */
-static bool ggml_backend_cann_supports_buft(
-    ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
-    if (ggml_backend_buft_is_cann(buft)) {
-        ggml_backend_cann_context * cann_ctx =
-                        (ggml_backend_cann_context *)backend->context;
-        ggml_backend_cann_buffer_type_context * buft_ctx =
-                        (ggml_backend_cann_buffer_type_context *)buft->context;
-        return buft_ctx->device == cann_ctx->device;
-    }
-    return false;
-}
-
 /**
  * @brief Determines if a tensor operation should be offloaded to the CANN
  * backend.
@@ -1839,54 +1831,14 @@ static bool ggml_backend_cann_supports_buft(
  * @return bool Returns true if the operation should be offloaded, otherwise
  * false.
  */
-static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
                                                    const ggml_tensor* op) {
     const int min_batch_size = 32;
-    GGML_UNUSED(backend);
+    GGML_UNUSED(dev);
 
     return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
 }
 
-/**
- * @brief Creates a new event for the CANN backend.
- *
- * This function initializes a new event for the CANN backend by setting the
- * device and creating an ACL runtime event. The created event is then wrapped
- * in a ggml_backend_event structure and returned.
- *
- * @param backend Pointer to the CANN backend.
- * @return ggml_backend_event_t Returns a pointer to the new event structure.
- */
-static ggml_backend_event_t ggml_backend_cann_event_new(
-    ggml_backend_t backend) {
-    ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)backend->context;
-
-    ggml_cann_set_device(cann_ctx->device);
-
-    aclrtEvent event;
-    ACL_CHECK(aclrtCreateEvent(&event));
-
-    return new ggml_backend_event{
-        /* .backend = */ backend,
-        /* .context = */ event,
-    };
-}
-
-/**
- * @brief Frees a CANN backend event.
- *
- * This function destroys the ACL runtime event associated with the given CANN
- * backend event and then deletes the event structure itself.
- *
- * @param event Pointer to the event structure to be freed.
- */
-static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
-
-    delete event;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *
@@ -1895,10 +1847,9 @@ static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
  *
  * @param event Pointer to the event structure to be recorded.
  */
-static void ggml_backend_cann_event_record(ggml_backend_event_t event) {
+static void ggml_backend_cann_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
     ggml_backend_cann_context* cann_ctx =
-        (ggml_backend_cann_context*)event->backend->context;
-
+        (ggml_backend_cann_context*)backend->context;
     ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
 }
 
@@ -1916,8 +1867,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
                                          ggml_backend_event_t event) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
-    if (ggml_backend_is_cann(event->backend)) {
+    if (ggml_backend_is_cann(backend)) {
         ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
                                        (aclrtEvent)event->context));
     } else {
@@ -1925,17 +1875,6 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
     }
 }
 
-/**
- * @brief Synchronizes the given event on the CANN backend.
- *
- * This function waits for the specified event to complete on the ACL runtime.
- *
- * @param event Pointer to the event structure to be synchronized.
- */
-static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
-    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
-}
-
 /**
  * @brief Structure defining the interface for the CANN backend.
  *
@@ -1943,10 +1882,9 @@ static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
  * supported by the CANN backend, including name retrieval, memory
  * management, tensor operations, synchronization, and event handling.
  */
-static ggml_backend_i ggml_backend_cann_interface = {
+static const ggml_backend_i ggml_backend_cann_interface = {
     /* .get_name                = */ ggml_backend_cann_name,
     /* .free                    = */ ggml_backend_cann_free,
-    /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
     /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
@@ -1956,9 +1894,6 @@ static ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
-    /* .supports_op             = */ ggml_backend_cann_supports_op,
-    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
-    /* .offload_op              = */ ggml_backend_cann_offload_op,
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
 };
@@ -1977,6 +1912,235 @@ static ggml_guid_t ggml_backend_cann_guid() {
     return &guid;
 }
 
+// backend device
+struct ggml_backend_cann_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    ggml_backend_cann_get_device_memory(ctx->device, free, total);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cann_device_get_name(dev);
+    props->description = ggml_backend_cann_device_get_description(dev);
+    props->type        = ggml_backend_cann_device_get_type(dev);
+    ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
+
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ host_buffer,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ true,
+    };
+}
+
+static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_init(ctx->device);
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_buft(
+    ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+    return false;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
+    return ggml_backend_cann_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cann_host_buffer_type();
+}
+
+/**
+ * @brief Creates a new event for the CANN backend device.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_device_event_new(
+    ggml_backend_dev_t dev) {
+    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+
+    ggml_cann_set_device(dev_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+    GGML_UNUSED(dev);
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+
+    GGML_UNUSED(dev);
+}
+
+static const ggml_backend_device_i ggml_backend_cann_device_interface = {
+    /* .get_name                = */ ggml_backend_cann_device_get_name,
+    /* .get_description         = */ ggml_backend_cann_device_get_description,
+    /* .get_memory              = */ ggml_backend_cann_device_get_memory,
+    /* .get_type                = */ ggml_backend_cann_device_get_type,
+    /* .get_props               = */ ggml_backend_cann_device_get_props,
+    /* .init_backend            = */ ggml_backend_cann_device_init,    // called for every card
+    /* .get_buffer_type         = */ ggml_backend_cann_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cann_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ NULL, // not supported for CANN
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_new               = */ ggml_backend_cann_device_event_new,
+    /* .event_free              = */ ggml_backend_cann_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cann_device_event_synchronize,
+};
+
+
+// backend reg
+struct ggml_backend_cann_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CANN_NAME;
+}
+
+static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    return ctx->devices.size();
+}
+
+static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
+
+static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    GGML_UNUSED(name);
+    // reserved for future use
+    return nullptr;
+}
+
+static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
+    /* .get_name          = */ ggml_backend_cann_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
+    /* .get_device        = */ ggml_backend_cann_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
+};
+
+// backend registry, called only once for cann backend
+ggml_backend_reg_t ggml_backend_cann_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            aclInit(nullptr);
+            ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
+
+            for (int i = 0; i < ggml_cann_info().device_count; i++) {
+                ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
+                dev_ctx->description = aclrtGetSocName();
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
+                ggml_cann_set_device(i);
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .iface   = */ ggml_backend_cann_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cann_reg_interface,
+                /* .context     = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
+}
+
 ggml_backend_t ggml_backend_cann_init(int32_t device) {
     aclInit(nullptr);
     if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
@@ -1993,7 +2157,7 @@ ggml_backend_t ggml_backend_cann_init(int32_t device) {
     ggml_backend_t cann_backend =
         new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
                          /* .interface = */ ggml_backend_cann_interface,
-                         /* .device    = */ nullptr,
+                         /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
                          /* .context   = */ ctx};
 
     return cann_backend;
@@ -2020,3 +2184,5 @@ void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
     ggml_cann_set_device(device);
     ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cann_reg)

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -1,7 +1,3 @@
-if (NOT SOC_TYPE)
-    set (SOC_TYPE "Ascend910B3")
-endif()
-
 file(GLOB SRC_FILES
     get_row_f32.cpp
     get_row_f16.cpp
@@ -13,7 +9,6 @@ file(GLOB SRC_FILES
     dup.cpp
 )
 
-string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
 set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
 
@@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
+message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
+ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/dup.cpp

@@ -5,6 +5,7 @@
 using namespace AscendC;
 
 #define BUFFER_NUM 2
+const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
 
 template <typename SRC_T, typename DST_T>
 class DupByRows {
@@ -51,24 +52,36 @@ class DupByRows {
 
     __aicore__ inline void copy_in() {
         LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
-        DataCopyPadExtParams<SRC_T> padParams;
-        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
-
+        const size_t elem_per_block = 32 / sizeof(SRC_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
+        DataCopy(src_local, src_gm, cpy_elements_len);
         src_queue.EnQue(src_local);
     }
 
     __aicore__ inline void copy_out() {
         LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-
+#ifdef ASCEND_310P
+        const size_t elem_per_block = 32 / sizeof(DST_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t len = num_elem & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(dst_gm, dst_local, len);
+        }
+        if(tail != 0) {
+            for (size_t i = tail; i < elem_per_block; i++) {
+                dst_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
+            SetAtomicNone();
+        }
+#else
         DataCopyExtParams dataCopyParams;
         dataCopyParams.blockCount = 1;
         dataCopyParams.blockLen = num_elem * sizeof(DST_T);
         DataCopyPad(dst_gm, dst_local, dataCopyParams);
-
+#endif
         dst_queue.FreeTensor(dst_local);
     }