release : v1.7.2

* Fixes broken build for the SYCL CUDA backend caused by non-explicit gemm call in outprod (merged in with RWKV6 in
Optimize RWKV6 Operator Naming and Implement Multi-core CPU/ SYCL Acceleration #10133)

* Marks permuted MUL_MAT as unsupported to be able to run test-backend-ops

* Fixes asserts in norm to fix debug builds.

.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
+    apt-get install -y build-essential git cmake libsdl2-dev
 
 WORKDIR /app
 
@@ -21,7 +21,7 @@ COPY . .
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV WHISPER_CUBLAS=1
+ENV GGML_CUDA=1
 
 RUN make
 

.devops/main-cuda.Dockerfile

@@ -14,10 +14,10 @@ ARG CUDA_DOCKER_ARCH=all
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV WHISPER_CUBLAS=1
+ENV GGML_CUDA=1
 
 RUN apt-get update && \
-    apt-get install -y build-essential \
+    apt-get install -y build-essential libsdl2-dev \
     && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 # Ref: https://stackoverflow.com/a/53464012

.devops/main.Dockerfile

@@ -12,7 +12,7 @@ FROM ubuntu:22.04 AS runtime
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
+  apt-get install -y curl ffmpeg libsdl2-dev \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app

.github/workflows/bindings-go.yml

@@ -13,10 +13,10 @@ jobs:
   ubuntu-latest:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/setup-go@v3
+      - uses: actions/setup-go@v5
         with:
-          go-version: '^1.19'
-      - uses: actions/checkout@v1
+          go-version: '^1.23'
+      - uses: actions/checkout@v4
       - run: |
           cd bindings/go
           make test

.github/workflows/bindings-ruby.yml

@@ -3,20 +3,73 @@ on:
   push:
     paths:
       - bindings/ruby/**
-      - whisper.h
+      - 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
   pull_request:
     paths:
       - bindings/ruby/**
-      - whisper.h
+      - 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
 
 jobs:
   ubuntu-latest:
     runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: bindings/ruby
     steps:
       - uses: ruby/setup-ruby@v1
         with:
           ruby-version: '3.0'
-      - uses: actions/checkout@v1
-      - run: |
-          cd bindings/ruby/ext
-          ruby extconf.rb && make
+      - 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:
@@ -15,10 +16,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -36,7 +37,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         run: |
@@ -53,13 +54,13 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
-        uses: cross-platform-actions/action@v0.15.0
+        uses: cross-platform-actions/action@v0.24.0
         with:
           operating_system: freebsd
-          version: '13.2'
+          version: '13.3'
           run: |
             sudo pkg update
             sudo pkg install -y gmake sdl2
@@ -77,10 +78,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -101,14 +102,17 @@ jobs:
       fail-fast: false
       matrix:
         build: [Debug, Release]
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        #arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        # TODO: arm/v7 disabled due to clang bug
+        #       https://github.com/ggerganov/whisper.cpp/actions/runs/9657764109/job/26637633042?pr=2256#step:4:1990
+        arch: [linux/amd64, linux/arm64, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -133,10 +137,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -165,7 +169,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -189,7 +193,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -197,7 +201,7 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
   ubuntu-22-cmake-sycl-fp16:
@@ -215,7 +219,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -239,7 +243,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -247,7 +251,7 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DGGML_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
   windows-msys2:
@@ -262,7 +266,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup ${{ matrix.sys }}
         uses: msys2/setup-msys2@v2
@@ -289,7 +293,7 @@ jobs:
       - name: Build using make w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            make WHISPER_OPENBLAS=1 -j $(nproc)
+            make GGML_OPENBLAS=1 -j $(nproc)
 
       - name: Build using CMake
         shell: msys2 {0}
@@ -305,7 +309,7 @@ jobs:
       - name: Build using CMake w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            cmake -B build -DWHISPER_OPENBLAS=ON
+            cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
@@ -328,10 +332,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1
+        uses: microsoft/setup-msbuild@v2
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -356,14 +360,14 @@ jobs:
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
       - name: Upload dll
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.jnaPath }}_whisper.dll
           path: build/bin/${{ matrix.build }}/whisper.dll
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v1
+        uses: actions/upload-artifact@v4
         with:
           name: whisper-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
@@ -379,32 +383,31 @@ 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
-            clblast: OFF
           - arch: x64
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
-            clblast: ON
-            clver: 1.6.1
           - sdl2: ON
             s2ver: 2.28.5
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        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@v1
+        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'
@@ -413,53 +416,37 @@ jobs:
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
-      - name: Install OpenCL
-        if: matrix.clblast == 'ON'
-        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
-
-      - name: Fetch CLBlast and set CLBlast_DIR
-        if: matrix.clblast == 'ON'
-        run: |
-          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
-          7z x clblast.zip
-          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
-          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
-
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_TOOLCHAIN_FILE="$env:VCPKG_INSTALLATION_ROOT/scripts/buildsystems/vcpkg.cmake"
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
+          -DGGML_BLAS=${{ matrix.blas }}
+          -DGGML_BLAS_VENDOR=OpenBLAS
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
-          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
           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'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
-      - name: Copy clblast.dll
-        if: matrix.clblast == 'ON'
-        run: copy "$env:CLBlast_DIR/../../clblast.dll" build/bin/${{ matrix.build }}
-
       - name: Upload binaries
         if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v1
+        uses: actions/upload-artifact@v4
         with:
-          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
+          name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   windows-cublas:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     strategy:
       matrix:
@@ -476,14 +463,14 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1
+        uses: microsoft/setup-msbuild@v2
 
       - name: Install CUDA Toolkit
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.11
+        uses: Jimver/cuda-toolkit@v0.2.15
         with:
           cuda: '${{ matrix.cuda-toolkit }}'
 
@@ -498,7 +485,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUBLAS=${{ matrix.cublas }}
+          -DGGML_CUDA=${{ matrix.cublas }}
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build ${{ matrix.cuda-toolkit }}
@@ -519,7 +506,7 @@ jobs:
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v1
+        uses: actions/upload-artifact@v4
         with:
           name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
@@ -533,10 +520,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup emsdk
-        uses: mymindstorm/setup-emsdk@v12
+        uses: mymindstorm/setup-emsdk@v14
 
       - name: Verify
         run: emcc -v
@@ -555,7 +542,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Configure
         run: |
@@ -573,24 +560,18 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           path: whisper
 
-      - name: Clone
-        uses: actions/checkout@v3
-        with:
-          repository: ggerganov/ggml
-          path: ggml
-
       - name: Install Java
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v4
         with:
           distribution: zulu
-          java-version: 17
+          java-version: 21
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
+        uses: android-actions/setup-android@v3
 
       - name: Build
         run: |
@@ -601,82 +582,84 @@ 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
 
-  android_java:
-    runs-on: ubuntu-latest
+# TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602
+#  android_java:
+#    runs-on: ubuntu-latest
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: set up JDK 11
+#        uses: actions/setup-java@v4
+#        with:
+#          java-version: '11'
+#          distribution: 'temurin'
+#          cache: gradle
+#
+#      - name: Setup Android SDK
+#        uses: android-actions/setup-android@v3
+#        with:
+#          cmdline-tools-version: 9.0
+#
+#      - name: Build
+#        run: |
+#          cd examples/whisper.android.java
+#          chmod +x ./gradlew
+#          ./gradlew assembleRelease
 
-    steps:
-      - name: Clone
-        uses: actions/checkout@v3
-
-      - name: set up JDK 11
-        uses: actions/setup-java@v3
-        with:
-          java-version: '11'
-          distribution: 'temurin'
-          cache: gradle
-
-      - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
-        with:
-          api-level: 30
-          build-tools-version: 30.0.3
-
-      - name: Build
-        run: |
-          cd examples/whisper.android.java
-          chmod +x ./gradlew
-          ./gradlew assembleRelease
-
-  java:
-    needs: [ 'windows' ]
-    runs-on: windows-latest
-    steps:
-      - uses: actions/checkout@v3
-
-      - name: Install Java
-        uses: actions/setup-java@v1
-        with:
-          java-version: 17
-
-      - name: Download Windows lib
-        uses: actions/download-artifact@v3
-        with:
-          name: win32-x86-64_whisper.dll
-          path: bindings/java/build/generated/resources/main/win32-x86-64
-
-      - name: Build
-        run: |
-          models\download-ggml-model.cmd tiny.en
-          cd bindings/java
-          chmod +x ./gradlew
-          ./gradlew build
-
-      - name: Upload jar
-        uses: actions/upload-artifact@v3
-        with:
-          name: whispercpp.jar
-          path: bindings/java/build/libs/whispercpp-*.jar
-
-      - name: Publish package
-        if: ${{ github.ref == 'refs/heads/master' }}
-        uses: gradle/gradle-build-action@v2.4.2
-        with:
-          arguments: publish
-          build-root-directory: bindings/java
-        env:
-          MAVEN_USERNAME: ${{ secrets.JIRA_USER }}
-          MAVEN_PASSWORD: ${{ secrets.JIRA_PASS }}
-          PGP_SECRET: ${{ secrets.GPG_PRIVATE_KEY }}
-          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+# TODO: disabled because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/9686220096/job/26735899598
+#  java:
+#    needs: [ 'windows' ]
+#    runs-on: windows-latest
+#    steps:
+#      - uses: actions/checkout@v4
+#
+#      - name: Install Java
+#        uses: actions/setup-java@v4
+#        with:
+#          distribution: zulu
+#          java-version: 20
+#
+#      - name: Download Windows lib
+#        uses: actions/download-artifact@v4
+#        with:
+#          name: win32-x86-64_whisper.dll
+#          path: bindings/java/build/generated/resources/main/win32-x86-64
+#
+#      - name: Build
+#        run: |
+#          models\download-ggml-model.cmd tiny.en
+#          cd bindings/java
+#          chmod +x ./gradlew
+#          ./gradlew build
+#
+#      - name: Upload jar
+#        uses: actions/upload-artifact@v4
+#        with:
+#          name: whispercpp.jar
+#          path: bindings/java/build/libs/whispercpp-*.jar
+#
+#      - name: Publish package
+#        if: ${{ github.ref == 'refs/heads/master' }}
+#        uses: gradle/gradle-build-action@v2.4.2
+#        with:
+#          arguments: publish
+#          build-root-directory: bindings/java
+#        env:
+#          MAVEN_USERNAME: ${{ secrets.JIRA_USER }}
+#          MAVEN_PASSWORD: ${{ secrets.JIRA_PASS }}
+#          PGP_SECRET: ${{ secrets.GPG_PRIVATE_KEY }}
+#          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
 
   quantize:
     runs-on: ubuntu-latest
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Test quantize
         run: |

.github/workflows/docker.yml

@@ -18,7 +18,9 @@ jobs:
       matrix:
         config:
           - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64,linux/arm64" }
-          - { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
+          #TODO: the cuda image keeps failing - disable for now
+          #      https://github.com/ggerganov/whisper.cpp/actions/runs/11019444428/job/30602020339
+          #- { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
 
     steps:
       - name: Check out the repo
@@ -43,7 +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 }}
 
@@ -52,6 +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 }}

.github/workflows/examples.yml

@@ -37,7 +37,7 @@ jobs:
         run: npm install
 
       - name: Compile addon.node
-        run: npx cmake-js compile -T whisper-addon -B Release
+        run: npx cmake-js compile -T addon.node -B Release
 
       - name: Download test model
         run: |

.gitignore

@@ -3,22 +3,16 @@
 .cache/
 .coreml/
 .test/
+.venv/
 .vs/
 .vscode/
 .DS_Store
 .vimspector.json
+/CMakeSettings.json
+/talk-llama.dSYM/
 
 build/
-build-coreml/
-build-em/
-build-debug/
-build-release/
-build-rwdi/
-build-static/
-build-cublas/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build-*/
 
 # SPM
 .build/

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "bindings/ios"]
-	path = bindings/ios
-	url = https://github.com/ggerganov/whisper.spm

CMakeLists.txt

@@ -1,25 +1,31 @@
-cmake_minimum_required (VERSION 3.5)
+cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
+project("whisper.cpp" C CXX)
+project("whisper.cpp" VERSION 1.7.2)
+include(CheckIncludeFileCXX)
 
-# Allow for the creation of solution folders.
-set_property(GLOBAL PROPERTY USE_FOLDERS ON)
-
-project(whisper.cpp VERSION 1.5.5)
 set(SOVERSION 1)
 
+#set(CMAKE_WARN_DEPRECATED YES)
+set(CMAKE_WARN_UNUSED_CLI YES)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(WHISPER_STANDALONE ON)
-    include(GitVars)
-    include(BuildTypes)
+
+    include(git-vars)
 
     # configure project version
-    if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
-        configure_file(${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl ${CMAKE_SOURCE_DIR}/bindings/ios/Makefile @ONLY)
-    endif()
     configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
 else()
     set(WHISPER_STANDALONE OFF)
@@ -29,6 +35,11 @@ if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
     option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
+
+    # TODO: without these, we get the following error:
+    #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
 else()
     if (MINGW)
         set(BUILD_SHARED_LIBS_DEFAULT OFF)
@@ -37,723 +48,136 @@ else()
     endif()
 endif()
 
-# options
+option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 
-if (APPLE)
-    set(WHISPER_METAL_DEFAULT ON)
-else()
-    set(WHISPER_METAL_DEFAULT OFF)
-endif()
+#
+# option list
+#
 
-option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
+# general
+option(WHISPER_CCACHE "whisper: use ccache if available" ON)
 
+# debug
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
 option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
 
-option(WHISPER_SANITIZE_THREAD        "whisper: enable thread sanitizer"    OFF)
-option(WHISPER_SANITIZE_ADDRESS       "whisper: enable address sanitizer"   OFF)
-option(WHISPER_SANITIZE_UNDEFINED     "whisper: enable undefined sanitizer" OFF)
-
-option(WHISPER_BUILD_TESTS            "whisper: build tests"    ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDALONE})
-
-option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
-
-option(WHISPER_NO_AVX                 "whisper: disable AVX"         OFF)
-option(WHISPER_NO_AVX2                "whisper: disable AVX2"        OFF)
-option(WHISPER_NO_AVX512              "whisper: disable AVX512"      ON)
-option(WHISPER_NO_AVX512_VBMI         "whisper: disable AVX512-VBMI" ON)
-option(WHISPER_NO_AVX512_VNNI         "whisper: disable AVX512-VNNI" ON)
-option(WHISPER_NO_FMA                 "whisper: disable FMA"         OFF)
-option(WHISPER_NO_F16C                "whisper: disable F16c"        OFF)
-
-option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
-
-if (APPLE)
-    option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
-    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
-    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
-    option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
-    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
-    option(WHISPER_METAL_EMBED_LIBRARY   "whisper: embed Metal library"          OFF)
-else()
-    option(WHISPER_BLAS                  "whisper: use BLAS libraries"                        OFF)
-    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor"                       Generic)
-    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"                           OFF)
-    option(WHISPER_OPENBLAS_INTERFACE64  "whisper: use OpenBLAS w/ 64-bit interface"          OFF)
-    option(WHISPER_CUDA                  "whisper: support for CUDA"                          OFF)
-    option(WHISPER_CUBLAS                "whisper: support for CUDA (deprecated)"             OFF)
-    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
-    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               OFF)
-    option(WHISPER_MKL                   "whisper: use Intel Math Kernel Library (MKL)"       OFF)
-    option(WHISPER_SYCL                  "whisper: use SYCL"                                  OFF)
-    option(WHISPER_SYCL_F16              "whisper: use 16 bit floats for sycl calculations"   OFF)
-endif()
-
-option(WHISPER_PERF "whisper: enable perf timings" OFF)
+# build
+option(WHISPER_FATAL_WARNINGS "whisper: enable -Werror flag" OFF)
 
 # sanitizers
+option(WHISPER_SANITIZE_THREAD    "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS   "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED "whisper: enable undefined sanitizer" OFF)
 
-if (NOT MSVC)
-    if (WHISPER_SANITIZE_THREAD)
-        set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
-    endif()
+# extra artifacts
+option(WHISPER_BUILD_TESTS    "whisper: build tests"          ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES "whisper: build examples"       ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_SERVER   "whisper: build server example" ${WHISPER_STANDALONE})
 
-    if (WHISPER_SANITIZE_ADDRESS)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=address -fno-omit-frame-pointer")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
-    endif()
+# 3rd party libs
+option(WHISPER_CURL "whisper: use libcurl to download model from an URL" OFF)
+option(WHISPER_SDL2 "whisper: support for libSDL2" OFF)
 
-    if (WHISPER_SANITIZE_UNDEFINED)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=undefined")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
-    endif()
-endif()
-
-#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
-#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
-
-# dependencies
-
-find_package(Threads REQUIRED)
-
-#compile flag sycl
-if (WHISPER_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
-
-# on APPLE
-if (APPLE)
-    # include Accelerate framework
-    if (NOT WHISPER_NO_ACCELERATE)
-        find_library(ACCELERATE_FRAMEWORK Accelerate)
-
-        if (ACCELERATE_FRAMEWORK)
-            message(STATUS "Accelerate framework found")
-
-            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64)
-        else()
-            message(FATAL_ERROR "Accelerate framework not found")
-        endif()
-    endif()
-
-    if (WHISPER_METAL)
-        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-
-        if (METAL_FRAMEWORK)
-            message(STATUS "Metal framework found")
-
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
-                ${FOUNDATION_LIBRARY}
-                ${METAL_FRAMEWORK}
-                ${METALKIT_FRAMEWORK}
-                )
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
-
-            if (WHISPER_METAL_NDEBUG)
-                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
-            endif()
-        else()
-            message(FATAL_ERROR "Metal framework not found")
-        endif()
-
-        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
-
-        # copy ggml-common.h and ggml-metal.metal to bin directory
-        configure_file(ggml-common.h    bin/ggml-common.h    COPYONLY)
-        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-        if (WHISPER_METAL_EMBED_LIBRARY)
-            enable_language(ASM)
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)
-
-            set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
-            set(COMMON_HEADER "${CMAKE_SOURCE_DIR}/ggml-common.h")
-
-            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-            set(EMBED_METALLIB_SOURCE "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-combined.metal")
-
-            add_custom_command(
-                OUTPUT ${EMBED_METALLIB_SOURCE}
-                COMMAND sed -e "/^#include \\\"ggml-common.h\\\"/r ${COMMON_HEADER}" -e "/^#include \\\"ggml-common.h\\\"/d" ${METALLIB_SOURCE} > ${EMBED_METALLIB_SOURCE}
-                DEPENDS ${METALLIB_SOURCE} ${COMMON_HEADER}
-                COMMENT "Generating combined Metal library for embedding"
-            )
-
-            add_custom_command(
-                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".incbin \\\"${EMBED_METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-                DEPENDS ${EMBED_METALLIB_SOURCE}
-                COMMENT "Generate assembly for embedded Metal library"
-            )
-
-            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-        endif()
-    endif()
-
-    if (WHISPER_COREML)
-        find_library(FOUNDATION_FRAMEWORK Foundation)
-        find_library(COREML_FRAMEWORK CoreML)
-
-        if (COREML_FRAMEWORK)
-            message(STATUS "CoreML framework found")
-
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
-        else()
-            message(FATAL_ERROR "CoreML framework not found")
-        endif()
-
-        if (WHISPER_COREML_ALLOW_FALLBACK)
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
-        endif()
-    endif()
-endif()
-
-if (WHISPER_OPENBLAS)
-    set(WHISPER_BLAS_VENDOR "OpenBLAS")
-    set(WHISPER_BLAS ON)
-    # BLA_PKGCONFIG_BLAS is supported since CMake 3.25.
-    # FindBLAS.cmake pkg-config logic seems incomplete, because when
-    # BLA_SIZEOF_INTEGER is 8, then it should search for blas64 instead of blas.
-    # blas.pc/blas64.pc are not always provided, so let's be more specific
-    # and go with openblas.pc/openblas64.pc if WHISPER_OPENBLAS is on.
-    if (WHISPER_OPENBLAS_INTERFACE64)
-        set(WHISPER_BLAS_LIB "openblas64")
-    else ()
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
-    set(BLA_PKGCONFIG_BLAS ${WHISPER_BLAS_LIB})
-    # OpenBLAS prebuilt libraries for Windows do not have "64" suffix in filename.
-    # (But .pc file has "64" suffix in filename for USE_64BITINT=1 Windows build.)
-    if (MSVC)
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
-endif()
-
-if (WHISPER_BLAS)
-    if (NOT "$ENV{OPENBLAS_PATH}" STREQUAL "")
-        if (WHISPER_STATIC)
-            set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_STATIC_LIBRARY_PREFIX})
-            set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_STATIC_LIBRARY_SUFFIX})
-        else ()
-            if (CMAKE_IMPORT_LIBRARY_SUFFIX)
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_IMPORT_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_IMPORT_LIBRARY_SUFFIX})
-            else ()
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_SHARED_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_SHARED_LIBRARY_SUFFIX})
-            endif ()
-        endif ()
-        # OpenBLAS prebuilt libraries hardcode "lib" prefix in filename even on Windows
-        if (WHISPER_OPENBLAS)
-            set(WHISPER_BLAS_LIB_PREFIX "lib")
-        endif ()
-        message(STATUS "BLAS compatible library path provided")
-        set(BLAS_LIBRARIES "$ENV{OPENBLAS_PATH}/lib/${WHISPER_BLAS_LIB_PREFIX}${WHISPER_BLAS_LIB}${WHISPER_BLAS_LIB_SUFFIX}")
-        message(STATUS "Libraries ${BLAS_LIBRARIES}")
-        set(BLAS_INCLUDE_DIRS "$ENV{OPENBLAS_PATH}/include")
-        message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
-        if (NOT EXISTS "${BLAS_LIBRARIES}")
-            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH misdefined.")
-        endif ()
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-        include_directories(${BLAS_INCLUDE_DIRS})
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
-    else ()
-        if (WHISPER_STATIC)
-            # FindBLAS.cmake pkg-config logic seems incomplete, because when
-            # BLA_STATIC is on, then it should use pkg_check_modules_static
-            # instead of pkg_check_modules.
-            # Some manual variable overriding may be necessary if you don't
-            # achieve desired results.
-            set(BLA_STATIC 1)
-        endif ()
-        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-        if (WHISPER_OPENBLAS_INTERFACE64)
-            set(BLA_SIZEOF_INTEGER 8)
-        else ()
-            set(BLA_SIZEOF_INTEGER 4)
-        endif()
-        set(BLA_PREFER_PKGCONFIG 1)
-        find_package(BLAS)
-
-        if(BLAS_FOUND)
-            message(STATUS "BLAS compatible library found")
-            message(STATUS "Libraries ${BLAS_LIBRARIES}")
-            if (NOT DEFINED BLAS_INCLUDE_DIRS)
-                if (PKGC_BLAS_FOUND)
-                    set(BLAS_INCLUDE_DIRS "${PKGC_BLAS_INCLUDE_DIRS}")
-                else ()
-                    find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas)
-                endif()
-            endif()
-            message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-            include_directories(${BLAS_INCLUDE_DIRS})
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
-        else()
-            message(FATAL_ERROR "BLAS library was not found")
-        endif()
-    endif ()
-endif ()
-
-if (WHISPER_MKL)
-    find_package(MKL CONFIG REQUIRED PATHS $ENV{MKLROOT})
-    message(STATUS "Imported oneMKL targets: ${MKL_IMPORTED_TARGETS}")
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_BLAS_USE_MKL)
-endif()
-
-if (WHISPER_CUBLAS)
-    message(WARNING "WHISPER_CUBLAS is deprecated and will be removed in the future.\nUse WHISPER_CUDA instead")
-    set(WHISPER_CUDA ON)
-endif()
-
-if (WHISPER_CUDA)
-    cmake_minimum_required(VERSION 3.17)
-
-    find_package(CUDAToolkit)
-
-    if (CUDAToolkit_FOUND)
-        message(STATUS "cuBLAS found")
-
-        enable_language(CUDA)
-
-        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.h)
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.cu)
-
-        add_compile_definitions(GGML_USE_CUDA)
-
-        if (WHISPER_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            else ()
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
-            endif()
-        else()
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
-        endif()
-
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
-    else()
-        message(FATAL_ERROR "cuBLAS not found")
-    endif()
-endif()
-
-
-if (WHISPER_HIPBLAS)
-    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
-    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
-    endif()
-    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-    endif()
-
-    find_package(hip)
-    find_package(hipblas)
-    find_package(rocblas)
-
-    if (${hipblas_FOUND} AND ${hip_FOUND})
-        message(STATUS "HIP and hipBLAS found")
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
-        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
-        set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
-        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
-        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
-
-        if (WHISPER_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
-    else()
-        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
-    endif()
-endif()
-
-if (WHISPER_CLBLAST)
-    find_package(CLBlast)
-    if (CLBlast_FOUND)
-        message(STATUS "CLBlast found")
-
-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
-
-        add_compile_definitions(GGML_USE_CLBLAST)
-
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
-    else()
-        message(FATAL_ERROR "CLBlast not found")
-    endif()
-endif()
-
-if( WHISPER_OPENVINO )
-    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
-endif()
-
-if (WHISPER_SYCL)
-    if ( NOT DEFINED ENV{ONEAPI_ROOT})
-        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
-    endif()
-    #todo: AOT
-
-    find_package(IntelSYCL REQUIRED)
-    if (WHISPER_SYCL_F16)
-        add_compile_definitions(GGML_SYCL_F16)
-    endif()
-    add_compile_definitions(GGML_USE_SYCL)
-
-    add_compile_options(-I./) #include DPCT
-    add_compile_options(-I/${SYCL_INCLUDE_DIR})
-
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
-
-    set(GGML_HEADERS_SYCL ggml-sycl.h)
-    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
-
-    set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-endif()
-# compiler flags
-
-if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
-    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
-    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
-endif ()
-
-if (WHISPER_ALL_WARNINGS)
-    if (NOT MSVC)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} \
-            -Wall                           \
-            -Wextra                         \
-            -Wpedantic                      \
-            -Wshadow                        \
-            -Wcast-qual                     \
-            -Wstrict-prototypes             \
-            -Wpointer-arith                 \
-            -Wno-unused-function            \
-        ")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
-            -Wall                           \
-            -Wextra                         \
-            -Wpedantic                      \
-            -Wcast-qual                     \
-        ")
-    else()
-        # todo : msvc
-    endif()
-endif()
-
-if (NOT MSVC)
-    # TODO: temporary disabled until we figure out ggml-metal.m
-    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
-    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
-endif()
-
-message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
-
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
-    message(STATUS "ARM detected")
-elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-    message(STATUS "PowerPC detected")
-else()
-    message(STATUS "x86 detected")
-    if (MSVC)
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
-        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
-        if(NOT WHISPER_NO_AVX512)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX512")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX512")
-            set(CMAKE_C_FLAGS "${CMAKE_C_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 (NOT WHISPER_NO_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
-            endif()
-            if (NOT WHISPER_NO_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
-            endif()
-        elseif(NOT WHISPER_NO_AVX2)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-        elseif(NOT WHISPER_NO_AVX)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
-        endif()
-    else()
-        if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
-        else()
-            if(NOT WHISPER_NO_AVX)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
-            endif()
-            if(NOT WHISPER_NO_AVX2)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
-            endif()
-            if(NOT WHISPER_NO_AVX512)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw")
-            endif()
-            if(NOT WHISPER_NO_AVX512_VBMI)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vbmi")
-            endif()
-            if(NOT WHISPER_NO_AVX512_VNNI)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vnni")
-            endif()
-            if(NOT WHISPER_NO_FMA)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
-            endif()
-            if(NOT WHISPER_NO_F16C)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
-            endif()
-        endif()
-    endif()
-endif()
-
-#
-# POSIX conformance
-#
-
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-add_compile_definitions(_XOPEN_SOURCE=600)
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    remove_definitions(-D_XOPEN_SOURCE=600)
-    add_compile_definitions(_XOPEN_SOURCE=700)
-endif()
-
-# Data types, macros and functions related to controlling CPU affinity
-# are available on Linux through GNU extensions in libc
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_compile_definitions(_GNU_SOURCE)
+    option(WHISPER_FFMPEG "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
 endif()
 
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
+option(WHISPER_COREML                "whisper: enable Core ML framework"  OFF)
+option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
+option(WHISPER_OPENVINO              "whisper: support for OpenVINO"      OFF)
 
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
-    add_compile_definitions(__BSD_VISIBLE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
-    add_compile_definitions(_NETBSD_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    add_compile_definitions(_BSD_SOURCE)
-endif()
+# Required for relocatable CMake package
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
-if (WHISPER_PERF)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
-endif()
+# override ggml options
+set(GGML_CCACHE             ${WHISPER_CCACHE})
+set(GGML_SANITIZE_THREAD    ${WHISPER_SANITIZE_THREAD})
+set(GGML_SANITIZE_ADDRESS   ${WHISPER_SANITIZE_ADDRESS})
+set(GGML_SANITIZE_UNDEFINED ${WHISPER_SANITIZE_UNDEFINED})
+set(GGML_ALL_WARNINGS       ${WHISPER_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS     ${WHISPER_FATAL_WARNINGS})
 
-#
-# whisper.coreml - Core ML support
-#
-
-if (WHISPER_COREML)
-    set(TARGET whisper.coreml)
-
-    add_library(${TARGET}
-        coreml/whisper-encoder.h
-        coreml/whisper-encoder.mm
-        coreml/whisper-encoder-impl.h
-        coreml/whisper-encoder-impl.m
-        )
-
-    include(DefaultTargetOptions)
-
-    target_include_directories(${TARGET} PUBLIC
-        .
-        )
-
-    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
-
-    set_target_properties(${TARGET} PROPERTIES
-        COMPILE_FLAGS "-fobjc-arc"
-        )
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
-endif()
-
-if (WHISPER_OPENVINO)
-    set(TARGET whisper.openvino)
-
-    add_library(${TARGET} OBJECT
-        openvino/whisper-openvino-encoder.h
-        openvino/whisper-openvino-encoder.cpp
-        )
-
-    target_include_directories(${TARGET} PUBLIC
-        .
-        )
-
-    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
-
-    target_link_libraries(${TARGET} PRIVATE openvino::runtime)
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
-endif()
-
-#
-# whisper - this is the main library of the project
-#
-
-set(TARGET whisper)
-
-add_library(${TARGET}
-    ggml.h
-    ggml.c
-    ggml-alloc.h
-    ggml-alloc.c
-    ggml-backend.h
-    ggml-backend.c
-    ggml-quants.h
-    ggml-quants.c
-    ${GGML_SOURCES_METAL}
-    ${GGML_SOURCES_CUDA}
-    ${GGML_SOURCES_OPENCL}
-    ${GGML_SOURCES_SYCL}
-    ${GGML_HEADERS_SYCL}
-    whisper.h
-    whisper.cpp
-    )
-
-# Set the version numbers
-set_target_properties(whisper PROPERTIES
-    VERSION ${PROJECT_VERSION}
-    SOVERSION ${SOVERSION}
-)
-
-include(DefaultTargetOptions)
-
-target_include_directories(${TARGET} PUBLIC
-    .
-    )
-
-if (WHISPER_COREML)
-    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
-endif()
-
-if (WHISPER_OPENVINO)
-    target_link_libraries(${TARGET} PRIVATE whisper.openvino)
-endif()
-
-if (WHISPER_MKL)
-    target_link_libraries(${TARGET} PUBLIC MKL::MKL)
-endif()
-
-if (MSVC)
-    target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
-
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
-else()
-    target_link_libraries(${TARGET} PRIVATE m ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
-endif()
-
-if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_link_libraries(${TARGET} PUBLIC
-        ${CMAKE_DL_LIBS}
-        )
-
-    target_compile_definitions(${TARGET} PUBLIC
-        WHISPER_SHARED
-        GGML_SHARED
-        )
-
-    target_compile_definitions(${TARGET} PRIVATE
-        WHISPER_BUILD
-        GGML_BUILD
-        )
-
-    if (WHISPER_METAL)
-        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
-        #       but for some reason it does not work here like it does in llama.cpp
-        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+# transition helpers
+function (whisper_option_depr TYPE OLD NEW)
+    if (${OLD})
+        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
+        set(${NEW} ON)
     endif()
+endfunction()
+
+whisper_option_depr(FATAL_ERROR WHISPER_CUBLAS              GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_CUDA                GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_KOMPUTE             GGML_KOMPUTE)
+whisper_option_depr(WARNING     WHISPER_METAL               GGML_METAL)
+whisper_option_depr(WARNING     WHISPER_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
+whisper_option_depr(WARNING     WHISPER_NATIVE              GGML_NATIVE)
+whisper_option_depr(WARNING     WHISPER_OPENMP              GGML_OPENMP)
+whisper_option_depr(WARNING     WHISPER_RPC                 GGML_RPC)
+whisper_option_depr(WARNING     WHISPER_SYCL                GGML_SYCL)
+whisper_option_depr(WARNING     WHISPER_SYCL_F16            GGML_SYCL_F16)
+
+#
+# build the library
+#
+
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+    # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
+add_subdirectory(src)
 
-if (GGML_SOURCES_CUDA)
-    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    # Only configure gmml CUDA architectures is not globally set
-    if (NOT DEFINED GGML_CUDA_ARCHITECTURES)
-        # Not overriden by user, so set defaults
-        set(GGML_CUDA_ARCHITECTURES 52 61 70)
-    endif()
-    message(STATUS "GGML Configuring CUDA architectures ${GGML_CUDA_ARCHITECTURES}")
-    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES ${GGML_CUDA_ARCHITECTURES})
-    set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-endif()
-
-if (EMSCRIPTEN)
-    set_target_properties(${TARGET} PROPERTIES COMPILE_FLAGS "-msimd128")
-endif()
-
-target_compile_definitions(${TARGET} PUBLIC
-    ${WHISPER_EXTRA_FLAGS}
-    )
-
-set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+#
+# install
+#
 
 include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
 
-install(TARGETS ${TARGET}
-    LIBRARY  DESTINATION lib
-    ARCHIVE  DESTINATION lib/static
-    RUNTIME  DESTINATION bin
-    RESOURCE DESTINATION bin
-    PUBLIC_HEADER DESTINATION include
-    )
+set(WHISPER_BUILD_NUMBER        ${BUILD_NUMBER})
+set(WHISPER_BUILD_COMMIT        ${BUILD_COMMIT})
+set(WHISPER_INSTALL_VERSION     ${CMAKE_PROJECT_VERSION})
 
-#
-# bindings
-#
+set(WHISPER_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(WHISPER_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(WHISPER_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
-add_subdirectory(bindings)
+get_directory_property(WHISPER_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+
+set_target_properties(whisper PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/whisper.h)
+install(TARGETS whisper LIBRARY PUBLIC_HEADER)
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/whisper-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper
+    PATH_VARS
+    WHISPER_INCLUDE_INSTALL_DIR
+    WHISPER_LIB_INSTALL_DIR
+    WHISPER_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+    ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+    VERSION ${WHISPER_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper)
+
+configure_file(cmake/whisper.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        DESTINATION lib/pkgconfig)
 
 #
 # programs, examples and tests
 #
 
 if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    enable_testing()
-    add_subdirectory(tests)
+    #include(CTest)
+    #add_subdirectory(tests)
 endif ()
 
 if (WHISPER_BUILD_EXAMPLES)

Package.swift

@@ -18,28 +18,29 @@ let package = Package(
             name: "whisper",
             path: ".",
             exclude: [
+               "build",
                "bindings",
                "cmake",
-               "coreml",
                "examples",
-               "extra",
+               "scripts",
                "models",
                "samples",
                "tests",
                "CMakeLists.txt",
-               "ggml-cuda.cu",
-               "ggml-cuda.h",
-               "Makefile"
+               "Makefile",
+               "ggml/src/ggml-metal-embed.metal"
             ],
             sources: [
-                "ggml.c",
-                "whisper.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m"
+                "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-cpu.c",
+                "ggml/src/ggml-quants.c",
+                "ggml/src/ggml-metal.m"
             ],
-            resources: [.process("ggml-metal.metal")],
+            resources: [.process("ggml/src/ggml-metal.metal")],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),

README.md

@@ -4,24 +4,26 @@
 
 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![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: [v1.5.5](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.5) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.7.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.2) / [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:
 
 - Plain C/C++ implementation without dependencies
-- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
-- [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
+- [Integer quantization support](#quantization)
 - Zero memory allocations at runtime
+- [Vulkan support](#vulkan-gpu-support)
 - Support for CPU-only inference
-- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
-- [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
-- [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
-- [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
+- [Efficient GPU support for NVIDIA](#nvidia-gpu-support)
+- [OpenVINO Support](#openvino-support)
+- [Ascend NPU Support](#ascend-npu-support)
+- [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/include/whisper.h)
 
 Supported platforms:
 
@@ -33,9 +35,9 @@ Supported platforms:
 - [x] [WebAssembly](examples/whisper.wasm)
 - [x] Windows ([MSVC](https://github.com/ggerganov/whisper.cpp/blob/master/.github/workflows/build.yml#L117-L144) and [MinGW](https://github.com/ggerganov/whisper.cpp/issues/168)]
 - [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/discussions/166)
-- [x] [docker](https://github.com/ggerganov/whisper.cpp/pkgs/container/whisper.cpp)
+- [x] [Docker](https://github.com/ggerganov/whisper.cpp/pkgs/container/whisper.cpp)
 
-The entire high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
+The entire high-level implementation of the model is contained in [whisper.h](include/whisper.h) and [whisper.cpp](src/whisper.cpp).
 The rest of the code is part of the [`ggml`](https://github.com/ggerganov/ggml) machine learning library.
 
 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
@@ -55,8 +57,8 @@ Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 
 ## Implementation details
 
-- The core tensor operations are implemented in C ([ggml.h](ggml.h) / [ggml.c](ggml.c))
-- The transformer model and the high-level C-style API are implemented in C++ ([whisper.h](whisper.h) / [whisper.cpp](whisper.cpp))
+- The core tensor operations are implemented in C ([ggml.h](ggml/include/ggml.h) / [ggml.c](ggml/src/ggml.c))
+- The transformer model and the high-level C-style API are implemented in C++ ([whisper.h](include/whisper.h) / [whisper.cpp](src/whisper.cpp))
 - Sample usage is demonstrated in [main.cpp](examples/main)
 - Sample real-time audio transcription from the microphone is demonstrated in [stream.cpp](examples/stream)
 - Various other examples are available in the [examples](examples) folder
@@ -71,17 +73,23 @@ First clone the repository:
 git clone https://github.com/ggerganov/whisper.cpp.git
 ```
 
+Navigate into the directory:
+
+```
+cd whisper.cpp
+```
+
 Then, download one of the Whisper [models](models/README.md) converted in [`ggml` format](#ggml-format). For example:
 
 ```bash
-bash ./models/download-ggml-model.sh base.en
+sh ./models/download-ggml-model.sh base.en
 ```
 
 Now build the [main](examples/main) example and transcribe an audio file like this:
 
 ```bash
 # build the main example
-make
+make -j
 
 # transcribe an audio file
 ./main -f samples/jfk.wav
@@ -92,7 +100,7 @@ make
 For a quick demo, simply run `make base.en`:
 
 ```text
-$ make base.en
+$ make -j base.en
 
 cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
 c++ -I. -I./examples -O3 -std=c++11 -pthread -c whisper.cpp -o whisper.o
@@ -145,7 +153,7 @@ options:
   -ng,       --no-gpu            [false  ] disable GPU
 
 
-bash ./models/download-ggml-model.sh base.en
+sh ./models/download-ggml-model.sh base.en
 Downloading ggml model base.en ...
 ggml-base.en.bin               100%[========================>] 141.11M  6.34MB/s    in 24s
 Done! Model 'base.en' saved in 'models/ggml-base.en.bin'
@@ -216,7 +224,7 @@ ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
 If you want some extra audio samples to play with, simply run:
 
 ```
-make samples
+make -j samples
 ```
 
 This will download a few more audio files from Wikipedia and convert them to 16-bit WAV format via `ffmpeg`.
@@ -224,17 +232,18 @@ This will download a few more audio files from Wikipedia and convert them to 16-
 You can download and run the other models as follows:
 
 ```
-make tiny.en
-make tiny
-make base.en
-make base
-make small.en
-make small
-make medium.en
-make medium
-make large-v1
-make large-v2
-make large-v3
+make -j tiny.en
+make -j tiny
+make -j base.en
+make -j base
+make -j small.en
+make -j small
+make -j medium.en
+make -j medium
+make -j large-v1
+make -j large-v2
+make -j large-v3
+make -j large-v3-turbo
 ```
 
 ## Memory usage
@@ -256,7 +265,7 @@ Here are the steps for creating and using a quantized model:
 
 ```bash
 # quantize a model with Q5_0 method
-make quantize
+make -j quantize
 ./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
@@ -418,31 +427,19 @@ Now build `whisper.cpp` with CUDA support:
 
 ```
 make clean
-WHISPER_CUDA=1 make -j
+GGML_CUDA=1 make -j
 ```
 
-## OpenCL GPU support via CLBlast
-
-For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APUs or low end devices for up to ~2x speedup.
-
-First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
-
-Now build `whisper.cpp` with CLBlast support:
+## Vulkan GPU support
+Cross-vendor solution which allows you to accelerate workload on your GPU.
+First, make sure your graphics card driver provides support for Vulkan API.
 
+Now build `whisper.cpp` with Vulkan support:
 ```
-Makefile:
-cd whisper.cpp
 make clean
-WHISPER_CLBLAST=1 make -j
-
-CMake:
-cd whisper.cpp
-cmake -B build -DWHISPER_CLBLAST=ON
-cmake --build build -j --config Release
+make GGML_VULKAN=1 -j
 ```
 
-Run all the examples as usual.
-
 ## BLAS CPU support via OpenBLAS
 
 Encoder processing can be accelerated on the CPU via OpenBLAS.
@@ -452,7 +449,7 @@ Now build `whisper.cpp` with OpenBLAS support:
 
 ```
 make clean
-WHISPER_OPENBLAS=1 make -j
+GGML_OPENBLAS=1 make -j
 ```
 
 ## BLAS CPU support via Intel MKL
@@ -470,6 +467,39 @@ cmake -DWHISPER_MKL=ON ..
 WHISPER_MKL=1 make -j
 ```
 
+## Ascend NPU support
+
+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:
+
+**Verified devices**
+| Ascend NPU                    | Status  |
+|:-----------------------------:|:-------:|
+| Atlas 300T A2                 | Support |
+
+Then, make sure you have installed [`CANN toolkit`](https://www.hiascend.com/en/software/cann/community) . The lasted version of CANN is recommanded.
+
+Now build `whisper.cpp` with CANN support:
+
+```
+mkdir build
+cd build
+cmake .. -D GGML_CANN=on
+make -j
+```
+
+Run the inference examples as usual, for example:
+
+```
+./build/bin/main -f samples/jfk.wav -m models/ggml-base.en.bin -t 8
+```
+
+*Notes:*
+
+- If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
+- If you run successfully with your Ascend NPU device, please help update the table `Verified devices`.
+
 ## Docker
 
 ### Prerequisites
@@ -502,6 +532,16 @@ docker run -it --rm \
   whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
 ```
 
+## Installing with Conan
+
+You can install pre-built binaries for whisper.cpp or build it from source using [Conan](https://conan.io/). Use the following command:
+
+```
+conan install --requires="whisper-cpp/[*]" --build=missing
+```
+
+For detailed instructions on how to use Conan, please refer to the [Conan documentation](https://docs.conan.io/2/).
+
 ## Limitations
 
 - Inference only
@@ -596,7 +636,7 @@ 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
+make stream -j
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
@@ -710,7 +750,7 @@ The [main](examples/main) example provides support for output of karaoke-style m
 currently pronounced word is highlighted. Use the `-wts` argument and run the generated bash script.
 This requires to have `ffmpeg` installed.
 
-Here are a few *"typical"* examples:
+Here are a few _"typical"_ examples:
 
 ```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
@@ -763,7 +803,7 @@ took to execute it. The results are summarized in the following Github issue:
 
 [Benchmark results](https://github.com/ggerganov/whisper.cpp/issues/89)
 
-Additionally a script to run whisper.cpp with different models and audio files is provided [bench.py](bench.py).
+Additionally a script to run whisper.cpp with different models and audio files is provided [bench.py](scripts/bench.py).
 
 You can run it with the following command, by default it will run against any standard model in the models folder.
 
@@ -808,7 +848,9 @@ For more details, see the conversion script [models/convert-pt-to-ggml.py](model
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
 - [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
   - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [AIWintermuteAI/whispercpp](https://github.com/AIWintermuteAI/whispercpp) (Updated fork of aarnphm/whispercpp)
   - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
+  - [abdeladim-s/pywhispercpp](https://github.com/abdeladim-s/pywhispercpp) (Pybind11)
 - [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
 - [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 

bindings/go/Makefile

@@ -14,9 +14,14 @@ GGML_METAL_PATH_RESOURCES := $(abspath ../..)
 BUILD_DIR := build
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
-INCLUDE_PATH := $(abspath ../..)
+INCLUDE_PATH := $(abspath ../../include):$(abspath ../../ggml/include)
 LIBRARY_PATH := $(abspath ../..)
 
+ifeq ($(GGML_CUDA),1)
+	LIBRARY_PATH := $(LIBRARY_PATH):$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib/
+	BUILD_FLAGS := -ldflags "-extldflags '-lcudart -lcuda -lcublas'"
+endif
+
 ifeq ($(UNAME_S),Darwin)
 	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit
 endif

bindings/go/README.md

@@ -62,6 +62,12 @@ This will compile a static `libwhisper.a` in a `build` folder, download a model
 make examples
 ```
 
+To build using cuda support add `GGML_CUDA=1`:
+
+```bash
+GGML_CUDA=1 make examples
+```
+
 The examples are placed in the `build` directory. Once built, you can download all the models with the following command:
 
 ```bash

bindings/go/examples/go-model-download/main.go

@@ -24,7 +24,7 @@ const (
 
 var (
 	// The models which will be downloaded, if no model is specified as an argument
-	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large-v3"}
+	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large-v3", "large-v3-turbo"}
 )
 
 var (

bindings/go/examples/go-whisper/flags.go

@@ -68,10 +68,6 @@ func (flags *Flags) GetOut() string {
 	return strings.ToLower(flags.Lookup("out").Value.String())
 }
 
-func (flags *Flags) IsSpeedup() bool {
-	return flags.Lookup("speedup").Value.String() == "true"
-}
-
 func (flags *Flags) IsTokens() bool {
 	return flags.Lookup("tokens").Value.String() == "true"
 }
@@ -111,10 +107,6 @@ func (flags *Flags) SetParams(context whisper.Context) error {
 		fmt.Fprintf(flags.Output(), "Setting duration to %v\n", duration)
 		context.SetDuration(duration)
 	}
-	if flags.IsSpeedup() {
-		fmt.Fprintf(flags.Output(), "Setting speedup to true\n")
-		context.SetSpeedup(true)
-	}
 	if threads := flags.GetThreads(); threads != 0 {
 		fmt.Fprintf(flags.Output(), "Setting threads to %d\n", threads)
 		context.SetThreads(threads)
@@ -146,7 +138,6 @@ func registerFlags(flag *Flags) {
 	flag.Duration("offset", 0, "Time offset")
 	flag.Duration("duration", 0, "Duration of audio to process")
 	flag.Uint("threads", 0, "Number of threads to use")
-	flag.Bool("speedup", false, "Enable speedup")
 	flag.Uint("max-len", 0, "Maximum segment length in characters")
 	flag.Uint("max-tokens", 0, "Maximum tokens per segment")
 	flag.Float64("word-thold", 0, "Maximum segment score")

bindings/go/go.mod

@@ -1,10 +1,10 @@
 module github.com/ggerganov/whisper.cpp/bindings/go
 
-go 1.19
+go 1.23
 
 require (
 	github.com/go-audio/wav v1.1.0
-	github.com/stretchr/testify v1.8.1
+	github.com/stretchr/testify v1.9.0
 )
 
 require (

bindings/go/go.sum

@@ -1,4 +1,3 @@
-github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/go-audio/audio v1.0.0 h1:zS9vebldgbQqktK4H0lUqWrG8P0NxCJVqcj7ZpNnwd4=
@@ -9,15 +8,9 @@ github.com/go-audio/wav v1.1.0 h1:jQgLtbqBzY7G+BM8fXF7AHUk1uHUviWS4X39d5rsL2g=
 github.com/go-audio/wav v1.1.0/go.mod h1:mpe9qfwbScEbkd8uybLuIpTgHyrISw/OTuvjUW2iGtE=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
-github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
-github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
-github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
-github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
-github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
-github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
-github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
+github.com/stretchr/testify v1.9.0 h1:HtqpIVDClZ4nwg75+f6Lvsy/wHu+3BoSGCbBAcpTsTg=
+github.com/stretchr/testify v1.9.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
-gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

bindings/go/params.go

@@ -47,10 +47,6 @@ func (p *Params) SetPrintTimestamps(v bool) {
 	p.print_timestamps = toBool(v)
 }
 
-func (p *Params) SetSpeedup(v bool) {
-	p.speed_up = toBool(v)
-}
-
 // Set language id
 func (p *Params) SetLanguage(lang int) error {
 	if lang == -1 {
@@ -123,6 +119,28 @@ func (p *Params) SetAudioCtx(n int) {
 	p.audio_ctx = C.int(n)
 }
 
+func (p *Params) SetMaxContext(n int) {
+	p.n_max_text_ctx = C.int(n)
+}
+
+func (p *Params) SetBeamSize(n int) {
+	p.beam_search.beam_size = C.int(n)
+}
+
+func (p *Params) SetEntropyThold(t float32) {
+	p.entropy_thold = C.float(t)
+}
+
+func (p *Params) SetTemperature(t float32) {
+	p.temperature = C.float(t)
+}
+
+// Sets the fallback temperature incrementation
+// Pass -1.0 to disable this feature
+func (p *Params) SetTemperatureFallback(t float32) {
+	p.temperature_inc = C.float(t)
+}
+
 // Set initial prompt
 func (p *Params) SetInitialPrompt(prompt string) {
 	p.initial_prompt = C.CString(prompt)
@@ -153,6 +171,10 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
 	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
 	str += fmt.Sprintf(" initial_prompt=%s", C.GoString(p.initial_prompt))
+	str += fmt.Sprintf(" entropy_thold=%f", p.entropy_thold)
+	str += fmt.Sprintf(" temperature=%f", p.temperature)
+	str += fmt.Sprintf(" temperature_inc=%f", p.temperature_inc)
+	str += fmt.Sprintf(" beam_size=%d", p.beam_search.beam_size)
 	if p.translate {
 		str += " translate"
 	}
@@ -177,9 +199,6 @@ func (p *Params) String() string {
 	if p.token_timestamps {
 		str += " token_timestamps"
 	}
-	if p.speed_up {
-		str += " speed_up"
-	}
 
 	return str + ">"
 }

bindings/go/pkg/whisper/context.go

@@ -76,11 +76,6 @@ func (context *context) SetTranslate(v bool) {
 	context.params.SetTranslate(v)
 }
 
-// Set speedup flag
-func (context *context) SetSpeedup(v bool) {
-	context.params.SetSpeedup(v)
-}
-
 func (context *context) SetSplitOnWord(v bool) {
 	context.params.SetSplitOnWord(v)
 }
@@ -130,6 +125,32 @@ func (context *context) SetAudioCtx(n uint) {
 	context.params.SetAudioCtx(int(n))
 }
 
+// Set maximum number of text context tokens to store
+func (context *context) SetMaxContext(n int) {
+	context.params.SetMaxContext(n)
+}
+
+// Set Beam Size
+func (context *context) SetBeamSize(n int) {
+	context.params.SetBeamSize(n)
+}
+
+// Set Entropy threshold
+func (context *context) SetEntropyThold(t float32) {
+	context.params.SetEntropyThold(t)
+}
+
+// Set Temperature
+func (context *context) SetTemperature(t float32) {
+	context.params.SetTemperature(t)
+}
+
+// Set the fallback temperature incrementation
+// Pass -1.0 to disable this feature
+func (context *context) SetTemperatureFallback(t float32) {
+	context.params.SetTemperatureFallback(t)
+}
+
 // Set initial prompt
 func (context *context) SetInitialPrompt(prompt string) {
 	context.params.SetInitialPrompt(prompt)

bindings/go/pkg/whisper/context_test.go

@@ -4,52 +4,90 @@ import (
 	"os"
 	"testing"
 
-	// Packages
-	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	"github.com/go-audio/wav"
 	assert "github.com/stretchr/testify/assert"
 )
 
-const (
-	ModelPath  = "../../models/ggml-tiny.bin"
-	SamplePath = "../../samples/jfk.wav"
-)
-
-func Test_Whisper_000(t *testing.T) {
+func TestSetLanguage(t *testing.T) {
 	assert := assert.New(t)
-	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
-		t.Skip("Skipping test, model not found:", ModelPath)
-	}
-	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
-		t.Skip("Skipping test, sample not found:", SamplePath)
-	}
 
-	// Load model
-	model, err := whisper.New(ModelPath)
-	assert.NoError(err)
-	assert.NotNil(model)
-	assert.NoError(model.Close())
-
-	t.Log("languages=", model.Languages())
-}
-
-func Test_Whisper_001(t *testing.T) {
-	assert := assert.New(t)
-	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
-		t.Skip("Skipping test, model not found:", ModelPath)
-	}
-	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
-		t.Skip("Skipping test, sample not found:", SamplePath)
-	}
-
-	// Load model
 	model, err := whisper.New(ModelPath)
 	assert.NoError(err)
 	assert.NotNil(model)
 	defer model.Close()
 
-	// Get context for decoding
-	ctx, err := model.NewContext()
+	context, err := model.NewContext()
 	assert.NoError(err)
-	assert.NotNil(ctx)
 
+	// This returns an error since
+	// the model 'models/ggml-small.en.bin'
+	// that is loaded is not multilingual
+	err = context.SetLanguage("en")
+	assert.Error(err)
+}
+
+func TestContextModelIsMultilingual(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	context, err := model.NewContext()
+	assert.NoError(err)
+
+	isMultilingual := context.IsMultilingual()
+
+	// This returns false since
+	// the model 'models/ggml-small.en.bin'
+	// that is loaded is not multilingual
+	assert.False(isMultilingual)
+}
+
+func TestLanguage(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	context, err := model.NewContext()
+	assert.NoError(err)
+
+	// This always returns en since
+	// the model 'models/ggml-small.en.bin'
+	// that is loaded is not multilingual
+	expectedLanguage := "en"
+	actualLanguage := context.Language()
+	assert.Equal(expectedLanguage, actualLanguage)
+}
+
+func TestProcess(t *testing.T) {
+	assert := assert.New(t)
+
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Decode the WAV file - load the full buffer
+	dec := wav.NewDecoder(fh)
+	buf, err := dec.FullPCMBuffer()
+	assert.NoError(err)
+	assert.Equal(uint16(1), dec.NumChans)
+
+	data := buf.AsFloat32Buffer().Data
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	context, err := model.NewContext()
+	assert.NoError(err)
+
+	err = context.Process(data, nil, nil)
+	assert.NoError(err)
 }

bindings/go/pkg/whisper/interface.go

@@ -38,18 +38,22 @@ type Context interface {
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language
 
-	SetOffset(time.Duration)        // Set offset
-	SetDuration(time.Duration)      // Set duration
-	SetThreads(uint)                // Set number of threads to use
-	SetSpeedup(bool)                // Set speedup flag
-	SetSplitOnWord(bool)            // Set split on word flag
-	SetTokenThreshold(float32)      // Set timestamp token probability threshold
-	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
-	SetMaxSegmentLength(uint)       // Set max segment length in characters
-	SetTokenTimestamps(bool)        // Set token timestamps flag
-	SetMaxTokensPerSegment(uint)    // Set max tokens per segment (0 = no limit)
-	SetAudioCtx(uint)               // Set audio encoder context
-	SetInitialPrompt(prompt string) // Set initial prompt
+	SetOffset(time.Duration)          // Set offset
+	SetDuration(time.Duration)        // Set duration
+	SetThreads(uint)                  // Set number of threads to use
+	SetSplitOnWord(bool)              // Set split on word flag
+	SetTokenThreshold(float32)        // Set timestamp token probability threshold
+	SetTokenSumThreshold(float32)     // Set timestamp token sum probability threshold
+	SetMaxSegmentLength(uint)         // Set max segment length in characters
+	SetTokenTimestamps(bool)          // Set token timestamps flag
+	SetMaxTokensPerSegment(uint)      // Set max tokens per segment (0 = no limit)
+	SetAudioCtx(uint)                 // Set audio encoder context
+	SetMaxContext(n int)              // Set maximum number of text context tokens to store
+	SetBeamSize(n int)                // Set Beam Size
+	SetEntropyThold(t float32)        // Set Entropy threshold
+	SetInitialPrompt(prompt string)   // Set initial prompt
+	SetTemperature(t float32)         // Set temperature
+	SetTemperatureFallback(t float32) // Set temperature incrementation
 
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the

bindings/go/pkg/whisper/model_test.go

@@ -0,0 +1,91 @@
+package whisper_test
+
+import (
+	"testing"
+
+	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	assert "github.com/stretchr/testify/assert"
+)
+
+func TestNew(t *testing.T) {
+	assert := assert.New(t)
+	t.Run("valid model path", func(t *testing.T) {
+		model, err := whisper.New(ModelPath)
+		assert.NoError(err)
+		assert.NotNil(model)
+		defer model.Close()
+
+	})
+
+	t.Run("invalid model path", func(t *testing.T) {
+		invalidModelPath := "invalid-model-path.bin"
+		model, err := whisper.New(invalidModelPath)
+		assert.Error(err)
+		assert.Nil(model)
+	})
+}
+
+func TestClose(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+
+	err = model.Close()
+	assert.NoError(err)
+}
+
+func TestNewContext(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	context, err := model.NewContext()
+	assert.NoError(err)
+	assert.NotNil(context)
+}
+
+func TestIsMultilingual(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	isMultilingual := model.IsMultilingual()
+
+	// This returns false since
+	// the model 'models/ggml-small.en.bin'
+	// that is loaded is not multilingual
+	assert.False(isMultilingual)
+}
+
+func TestLanguages(t *testing.T) {
+	assert := assert.New(t)
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	expectedLanguages := []string{
+		"en", "zh", "de", "es", "ru", "ko", "fr", "ja", "pt", "tr", "pl",
+		"ca", "nl", "ar", "sv", "it", "id", "hi", "fi", "vi", "he", "uk",
+		"el", "ms", "cs", "ro", "da", "hu", "ta", "no", "th", "ur", "hr",
+		"bg", "lt", "la", "mi", "ml", "cy", "sk", "te", "fa", "lv", "bn",
+		"sr", "az", "sl", "kn", "et", "mk", "br", "eu", "is", "hy", "ne",
+		"mn", "bs", "kk", "sq", "sw", "gl", "mr", "pa", "si", "km", "sn",
+		"yo", "so", "af", "oc", "ka", "be", "tg", "sd", "gu", "am", "yi",
+		"lo", "uz", "fo", "ht", "ps", "tk", "nn", "mt", "sa", "lb", "my",
+		"bo", "tl", "mg", "as", "tt", "haw", "ln", "ha", "ba", "jw", "su",
+	}
+
+	actualLanguages := model.Languages()
+
+	assert.Equal(expectedLanguages, actualLanguages)
+}

bindings/go/pkg/whisper/util_test.go

@@ -0,0 +1,6 @@
+package whisper_test
+
+const (
+	ModelPath  = "../../models/ggml-small.en.bin"
+	SamplePath = "../../samples/jfk.wav"
+)

bindings/go/whisper.go

@@ -9,7 +9,7 @@ import (
 // CGO
 
 /*
-#cgo LDFLAGS: -lwhisper -lm -lstdc++
+#cgo LDFLAGS: -lwhisper -lm -lstdc++ -fopenmp
 #cgo darwin LDFLAGS: -framework Accelerate -framework Metal -framework Foundation -framework CoreGraphics
 #include <whisper.h>
 #include <stdlib.h>

bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java

@@ -20,7 +20,7 @@ public interface WhisperCppJnaLibrary extends Library {
      * @return Whisper context on success, null on failure
      */
     Pointer whisper_init_from_file(String path_model);
-    
+
     /**
      * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
      * Because this function allocates memory for the params, the caller must call either:
@@ -304,14 +304,6 @@ public interface WhisperCppJnaLibrary extends Library {
     /** Language id associated with the provided state */
     int whisper_full_lang_id_from_state(Pointer state);
 
-    /**
-     * Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
-     * The resulting spectrogram is stored inside the default state of the provided whisper context.
-     * @return 0 on success
-     */
-    int whisper_pcm_to_mel_phase_vocoder(Pointer ctx, final float[] samples, int n_samples, int n_threads);
-
-    int whisper_pcm_to_mel_phase_vocoder_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);
 
     /** Get the start time of the specified segment. */
     long whisper_full_get_segment_t0(Pointer ctx, int i_segment);

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java

@@ -129,14 +129,6 @@ public class WhisperFullParams extends Structure {
     /** Maximum tokens per segment (0, default = no limit) */
     public int max_tokens;
 
-    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
-    public CBool speed_up;
-
-    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
-    public void speedUp(boolean enable) {
-        speed_up = enable ? CBool.TRUE : CBool.FALSE;
-    }
-
     /** Overwrite the audio context size (0 = use default). */
     public int audio_ctx;
 
@@ -321,7 +313,7 @@ public class WhisperFullParams extends Structure {
         return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
                 "no_context", "single_segment", "no_timestamps",
                 "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
-                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "audio_ctx",
                 "tdrz_enable", "suppress_regex", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
                 "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
                 "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.5.5",
+  "version": "1.7.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/.gitignore

@@ -0,0 +1,3 @@
+LICENSE
+pkg/
+lib/whisper.*

bindings/ruby/README.md

@@ -0,0 +1,169 @@
+whispercpp
+==========
+
+![whisper.cpp](https://user-images.githubusercontent.com/1991296/235238348-05d0f6a4-da44-4900-a1de-d0707e75b763.jpeg)
+
+Ruby bindings for [whisper.cpp][], an interface of automatic speech recognition model.
+
+Installation
+------------
+
+Install the gem and add to the application's Gemfile by executing:
+
+    $ bundle add whispercpp
+
+If bundler is not being used to manage dependencies, install the gem by executing:
+
+    $ gem install whispercpp
+
+Usage
+-----
+
+```ruby
+require "whisper"
+
+whisper = Whisper::Context.new("path/to/model.bin")
+
+params = Whisper::Params.new
+params.language = "en"
+params.offset = 10_000
+params.duration = 60_000
+params.max_text_tokens = 300
+params.translate = true
+params.print_timestamps = false
+params.initial_prompt = "Initial prompt here."
+
+whisper.transcribe("path/to/audio.wav", params) do |whole_text|
+  puts whole_text
+end
+
+```
+
+### Preparing model ###
+
+Use script to download model file(s):
+
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+cd whisper.cpp
+sh ./models/download-ggml-model.sh base.en
+```
+
+There are some types of models. See [models][] page for details.
+
+### Preparing audio file ###
+
+Currently, whisper.cpp accepts only 16-bit WAV files.
+
+### API ###
+
+Once `Whisper::Context#transcribe` called, you can retrieve segments by `#each_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
+
+whisper.transcribe("path/to/audio.wav", params)
+
+whisper.each_segment.with_index do |segment, index|
+  line = "[%{nth}: %{st} --> %{ed}] %{text}" % {
+    nth: index + 1,
+    st: format_time(segment.start_time),
+    ed: format_time(segment.end_time),
+    text: segment.text
+  }
+  line << " (speaker turned)" if segment.speaker_next_turn?
+  puts line
+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}" % {
+    st: format_time(segment.start_time),
+    ed: format_time(segment.end_time),
+    text: segment.text
+  }
+  line << " (speaker turned)" if segment.speaker_next_turn?
+  puts line
+end
+
+whisper.transcribe("path/to/audio.wav", params)
+
+```
+
+You can see model information:
+
+```ruby
+whisper = Whisper::Context.new("path/to/model.bin")
+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"
+
+```
+
+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(MODEL)
+```
+
+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

@@ -0,0 +1,68 @@
+require 'rake/clean'
+require "bundler/gem_tasks"
+require "pathname"
+require "yaml"
+require "rake/testtask"
+
+extsources = YAML.load_file("extsources.yaml")
+SOURCES = FileList[]
+extsources.each do |src|
+  basename = src.pathmap("%f")
+  dest = basename == "LICENSE" ? basename : basename.pathmap("ext/%f")
+  file src
+  file dest => src 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: FileList[
+       "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)
+
+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 "make"
+  end
+end
+CLEAN.include LIB_FILE
+
+directory "lib"
+file LIB_FILE => [SO_FILE, "lib"] do |t|
+  copy t.source, t.name
+end
+
+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"
+  end
+end

bindings/ruby/ext/.gitignore

@@ -3,7 +3,33 @@ ggml.c
 ggml.h
 ggml-alloc.c
 ggml-alloc.h
-whisper.bundle
+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.dll

bindings/ruby/ext/extconf.rb

@@ -1,22 +1,10 @@
 require 'mkmf'
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-common.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
-
 
 # need to use c++ compiler flags
 $CXXFLAGS << ' -std=c++11'
+
+$LDFLAGS << ' -lstdc++'
+
 # Set to true when building binary gems
 if enable_config('static-stdlib', false)
   $LDFLAGS << ' -static-libgcc -static-libstdc++'
@@ -27,4 +15,180 @@ if enable_config('march-tune-native', false)
   $CXXFLAGS << ' -march=native -mtune=native'
 end
 
+if ENV['WHISPER_METAL']
+  $GGML_METAL ||= true
+  $DEPRECATE_WARNING ||= true
+end
+
+$UNAME_S = `uname -s`.chomp
+$UNAME_P = `uname -p`.chomp
+$UNAME_M = `uname -m`.chomp
+
+if $UNAME_S == 'Darwin'
+  unless ENV['GGML_NO_METAL']
+    $GGML_METAL ||= true
+  end
+  $GGML_NO_OPENMP ||= true
+end
+
+if $GGML_METAL
+  $GGML_METAL_EMBED_LIBRARY = true
+end
+
+$MK_CPPFLAGS = ''
+$MK_CFLAGS   = '-std=c11   -fPIC'
+$MK_CXXFLAGS = '-std=c++11 -fPIC'
+$MK_NVCCFLAGS = '-std=c++11'
+$MK_LDFLAGS = ''
+
+$OBJ_GGML = []
+$OBJ_WHISPER = []
+$OBJ_COMMON = []
+$OBJ_SDL = []
+
+$MK_CPPFLAGS << ' -D_XOPEN_SOURCE=600'
+
+if $UNAME_S == 'Linux'
+  $MK_CPPFLAGS << ' -D_GNU_SOURCE'
+end
+
+if $UNAME_S == 'Darwin'
+  $MK_CPPFLAGS << ' -D_DARWIN_C_SOURCE'
+end
+
+if ENV['WHISPER_DEBUG']
+  $MK_CFLAGS    << ' -O0 -g'
+  $MK_CXXFLAGS  << ' -O0 -g'
+  $MK_LDFLAGS   << ' -g'
+  $MK_NVCCFLAGS << ' -O0 -g'
+else
+  $MK_CPPFLAGS   << ' -DNDEBUG'
+  $MK_CFLAGS     << ' -O3'
+  $MK_CXXFLAGS   << ' -O3'
+  $MK_NVCCFLAGS  << ' -O3'
+end
+
+$WARN_FLAGS =
+  ' -Wall' <<
+  ' -Wextra' <<
+  ' -Wpedantic' <<
+  ' -Wcast-qual' <<
+  ' -Wno-unused-function'
+
+$MK_CFLAGS <<
+  $WARN_FLAGS <<
+  ' -Wshadow' <<
+  ' -Wstrict-prototypes' <<
+  ' -Wpointer-arith' <<
+  ' -Wmissing-prototypes' <<
+  ' -Werror=implicit-int' <<
+  ' -Werror=implicit-function-declaration'
+
+$MK_CXXFLAGS <<
+  $WARN_FLAGS <<
+  ' -Wmissing-declarations' <<
+  ' -Wmissing-noreturn'
+
+unless `#{cc_command} #{$LDFLAGS} -Wl,-v 2>&1`.chomp.include? 'dyld-1015.7'
+  $MK_CPPFLAGS << ' -DHAVE_BUGGY_APPLE_LINKER'
+end
+
+if %w[Linux Darwin FreeBSD NetBSD OpenBSD Haiku].include? $UNAME_S
+  $MK_CFLAGS   << ' -pthread'
+  $MK_CXXFLAGS << ' -pthread'
+end
+
+unless $_WIN32
+  $DSO_EXT = '.so'
+else
+  $DSO_EXT = '.dll'
+end
+
+unless ENV['RISCV']
+  if %w[x86_64 i686 amd64].include? $UNAME_M
+    $HOST_CXXFLAGS ||= ''
+
+    $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'
+end
+
+unless ENV['GGML_NO_ACCELERATE']
+  if $UNAME_S == 'Darwin'
+    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS'
+    $MK_CPPFLAGS << ' -DACCELERATE_NEW_LAPACK'
+    $MK_CPPFLAGS << ' -DACCELERATE_LAPACK_ILP64'
+    $MK_LDFLAGS  << ' -framework Accelerate'
+    $OBJ_GGML    << 'ggml-blas.o'
+  end
+end
+
+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'
+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'
+end
+
+if $GGML_METAL
+  $MK_CPPFLAGS << ' -DGGML_USE_METAL'
+  $MK_LDFLAGS  << ' -framework Foundation -framework Metal -framework MetalKit'
+  $OBJ_GGML    << 'ggml-metal.o'
+
+  if ENV['GGML_METAL_NDEBUG']
+    $MK_CPPFLAGS << ' -DGGML_METAL_NDEBUG'
+  end
+
+  if $GGML_METAL_EMBED_LIBRARY
+    $MK_CPPFLAGS << ' -DGGML_METAL_EMBED_LIBRARY'
+    $OBJ_GGML    << 'ggml-metal-embed.o'
+  end
+end
+
+$OBJ_GGML <<
+  'ggml.o' <<
+  'ggml-cpu.o' <<
+  'ggml-alloc.o' <<
+  'ggml-backend.o' <<
+  'ggml-quants.o' <<
+  'ggml-aarch64.o'
+
+$OBJ_WHISPER <<
+  'whisper.o'
+
+$objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
+$objs << "ruby_whisper.o"
+
+$CPPFLAGS  = "#{$MK_CPPFLAGS} #{$CPPFLAGS}"
+$CFLAGS    = "#{$CPPFLAGS} #{$MK_CFLAGS} #{$GF_CFLAGS} #{$CFLAGS}"
+$BASE_CXXFLAGS = "#{$MK_CXXFLAGS} #{$CXXFLAGS}"
+$CXXFLAGS  = "#{$BASE_CXXFLAGS} #{$HOST_CXXFLAGS} #{$GF_CXXFLAGS} #{$CPPFLAGS}"
+$NVCCFLAGS = "#{$MK_NVCCFLAGS} #{$NVCCFLAGS}"
+$LDFLAGS   = "#{$MK_LDFLAGS} #{$LDFLAGS}"
+
 create_makefile('whisper')
+
+File.open 'Makefile', 'a' do |file|
+  file.puts 'include get-flags.mk'
+
+  if $GGML_METAL
+    if $GGML_METAL_EMBED_LIBRARY
+      file.puts 'include metal-embed.mk'
+    end
+  end
+end

bindings/ruby/ext/ggml-backend-impl.h

@@ -1,87 +0,0 @@
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
-    };
-
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
-        ggml_backend_buffer_context_t context;
-
-        size_t size;
-    };
-
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*get_name)(ggml_backend_t backend);
-
-        void (*free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
-
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
-        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);
-        void (*synchronize)     (ggml_backend_t backend);
-
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph without a plan
-        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend supports an operation
-        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-    };
-
-    struct ggml_backend {
-        struct ggml_backend_i iface;
-
-        ggml_backend_context_t context;
-    };
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-backend.c

@@ -1,950 +0,0 @@
-#include "ggml-backend-impl.h"
-#include "ggml-alloc.h"
-#include "ggml-impl.h"
-
-#include <assert.h>
-#include <limits.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define UNUSED GGML_UNUSED
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-// backend buffer
-
-ggml_backend_buffer_t ggml_backend_buffer_init(
-        struct ggml_backend                  * backend,
-        struct ggml_backend_buffer_i           iface,
-               ggml_backend_buffer_context_t   context,
-               size_t                          size) {
-    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
-
-    GGML_ASSERT(iface.get_base != NULL);
-
-    (*buffer) = (struct ggml_backend_buffer) {
-        /* .interface = */ iface,
-        /* .backend   = */ backend,
-        /* .context   = */ context,
-        /* .size      = */ size,
-    };
-
-    return buffer;
-}
-
-void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
-    if (buffer == NULL) {
-        return;
-    }
-
-    if (buffer->iface.free_buffer != NULL) {
-        buffer->iface.free_buffer(buffer);
-    }
-    free(buffer);
-}
-
-size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
-    return ggml_backend_get_alignment(buffer->backend);
-}
-
-size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-    return buffer->size;
-}
-
-void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
-    void * base = buffer->iface.get_base(buffer);
-
-    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
-
-    return base;
-}
-
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // get_alloc_size is optional, defaults to ggml_nbytes
-    if (buffer->iface.get_alloc_size) {
-        return buffer->iface.get_alloc_size(buffer, tensor);
-    }
-    return ggml_nbytes(tensor);
-}
-
-void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // init_tensor is optional
-    if (buffer->iface.init_tensor) {
-        buffer->iface.init_tensor(buffer, tensor);
-    }
-}
-
-void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // free_tensor is optional
-    if (buffer->iface.free_tensor) {
-        buffer->iface.free_tensor(buffer, tensor);
-    }
-}
-
-// backend
-
-ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
-    return tensor->buffer ? tensor->buffer->backend : NULL;
-}
-
-const char * ggml_backend_name(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return "NULL";
-    }
-    return backend->iface.get_name(backend);
-}
-
-void ggml_backend_free(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return;
-    }
-
-    backend->iface.free(backend);
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
-    return backend->iface.alloc_buffer(backend, size);
-}
-
-size_t ggml_backend_get_alignment(ggml_backend_t backend) {
-    return backend->iface.get_alignment(backend);
-}
-
-void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
-
-    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
-}
-
-void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
-
-    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
-}
-
-void ggml_backend_synchronize(ggml_backend_t backend) {
-    backend->iface.synchronize(backend);
-}
-
-ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_plan_create(backend, cgraph);
-}
-
-void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_free(backend, plan);
-}
-
-void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_compute(backend, plan);
-}
-
-bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_compute(backend, cgraph);
-}
-
-bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return backend->iface.supports_op(backend, op);
-}
-
-// backend copy
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
-    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
-    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
-    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
-
-    // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
-
-    if (src == dst) {
-        return;
-    }
-
-    // TODO: allow backends to support copy to/from same backend
-
-    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
-        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
-    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
-        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
-    } else {
-        // shouldn't be hit when copying from/to CPU
-        #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
-        #endif
-        size_t nbytes = ggml_nbytes(src);
-        void * data = malloc(nbytes);
-        ggml_backend_tensor_get(src, data, 0, nbytes);
-        ggml_backend_tensor_set(dst, data, 0, nbytes);
-        free(data);
-    }
-}
-
-// backend CPU
-
-struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
-};
-
-static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
-    return "CPU";
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    free(cpu_ctx->work_data);
-    free(cpu_ctx);
-    free(backend);
-}
-
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
-
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-    UNUSED(buffer);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL, // no initialization required
-    /* .free_tensor    = */ NULL, // no cleanup required
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL,
-    /* .free_tensor    = */ NULL,
-};
-
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-
-static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
-
-    GGML_ASSERT(data != NULL && "failed to allocate buffer");
-
-    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
-    return TENSOR_ALIGNMENT;
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph;
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-    }
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    free(cpu_plan->cplan.work_data);
-    free(cpu_plan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
-        cpu_ctx->work_size = cplan.work_size;
-    }
-
-    cplan.work_data = cpu_ctx->work_data;
-
-    ggml_graph_compute(cgraph, &cplan);
-}
-
-static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return true;
-    UNUSED(backend);
-    UNUSED(op);
-}
-
-static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name            = */ ggml_backend_cpu_name,
-    /* .free                = */ ggml_backend_cpu_free,
-    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_cpu_synchronize,
-    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
-    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
-    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
-    /* .supports_op         = */ ggml_backend_cpu_supports_op,
-};
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
-
-    ctx->n_threads = GGML_DEFAULT_N_THREADS;
-    ctx->work_data = NULL;
-    ctx->work_size = 0;
-
-    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
-
-    *cpu_backend = (struct ggml_backend) {
-        /* .interface = */ cpu_backend_i,
-        /* .context   = */ ctx
-    };
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_cpu_name;
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
-    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
-}
-
-// scheduler
-
-#define GGML_MAX_BACKENDS 4
-#define GGML_MAX_SPLITS 256
-#define GGML_MAX_SPLIT_INPUTS 16
-
-struct ggml_backend_sched_split {
-    ggml_tallocr_t tallocr;
-    int i_start;
-    int i_end;
-    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
-    int n_inputs;
-    struct ggml_cgraph * graph;
-};
-
-struct ggml_backend_sched {
-    int n_backends;
-    ggml_backend_t backends[GGML_MAX_BACKENDS];
-    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
-
-    ggml_gallocr_t galloc;
-
-    struct ggml_hash_set    hash_set;
-    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
-    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
-
-    struct ggml_cgraph * graph;
-    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
-    int n_splits;
-
-    struct ggml_context * ctx;
-
-    // align context_buffer to GGML_MEM_ALIGN
-    #ifdef _MSC_VER
-    __declspec(align(GGML_MEM_ALIGN))
-    #else
-    __attribute__((aligned(GGML_MEM_ALIGN)))
-    #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
-};
-
-#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
-#define node_allocr(node) sched->node_talloc[hash_id(node)]
-
-static bool ggml_is_view_op(enum ggml_op op) {
-    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
-}
-
-// returns the priority of the backend, lower is better
-static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->backends[i] == backend) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->tallocs[i] == allocr) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-// returns the backend that should be used for the node based on the current locations
-char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
-static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
-    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
-    // ie. kv cache updates
-    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
-    // dst
-    ggml_backend_t cur_backend = ggml_get_backend(node);
-    if (cur_backend != NULL) {
-        sprintf(causes[hash_id(node)], "1.dst");
-        return cur_backend;
-    }
-
-    // view_src
-    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
-        sprintf(causes[hash_id(node)], "1.vsrc");
-        return ggml_get_backend(node->view_src);
-    }
-
-    // src
-    int cur_prio = INT_MAX;
-    size_t cur_size = 0;
-
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        const struct ggml_tensor * src = node->src[i];
-        if (src == NULL) {
-            break;
-        }
-        ggml_backend_t src_backend = ggml_get_backend(src);
-        if (src_backend != NULL) {
-            int src_prio = sched_backend_prio(sched, src_backend);
-            size_t src_size = ggml_nbytes(src);
-            if (src_prio < cur_prio && src_size >= cur_size) {
-                cur_prio = src_prio;
-                cur_size = src_size;
-                cur_backend = src_backend;
-                sprintf(causes[hash_id(node)], "1.src%d", i);
-            }
-        }
-    }
-    return cur_backend;
-}
-
-static char * fmt_size(size_t size) {
-    static char buffer[128];
-    if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
-    } else {
-        sprintf(buffer, "%zuK", size/1024);
-    }
-    return buffer;
-}
-
-static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
-            ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
-            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs);
-            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
-                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
-            }
-            fprintf(stderr, "\n");
-            cur_split++;
-        }
-        struct ggml_tensor * node = graph->nodes[i];
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
-        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
-            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
-        }
-        fprintf(stderr, "\n");
-    }
-}
-
-// creates a copy of the tensor with the same memory layout
-static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
-    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        dup->nb[i] = tensor->nb[i];
-    }
-    return dup;
-}
-
-// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
-// TODO: merge passes
-static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    // reset state
-    size_t hash_size = sched->hash_set.size;
-    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
-    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
-    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
-    sched->n_splits = 0;
-
-    struct ggml_init_params params = {
-        /*.mem_size =   */ sizeof(sched->context_buffer),
-        /*.mem_buffer = */ sched->context_buffer,
-        /*.no_alloc =   */ true
-    };
-
-    if (sched->ctx != NULL) {
-        ggml_free(sched->ctx);
-    }
-
-    sched->ctx = ggml_init(params);
-
-    // pass 1: assign backends to ops with allocated inputs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        if (node_allocr(leaf) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
-        if (leaf_backend == NULL && leaf->view_src != NULL) {
-            leaf_backend = ggml_get_backend(leaf->view_src);
-        }
-        if (leaf_backend != NULL) {
-            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
-        }
-    }
-
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node_allocr(node) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
-        if (node_backend != NULL) {
-            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
-        }
-    }
-    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 2: assign backends to ops from current assignments
-    // TODO:
-    //  - reuse sched_backend_from_cur
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            int    cur_prio = INT_MAX;
-            size_t cur_size = 0;
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * src = node->src[j];
-                if (src == NULL) {
-                    break;
-                }
-                ggml_tallocr_t src_allocr = node_allocr(src);
-                if (src_allocr != NULL) {
-                    int    src_prio = sched_allocr_prio(sched, src_allocr);
-                    size_t src_size = ggml_nbytes(src);
-                    if (src_prio < cur_prio && src_size >= cur_size) {
-                        cur_prio = src_prio;
-                        cur_size = src_size;
-                        node_allocr = src_allocr;
-                        sprintf(causes[hash_id(node)], "2.src%d", j);
-                    }
-                }
-            }
-            if (node_allocr != NULL) {
-                node_allocr(node) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 3: assign backends to remaining src from dst (should only be leafs)
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr == NULL) {
-                node_allocr(src) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 4: split graph, find tensors that need to be copied
-    // TODO:
-    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
-    // find first backend
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node->view_src == NULL) {
-            sched->splits[0].tallocr = node_allocr(node);
-            break;
-        }
-    }
-    sched->splits[0].i_start = 0;
-    sched->splits[0].n_inputs = 0;
-    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
-    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
-    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-
-        ggml_tallocr_t node_allocr = node_allocr(node);
-
-        if (node_allocr != cur_allocr) {
-            sched->splits[cur_split].i_end = i;
-            cur_split++;
-            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
-            sched->splits[cur_split].tallocr = node_allocr;
-            sched->splits[cur_split].i_start = i;
-            sched->splits[cur_split].n_inputs = 0;
-            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
-            cur_allocr = node_allocr;
-            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-        }
-
-        // find inputs that are not on the same backend
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr) {
-                int n_inputs = sched->splits[cur_split].n_inputs++;
-                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
-                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
-
-                // create copies
-                size_t id = hash_id(src);
-                if (sched->node_copies[id][cur_backend_id] == NULL) {
-                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                    sched->node_copies[id][cur_backend_id] = tensor_copy;
-                    node_allocr(tensor_copy) = cur_allocr;
-                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
-                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-                }
-                node->src[j] = sched->node_copies[id][cur_backend_id];
-            }
-        }
-    }
-    sched->splits[cur_split].i_end = graph->n_nodes;
-    sched->n_splits = cur_split + 1;
-
-    //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout);
-
-#if 1
-    // sanity check: all sources should have the same backend as the node
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
-        }
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
-                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
-                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
-                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
-            }
-        }
-    }
-#endif
-
-    // create copies of the graph for each split
-    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &sched->splits[i];
-        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
-
-        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
-            input_cpy->src[0] = input;
-            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
-        }
-
-        for (int j = split->i_start; j < split->i_end; j++) {
-            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
-        }
-    }
-    sched->graph = graph_copy;
-}
-
-static void sched_alloc_splits(ggml_backend_sched_t sched) {
-    ggml_gallocr_alloc_graph_n(
-        sched->galloc,
-        sched->graph,
-        sched->hash_set,
-        sched->node_talloc);
-}
-
-static void sched_compute_splits(ggml_backend_sched_t sched) {
-    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
-    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
-
-    struct ggml_backend_sched_split * splits = sched->splits;
-
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
-        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
-        int split_backend_id = sched_backend_prio(sched, split_backend);
-
-        // copy the input tensors to the split backend
-        uint64_t copy_start_us = ggml_time_us();
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
-            if (split->inputs[j]->buffer == NULL) {
-                if (split->inputs[j]->view_src == NULL) {
-                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
-                    exit(1);
-                }
-                struct ggml_tensor * view = split->inputs[j];
-                view->backend = view->view_src->backend;
-                view->buffer  = view->view_src->buffer;
-                view->data    = (char *)view->view_src->data + view->view_offs;
-                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
-            }
-            if (input_cpy->buffer == NULL) {
-                fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
-                exit(1);
-            }
-            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
-            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
-            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
-        }
-        // ggml_backend_synchronize(split_backend);
-        int64_t copy_end_us = ggml_time_us();
-        copy_us[split_backend_id] += copy_end_us - copy_start_us;
-
-#if 0
-        char split_filename[GGML_MAX_NAME];
-        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
-        ggml_graph_dump_dot(split->graph, NULL, split_filename);
-#endif
-
-        uint64_t compute_start_us = ggml_time_us();
-        ggml_backend_graph_compute(split_backend, split->graph);
-        // ggml_backend_synchronize(split_backend);
-        uint64_t compute_end_us = ggml_time_us();
-        compute_us[split_backend_id] += compute_end_us - compute_start_us;
-    }
-
-#if 0
-    // per-backend timings
-    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (copy_us[i] > 0 || compute_us[i] > 0) {
-            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
-        }
-    }
-#endif
-}
-
-static void sched_reset(ggml_backend_sched_t sched) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_reset(sched->tallocs[i]);
-    }
-}
-
-ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
-    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
-
-    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
-    memset(sched, 0, sizeof(struct ggml_backend_sched));
-
-    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
-
-    sched->n_backends = n_backends;
-    for (int i = 0; i < n_backends; i++) {
-        sched->backends[i] = backends[i];
-    }
-
-    sched->galloc = ggml_gallocr_new();
-
-    // init measure allocs for each backend
-    for (int i = 0; i < n_backends; i++) {
-        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
-    }
-
-    return sched;
-}
-
-void ggml_backend_sched_free(ggml_backend_sched_t sched) {
-    if (sched == NULL) {
-        return;
-    }
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_free(sched->tallocs[i]);
-    }
-    ggml_gallocr_free(sched->galloc);
-    free(sched->hash_set.keys);
-    free(sched->node_talloc);
-    free(sched->node_copies);
-    free(sched);
-}
-
-void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    // initialize hash tables
-    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
-    sched->hash_set.size = hash_size;
-    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
-    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
-    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
-
-    sched_split_graph(sched, measure_graph);
-    sched_alloc_splits(sched);
-
-    // allocate buffers and reset allocators
-    for (int i = 0; i < sched->n_backends; i++) {
-        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
-        ggml_tallocr_free(sched->tallocs[i]);
-        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
-    }
-
-    sched_reset(sched);
-}
-
-void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
-
-    sched_split_graph(sched, graph);
-    sched_alloc_splits(sched);
-    sched_compute_splits(sched);
-    sched_reset(sched);
-}
-
-ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return sched->tallocs[backend_index];
-}
-
-ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
-}
-
-void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
-    node_allocr(node) = sched->tallocs[backend_index];
-}

bindings/ruby/ext/ggml-backend.h

@@ -1,136 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    struct ggml_backend_buffer;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-
-    // backend buffer functions
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-
-    //
-    // Backend
-    //
-
-    struct ggml_backend;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
-
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-
-    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    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_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    //
-    // 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);
-
-    // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
-
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
-        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
-
-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
-
-        // in build_graph:
-        build_graph(...) {
-            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
-            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
-            ggml_allocr_alloc(alloc_cpu, tensor);
-
-            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
-            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
-        }
-
-        // allocate backend buffers from measure graph
-        ggml_backend_sched_init_measure(sched, measure_graph);
-
-        // the scheduler is now ready to compute graphs
-
-        // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
-
-    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
-
-    // Initialize backend buffers from a measure graph
-    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-
-    // Allocate a graph on the backend scheduler
-    GGML_API void ggml_backend_sched_graph_compute(
-            ggml_backend_sched_t sched,
-            struct ggml_cgraph * graph);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-impl.h

@@ -1,249 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#include <assert.h>
-#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// static_assert should be a #define, but if it's not,
-// fall back to the _Static_assert C11 keyword.
-// if C99 - static_assert is noop
-// ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#endif
-
-#undef MIN
-#undef MAX
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
-
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // __ARM_NEON
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
-
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-#endif
-
-#define GGML_HASHTABLE_FULL ((size_t)-1)
-#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
-
-bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
-size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
-size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// return index, asserts if table is full
-size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-#ifdef __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-quants.h

@@ -1,224 +0,0 @@
-#pragma once
-
-#include "ggml-impl.h"
-
-// GGML internal header
-
-#include <stdint.h>
-#include <stddef.h>
-
-#define QK4_0 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    ggml_fp16_t m;          // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // min
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    int8_t  qs[QK8_0];     // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    float d;               // delta
-    float s;               // d * sum(qs[i])
-    int8_t  qs[QK8_1];     // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.5625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    ggml_fp16_t d;           // super-block scale for quantized scales
-    ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;             // super-block scale for quantized scales
-    ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;               // super-block scale for quantized scales
-    ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-
-// Quantization
-void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
-void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
-void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
-void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
-void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
-void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
-
-void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
-void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
-void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
-void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
-void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
-void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
-
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
-
-void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
-
-// Dequantization
-void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
-void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
-void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
-void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
-void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
-//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
-
-void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
-void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
-void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
-void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

bindings/ruby/ext/metal-embed.mk

@@ -0,0 +1,14 @@
+ggml-metal-embed.o: \
+	ggml-metal.metal \
+	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}

bindings/ruby/ext/ruby_whisper.h

@@ -3,6 +3,13 @@
 
 #include "whisper.h"
 
+typedef struct {
+  VALUE *context;
+  VALUE user_data;
+  VALUE callback;
+  VALUE callbacks;
+} ruby_whisper_callback_container;
+
 typedef struct {
   struct whisper_context *context;
 } ruby_whisper;
@@ -10,6 +17,9 @@ typedef struct {
 typedef struct {
   struct whisper_full_params params;
   bool diarize;
+  ruby_whisper_callback_container *new_segment_callback_container;
+  ruby_whisper_callback_container *progress_callback_container;
+  ruby_whisper_callback_container *abort_callback_container;
 } ruby_whisper_params;
 
 #endif

bindings/ruby/extsources.yaml

@@ -0,0 +1,31 @@
+---
+- ../../src/whisper.cpp
+- ../../include/whisper.h
+- ../../ggml/src/ggml.c
+- ../../ggml/src/ggml-cpu.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-cpu.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/tests/helper.rb

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

bindings/ruby/tests/test_callback.rb

@@ -0,0 +1,163 @@
+require "test/unit"
+require "whisper"
+
+class TestCallback < Test::Unit::TestCase
+  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+
+  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')
+  end
+
+  def test_new_segment_callback
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_kind_of Integer, n_new
+      assert n_new > 0
+      assert_same @whisper, context
+
+      n_segments = context.full_n_segments
+      n_new.times do |i|
+        i_segment = n_segments - 1 + i
+        start_time = context.full_get_segment_t0(i_segment) * 10
+        end_time = context.full_get_segment_t1(i_segment) * 10
+        text = context.full_get_segment_text(i_segment)
+
+        assert_kind_of Integer, start_time
+        assert start_time >= 0
+        assert_kind_of Integer, end_time
+        assert end_time > 0
+        assert_match /ask not what your country can do for you, ask what you can do for your country/, text if i_segment == 0
+      end
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_new_segment_callback_closure
+    search_word = "what"
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      n_segments = context.full_n_segments
+      n_new.times do |i|
+        i_segment = n_segments - 1 + i
+        text = context.full_get_segment_text(i_segment)
+        if text.include?(search_word)
+          t0 = context.full_get_segment_t0(i_segment)
+          t1 = context.full_get_segment_t1(i_segment)
+          raise "search word '#{search_word}' found at between #{t0} and #{t1}"
+        end
+      end
+    }
+
+    assert_raise RuntimeError do
+      @whisper.transcribe(@audio, @params)
+    end
+  end
+
+  def test_new_segment_callback_user_data
+    udata = Object.new
+    @params.new_segment_callback_user_data = udata
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_same udata, user_data
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_new_segment_callback_user_data_gc
+    @params.new_segment_callback_user_data = "My user data"
+    @params.new_segment_callback = ->(context, state, n_new, user_data) {
+      assert_equal "My user data", user_data
+    }
+    GC.start
+
+    assert_same @whisper, @whisper.transcribe(@audio, @params)
+  end
+
+  def test_progress_callback
+    first = nil
+    last = nil
+    @params.progress_callback = ->(context, state, progress, user_data) {
+      assert_kind_of Integer, progress
+      assert 0 <= progress && progress <= 100
+      assert_same @whisper, context
+      first = progress if first.nil?
+      last = progress
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_equal 0, first
+    assert_equal 100, last
+  end
+
+  def test_progress_callback_user_data
+    udata = Object.new
+    @params.progress_callback_user_data = udata
+    @params.progress_callback = ->(context, state, n_new, user_data) {
+      assert_same udata, user_data
+    }
+
+    @whisper.transcribe(@audio, @params)
+  end
+
+  def test_on_progress
+    first = nil
+    last = nil
+    @params.on_progress do |progress|
+      assert_kind_of Integer, progress
+      assert 0 <= progress && progress <= 100
+      first = progress if first.nil?
+      last = progress
+    end
+    @whisper.transcribe(@audio, @params)
+    assert_equal 0, first
+    assert_equal 100, last
+  end
+
+  def test_abort_callback
+    i = 0
+    @params.abort_callback = ->(user_data) {
+      assert_nil user_data
+      i += 1
+      return false
+    }
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+
+  def test_abort_callback_abort
+    i = 0
+    @params.abort_callback = ->(user_data) {
+      i += 1
+      return i == 3
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_equal 3, i
+  end
+
+  def test_abort_callback_user_data
+    udata = Object.new
+    @params.abort_callback_user_data = udata
+    yielded = nil
+    @params.abort_callback = ->(user_data) {
+      yielded = user_data
+    }
+    @whisper.transcribe(@audio, @params)
+    assert_same udata, yielded
+  end
+
+  def test_abort_on
+    do_abort = false
+    aborted_from_callback = false
+    @params.on_new_segment do |segment|
+      do_abort = true if segment.text.match? /ask/
+    end
+    i = 0
+    @params.abort_on do
+      i += 1
+      do_abort
+    end
+    @whisper.transcribe(@audio, @params)
+    assert i > 0
+  end
+end

bindings/ruby/tests/test_model.rb

@@ -0,0 +1,44 @@
+require_relative "helper"
+
+class TestModel < TestBase
+  def test_model
+    whisper = Whisper::Context.new(MODEL)
+    assert_instance_of Whisper::Model, whisper.model
+  end
+
+  def test_attributes
+    whisper = Whisper::Context.new(MODEL)
+    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(MODEL).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
+end

bindings/ruby/tests/test_package.rb

@@ -0,0 +1,31 @@
+require_relative "helper"
+require 'tempfile'
+require 'tmpdir'
+require 'shellwords'
+
+class TestPackage < TestBase
+  def test_build
+    Tempfile.create do |file|
+      assert system("gem", "build", "whispercpp.gemspec", "--output", file.to_path.shellescape, exception: true)
+      assert file.size > 0
+      assert_path_exist file.to_path
+    end
+  end
+
+  sub_test_case "Building binary on installation" do
+    def setup
+      system "rake", "build", exception: true
+    end
+
+    def test_install
+      match_data = `rake -Tbuild`.match(/(whispercpp-(.+)\.gem)/)
+      filename = match_data[1]
+      version = match_data[2]
+      basename = "whisper.#{RbConfig::CONFIG["DLEXT"]}"
+      Dir.mktmpdir do |dir|
+        system "gem", "install", "--install-dir", dir.shellescape, "pkg/#{filename.shellescape}", exception: true
+        assert_path_exist File.join(dir, "gems/whispercpp-#{version}/lib", basename)
+      end
+    end
+  end
+end

bindings/ruby/tests/test_params.rb

@@ -0,0 +1,154 @@
+require_relative "helper"
+
+class TestParams < TestBase
+  def setup
+    @params  = Whisper::Params.new
+  end
+
+  def test_language
+    @params.language = "en"
+    assert_equal @params.language, "en"
+    @params.language = "auto"
+    assert_equal @params.language, "auto"
+  end
+
+  def test_offset
+    @params.offset = 10_000
+    assert_equal @params.offset, 10_000
+    @params.offset = 0
+    assert_equal @params.offset, 0
+  end
+
+  def test_duration
+    @params.duration = 60_000
+    assert_equal @params.duration, 60_000
+    @params.duration = 0
+    assert_equal @params.duration, 0
+  end
+
+  def test_max_text_tokens
+    @params.max_text_tokens = 300
+    assert_equal @params.max_text_tokens, 300
+    @params.max_text_tokens = 0
+    assert_equal @params.max_text_tokens, 0
+  end
+
+  def test_translate
+    @params.translate = true
+    assert @params.translate
+    @params.translate = false
+    assert !@params.translate
+  end
+
+  def test_no_context
+    @params.no_context = true
+    assert @params.no_context
+    @params.no_context = false
+    assert !@params.no_context
+  end
+
+  def test_single_segment
+    @params.single_segment = true
+    assert @params.single_segment
+    @params.single_segment = false
+    assert !@params.single_segment
+  end
+
+  def test_print_special
+    @params.print_special = true
+    assert @params.print_special
+    @params.print_special = false
+    assert !@params.print_special
+  end
+
+  def test_print_progress
+    @params.print_progress = true
+    assert @params.print_progress
+    @params.print_progress = false
+    assert !@params.print_progress
+  end
+
+  def test_print_realtime
+    @params.print_realtime = true
+    assert @params.print_realtime
+    @params.print_realtime = false
+    assert !@params.print_realtime
+  end
+
+  def test_print_timestamps
+    @params.print_timestamps = true
+    assert @params.print_timestamps
+    @params.print_timestamps = false
+    assert !@params.print_timestamps
+  end
+
+  def test_suppress_blank
+    @params.suppress_blank = true
+    assert @params.suppress_blank
+    @params.suppress_blank = false
+    assert !@params.suppress_blank
+  end
+
+  def test_suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = true
+    assert @params.suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = false
+    assert !@params.suppress_non_speech_tokens
+  end
+
+  def test_token_timestamps
+    @params.token_timestamps = true
+    assert @params.token_timestamps
+    @params.token_timestamps = false
+    assert !@params.token_timestamps
+  end
+
+  def test_split_on_word
+    @params.split_on_word = true
+    assert @params.split_on_word
+    @params.split_on_word = false
+    assert !@params.split_on_word
+  end
+
+  def test_initial_prompt
+    assert_nil @params.initial_prompt
+    @params.initial_prompt = "You are a polite person."
+    assert_equal "You are a polite person.", @params.initial_prompt
+  end
+
+  def test_temperature
+    assert_equal 0.0, @params.temperature
+    @params.temperature = 0.5
+    assert_equal 0.5, @params.temperature
+  end
+
+  def test_max_initial_ts
+    assert_equal 1.0, @params.max_initial_ts
+    @params.max_initial_ts = 600.0
+    assert_equal 600.0, @params.max_initial_ts
+  end
+
+  def test_length_penalty
+    assert_equal -1.0, @params.length_penalty
+    @params.length_penalty = 0.5
+    assert_equal 0.5, @params.length_penalty
+  end
+
+  def test_temperature_inc
+    assert_in_delta 0.2, @params.temperature_inc
+    @params.temperature_inc = 0.5
+    assert_in_delta 0.5, @params.temperature_inc
+  end
+
+  def test_entropy_thold
+    assert_in_delta 2.4, @params.entropy_thold
+    @params.entropy_thold = 3.0
+    assert_in_delta 3.0, @params.entropy_thold
+  end
+
+  def test_logprob_thold
+    assert_in_delta -1.0, @params.logprob_thold
+    @params.logprob_thold = -0.5
+    assert_in_delta -0.5, @params.logprob_thold
+  end
+end

bindings/ruby/tests/test_segment.rb

@@ -0,0 +1,83 @@
+require_relative "helper"
+
+class TestSegment < TestBase
+  class << self
+    attr_reader :whisper
+
+    def startup
+      @whisper = Whisper::Context.new(TestBase::MODEL)
+      params = Whisper::Params.new
+      params.print_timestamps = false
+      @whisper.transcribe(TestBase::AUDIO, params)
+    end
+  end
+
+  def test_iteration
+    whisper.each_segment do |segment|
+      assert_instance_of Whisper::Segment, segment
+    end
+  end
+
+  def test_enumerator
+    enum = whisper.each_segment
+    assert_instance_of Enumerator, enum
+    enum.to_a.each_with_index do |segment, index|
+      assert_instance_of Whisper::Segment, segment
+      assert_kind_of Integer, index
+    end
+  end
+
+  def test_start_time
+    i = 0
+    whisper.each_segment do |segment|
+      assert_equal 0, segment.start_time if i == 0
+      i += 1
+    end
+  end
+
+  def test_end_time
+    i = 0
+    whisper.each_segment do |segment|
+      assert_equal whisper.full_get_segment_t1(i) * 10, segment.end_time
+      i += 1
+    end
+  end
+
+  def test_on_new_segment
+    params = Whisper::Params.new
+    seg = nil
+    index = 0
+    params.on_new_segment do |segment|
+      assert_instance_of Whisper::Segment, segment
+      if index == 0
+        seg = segment
+        assert_equal 0, segment.start_time
+        assert_match /ask not what your country can do for you, ask what you can do for your country/, segment.text
+      end
+      index += 1
+    end
+    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
+
+  def test_on_new_segment_twice
+    params = Whisper::Params.new
+    seg = nil
+    params.on_new_segment do |segment|
+      seg = segment
+      return
+    end
+    params.on_new_segment do |segment|
+      assert_same seg, segment
+      return
+    end
+    whisper.transcribe(AUDIO, params)
+  end
+
+  private
+
+  def whisper
+    self.class.whisper
+  end
+end

bindings/ruby/tests/test_whisper.rb

@@ -1,138 +1,127 @@
-TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
-EXTDIR = File.join(TOPDIR, 'ext')
-#$LIBDIR = File.join(TOPDIR, 'lib')
-#$:.unshift(LIBDIR)
-$:.unshift(EXTDIR)
+require_relative "helper"
+require "stringio"
 
-require 'whisper'
-require 'test/unit'
+# Exists to detect memory-related bug
+Whisper.log_set ->(level, buffer, user_data) {}, nil
 
-class TestWhisper < Test::Unit::TestCase
+class TestWhisper < TestBase
   def setup
     @params  = Whisper::Params.new
   end
 
-  def test_language
-    @params.language = "en"
-    assert_equal @params.language, "en"
-    @params.language = "auto"
-    assert_equal @params.language, "auto"
-  end
-
-  def test_offset
-    @params.offset = 10_000
-    assert_equal @params.offset, 10_000
-    @params.offset = 0
-    assert_equal @params.offset, 0
-  end
-
-  def test_duration
-    @params.duration = 60_000
-    assert_equal @params.duration, 60_000
-    @params.duration = 0
-    assert_equal @params.duration, 0
-  end
-
-  def test_max_text_tokens
-    @params.max_text_tokens = 300
-    assert_equal @params.max_text_tokens, 300
-    @params.max_text_tokens = 0
-    assert_equal @params.max_text_tokens, 0
-  end
-
-  def test_translate
-    @params.translate = true
-    assert @params.translate
-    @params.translate = false
-    assert !@params.translate
-  end
-
-  def test_no_context
-    @params.no_context = true
-    assert @params.no_context
-    @params.no_context = false
-    assert !@params.no_context
-  end
-
-  def test_single_segment
-    @params.single_segment = true
-    assert @params.single_segment
-    @params.single_segment = false
-    assert !@params.single_segment
-  end
-
-  def test_print_special
-    @params.print_special = true
-    assert @params.print_special
-    @params.print_special = false
-    assert !@params.print_special
-  end
-
-  def test_print_progress
-    @params.print_progress = true
-    assert @params.print_progress
-    @params.print_progress = false
-    assert !@params.print_progress
-  end
-
-  def test_print_realtime
-    @params.print_realtime = true
-    assert @params.print_realtime
-    @params.print_realtime = false
-    assert !@params.print_realtime
-  end
-
-  def test_print_timestamps
-    @params.print_timestamps = true
-    assert @params.print_timestamps
-    @params.print_timestamps = false
-    assert !@params.print_timestamps
-  end
-
-  def test_suppress_blank
-    @params.suppress_blank = true
-    assert @params.suppress_blank
-    @params.suppress_blank = false
-    assert !@params.suppress_blank
-  end
-
-  def test_suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = true
-    assert @params.suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = false
-    assert !@params.suppress_non_speech_tokens
-  end
-
-  def test_token_timestamps
-    @params.token_timestamps = true
-    assert @params.token_timestamps
-    @params.token_timestamps = false
-    assert !@params.token_timestamps
-  end
-
-  def test_split_on_word
-    @params.split_on_word = true
-    assert @params.split_on_word
-    @params.split_on_word = false
-    assert !@params.split_on_word
-  end
-
-  def test_speed_up
-    @params.speed_up = true
-    assert @params.speed_up
-    @params.speed_up = false
-    assert !@params.speed_up
-  end
-
   def test_whisper
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    @whisper = Whisper::Context.new(MODEL)
     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
 
+  sub_test_case "After transcription" do
+    class << self
+      attr_reader :whisper
+
+      def startup
+        @whisper = Whisper::Context.new(TestBase::MODEL)
+        params = Whisper::Params.new
+        params.print_timestamps = false
+        @whisper.transcribe(TestBase::AUDIO, params)
+      end
+    end
+
+    def whisper
+      self.class.whisper
+    end
+
+    def test_full_n_segments
+      assert_equal 1, whisper.full_n_segments
+    end
+
+    def test_full_lang_id
+      assert_equal 0, whisper.full_lang_id
+    end
+
+    def test_full_get_segment_t0
+      assert_equal 0, whisper.full_get_segment_t0(0)
+      assert_raise IndexError do
+        whisper.full_get_segment_t0(whisper.full_n_segments)
+      end
+      assert_raise IndexError do
+        whisper.full_get_segment_t0(-1)
+      end
+    end
+
+    def test_full_get_segment_t1
+      t1 = whisper.full_get_segment_t1(0)
+      assert_kind_of Integer, t1
+      assert t1 > 0
+      assert_raise IndexError do
+        whisper.full_get_segment_t1(whisper.full_n_segments)
+      end
+    end
+
+    def test_full_get_segment_speaker_turn_next
+      assert_false whisper.full_get_segment_speaker_turn_next(0)
+    end
+
+    def test_full_get_segment_text
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, whisper.full_get_segment_text(0)
+    end
+  end
+
+  def test_lang_max_id
+    assert_kind_of Integer, Whisper.lang_max_id
+  end
+
+  def test_lang_id
+    assert_equal 0, Whisper.lang_id("en")
+    assert_raise ArgumentError do
+      Whisper.lang_id("non existing language")
+    end
+  end
+
+  def test_lang_str
+    assert_equal "en", Whisper.lang_str(0)
+    assert_raise IndexError do
+      Whisper.lang_str(Whisper.lang_max_id + 1)
+    end
+  end
+
+  def test_lang_str_full
+    assert_equal "english", Whisper.lang_str_full(0)
+    assert_raise IndexError do
+      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(MODEL)
+
+    assert logs.length > 30
+    logs.each do |log|
+      assert_equal Whisper::LOG_LEVEL_INFO, 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(MODEL)
+    assert_empty dev.string
+  ensure
+    $stderr = stderr
+  end
 end

bindings/ruby/whispercpp.gemspec

@@ -0,0 +1,36 @@
+require "yaml"
+
+Gem::Specification.new do |s|
+  s.name    = "whispercpp"
+  s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
+  s.version = '1.3.0'
+  s.date    = '2024-05-14'
+  s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
+  s.email   = 'todd.fisher@gmail.com'
+  s.extra_rdoc_files = ['LICENSE', 'README.md']
+  
+  s.files = `git ls-files . -z`.split("\x0") +
+              YAML.load_file("extsources.yaml").collect {|file|
+                basename = File.basename(file)
+                if s.extra_rdoc_files.include?(basename)
+                  basename
+                else
+                  File.join("ext", basename)
+                end
+              }
+
+  s.summary = %q{Ruby whisper.cpp bindings}
+  s.test_files = ["tests/test_whisper.rb"]
+  
+  s.extensions << 'ext/extconf.rb'
+  
+
+  #### Documentation and testing.
+  s.homepage = 'https://github.com/ggerganov/whisper.cpp'
+  s.rdoc_options = ['--main', '../../README.md']
+
+  
+    s.platform = Gem::Platform::RUBY
+  
+  s.licenses = ['MIT']
+end

cmake/BuildTypes.cmake

@@ -1,54 +0,0 @@
-# Add new build types
-
-# ReleaseGG - Release with enabled asserts
-
-SET(CMAKE_CXX_FLAGS_RELEASEGG
-    "-O3"
-    CACHE STRING "Flags used by the c++ compiler during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_C_FLAGS_RELEASEGG
-    "-O3"
-    CACHE STRING "Flags used by the compiler during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_EXE_LINKER_FLAGS_RELEASEGG
-    ""
-    CACHE STRING "Flags used for linking binaries during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_SHARED_LINKER_FLAGS_RELEASEGG
-    ""
-    CACHE STRING "Flags used by the shared libraries linker during release builds with enabled asserts."
-    FORCE )
-MARK_AS_ADVANCED(
-    CMAKE_CXX_FLAGS_RELEASEGG
-    CMAKE_C_FLAGS_RELEASEGG
-    CMAKE_EXE_LINKER_FLAGS_RELEASEGG
-    CMAKE_SHARED_LINKER_FLAGS_RELEASEGG )
-
-# RelWithDebInfoGG - RelWithDebInfo with enabled asserts
-
-SET(CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
-    "-O2 -g"
-    CACHE STRING "Flags used by the c++ compiler during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_C_FLAGS_RELWITHDEBINFOGG
-    "-O2 -g"
-    CACHE STRING "Flags used by the compiler during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
-    ""
-    CACHE STRING "Flags used for linking binaries during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG
-    ""
-    CACHE STRING "Flags used by the shared libraries linker during release builds with debug symbols and enabled asserts."
-    FORCE )
-MARK_AS_ADVANCED(
-    CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
-    CMAKE_C_FLAGS_RELWITHDEBINFOGG
-    CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
-    CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG )
-
-if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
-    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
-    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo" "ReleaseGG" "RelWithDebInfoGG")
-endif()

cmake/DefaultTargetOptions.cmake

@@ -13,5 +13,5 @@ set_target_properties(${TARGET}
     PROPERTIES
         EXPORT_COMPILE_COMMANDS ON
         RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin"
-        INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib"
+        INSTALL_RPATH            "${CMAKE_INSTALL_PREFIX}/lib"
 )

cmake/FindFFmpeg.cmake

@@ -0,0 +1,163 @@
+# From
+# https://github.com/snikulov/cmake-modules/blob/master/FindFFmpeg.cmake
+#
+# vim: ts=2 sw=2
+# - Try to find the required ffmpeg components(default: AVFORMAT, AVUTIL, AVCODEC)
+#
+# Once done this will define
+#  FFMPEG_FOUND         - System has the all required components.
+#  FFMPEG_INCLUDE_DIRS  - Include directory necessary for using the required components headers.
+#  FFMPEG_LIBRARIES     - Link these to use the required ffmpeg components.
+#  FFMPEG_DEFINITIONS   - Compiler switches required for using the required ffmpeg components.
+#
+# For each of the components it will additionally set.
+#   - AVCODEC
+#   - AVDEVICE
+#   - AVFORMAT
+#   - AVFILTER
+#   - AVUTIL
+#   - POSTPROC
+#   - SWSCALE
+# the following variables will be defined
+#  <component>_FOUND        - System has <component>
+#  <component>_INCLUDE_DIRS - Include directory necessary for using the <component> headers
+#  <component>_LIBRARIES    - Link these to use <component>
+#  <component>_DEFINITIONS  - Compiler switches required for using <component>
+#  <component>_VERSION      - The components version
+#
+# Copyright (c) 2006, Matthias Kretz, <kretz@kde.org>
+# Copyright (c) 2008, Alexander Neundorf, <neundorf@kde.org>
+# Copyright (c) 2011, Michael Jansen, <kde@michael-jansen.biz>
+#
+# Redistribution and use is allowed according to the terms of the BSD license.
+# For details see the accompanying COPYING-CMAKE-SCRIPTS file.
+
+include(FindPackageHandleStandardArgs)
+
+# The default components were taken from a survey over other FindFFMPEG.cmake files
+if (NOT FFmpeg_FIND_COMPONENTS)
+  set(FFmpeg_FIND_COMPONENTS AVFORMAT AVCODEC AVUTIL SWRESAMPLE)
+endif()
+
+#
+### Macro: set_component_found
+#
+# Marks the given component as found if both *_LIBRARIES AND *_INCLUDE_DIRS is present.
+#
+macro(set_component_found _component )
+  if (${_component}_LIBRARIES AND ${_component}_INCLUDE_DIRS)
+    message(DEBUG "  - ${_component} found.")
+    set(${_component}_FOUND TRUE)
+  else ()
+  message(DEBUG "  - ${_component} not found.")
+  endif ()
+endmacro()
+
+#
+### Macro: find_component
+#
+# Checks for the given component by invoking pkgconfig and then looking up the libraries and
+# include directories.
+#
+macro(find_component _component _pkgconfig _library _header)
+
+  if (NOT WIN32)
+     # use pkg-config to get the directories and then use these values
+     # in the FIND_PATH() and FIND_LIBRARY() calls
+     find_package(PkgConfig)
+     if (PKG_CONFIG_FOUND)
+       pkg_check_modules(PC_${_component} ${_pkgconfig})
+       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDEDIR}")
+       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDE_DIRS}")
+       message(STATUS "${PC_${_component}_CFLAGS}")
+     endif ()
+  endif (NOT WIN32)
+
+
+  find_path(${_component}_INCLUDE_DIRS ${_header}
+    HINTS
+      ${PC_${_component}_INCLUDEDIR}
+      ${PC_${_component}_INCLUDE_DIRS}
+    PATH_SUFFIXES
+      ffmpeg
+  )
+
+  # CMake's default is to search first for shared libraries and then for static libraries.
+  # Todo later: add option to prefer static libs over dynamic:
+  find_library(${_component}_LIBRARIES NAMES ${_library} lib${_library}.a
+      HINTS
+      ${PC_${_component}_LIBDIR}
+      ${PC_${_component}_LIBRARY_DIRS}
+  )
+
+  set(${_component}_DEFINITIONS  ${PC_${_component}_CFLAGS_OTHER} CACHE STRING "The ${_component} CFLAGS.")
+  set(${_component}_VERSION      ${PC_${_component}_VERSION}      CACHE STRING "The ${_component} version number.")
+
+  set_component_found(${_component})
+
+  mark_as_advanced(
+    ${_component}_INCLUDE_DIRS
+    ${_component}_LIBRARIES
+    ${_component}_DEFINITIONS
+    ${_component}_VERSION)
+
+endmacro()
+
+
+# Check for cached results. If there are skip the costly part.
+if (NOT FFMPEG_LIBRARIES)
+
+  # Check for all possible component.
+  find_component(AVCODEC    libavcodec    avcodec  libavcodec/avcodec.h)
+  find_component(AVFORMAT   libavformat   avformat libavformat/avformat.h)
+  find_component(AVDEVICE   libavdevice   avdevice libavdevice/avdevice.h)
+  #find_component(AVRESAMPLE libavresample avresample libavresample/avresample.h) # old name for swresample
+  find_component(AVUTIL     libavutil     avutil   libavutil/avutil.h)
+  find_component(AVFILTER   libavfilter   avfilter libavfilter/avfilter.h)
+  find_component(SWSCALE    libswscale    swscale  libswscale/swscale.h)
+  find_component(POSTPROC   libpostproc   postproc libpostproc/postprocess.h)
+  find_component(SWRESAMPLE libswresample swresample libswresample/swresample.h)
+
+  # Check if the required components were found and add their stuff to the FFMPEG_* vars.
+  foreach (_component ${FFmpeg_FIND_COMPONENTS})
+    if (${_component}_FOUND)
+      # message(STATUS "Required component ${_component} present.")
+      set(FFMPEG_LIBRARIES   ${FFMPEG_LIBRARIES}   ${${_component}_LIBRARIES})
+      set(FFMPEG_DEFINITIONS ${FFMPEG_DEFINITIONS} ${${_component}_DEFINITIONS})
+      list(APPEND FFMPEG_INCLUDE_DIRS ${${_component}_INCLUDE_DIRS})
+    else ()
+      # message(STATUS "Required component ${_component} missing.")
+    endif ()
+  endforeach ()
+
+  # Build the include path with duplicates removed.
+  if (FFMPEG_INCLUDE_DIRS)
+    list(REMOVE_DUPLICATES FFMPEG_INCLUDE_DIRS)
+  endif ()
+
+  # cache the vars.
+  set(FFMPEG_INCLUDE_DIRS ${FFMPEG_INCLUDE_DIRS} CACHE STRING "The FFmpeg include directories." FORCE)
+  set(FFMPEG_LIBRARIES    ${FFMPEG_LIBRARIES}    CACHE STRING "The FFmpeg libraries." FORCE)
+  set(FFMPEG_DEFINITIONS  ${FFMPEG_DEFINITIONS}  CACHE STRING "The FFmpeg cflags." FORCE)
+
+  mark_as_advanced(FFMPEG_INCLUDE_DIRS
+                   FFMPEG_LIBRARIES
+                   FFMPEG_DEFINITIONS)
+
+endif ()
+
+# Now set the noncached _FOUND vars for the components.
+# whisper.cpp does not need SWSCALE
+foreach (_component AVCODEC AVDEVICE AVFORMAT AVRESAMPLE AVUTIL POSTPROCESS)
+  set_component_found(${_component})
+endforeach ()
+
+# Compile the list of required vars
+set(_FFmpeg_REQUIRED_VARS FFMPEG_LIBRARIES FFMPEG_INCLUDE_DIRS)
+foreach (_component ${FFmpeg_FIND_COMPONENTS})
+  list(APPEND _FFmpeg_REQUIRED_VARS ${_component}_LIBRARIES ${_component}_INCLUDE_DIRS)
+endforeach ()
+
+# Give a nice error message if some of the required vars are missing.
+find_package_handle_standard_args(FFmpeg DEFAULT_MSG ${_FFmpeg_REQUIRED_VARS})
+

cmake/build-info.cmake

@@ -0,0 +1,58 @@
+set(BUILD_NUMBER 0)
+set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")
+
+# Look for git
+find_package(Git)
+if(NOT Git_FOUND)
+    find_program(GIT_EXECUTABLE NAMES git git.exe)
+    if(GIT_EXECUTABLE)
+        set(Git_FOUND TRUE)
+        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
+    else()
+        message(WARNING "Git not found. Build info will not be accurate.")
+    endif()
+endif()
+
+# Get the commit count and hash
+if(Git_FOUND)
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE HEAD
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_COMMIT ${HEAD})
+    endif()
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE COUNT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_NUMBER ${COUNT})
+    endif()
+endif()
+
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    execute_process(
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_COMPILER ${OUT})
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
+endif()

cmake/whisper-config.cmake.in

@@ -0,0 +1,65 @@
+set(WHISPER_VERSION      @WHISPER_INSTALL_VERSION@)
+set(WHISPER_BUILD_COMMIT @WHISPER_BUILD_COMMIT@)
+set(WHISPER_BUILD_NUMBER @WHISPER_BUILD_NUMBER@)
+set(WHISPER_SHARED_LIB   @BUILD_SHARED_LIBS@)
+
+set(GGML_BLAS       @GGML_BLAS@)
+set(GGML_CUDA       @GGML_CUDA@)
+set(GGML_METAL      @GGML_METAL@)
+set(GGML_HIPBLAS    @GGML_HIPBLAS@)
+set(GGML_ACCELERATE @GGML_ACCELERATE@)
+
+@PACKAGE_INIT@
+
+set_and_check(WHISPER_INCLUDE_DIR "@PACKAGE_WHISPER_INCLUDE_INSTALL_DIR@")
+set_and_check(WHISPER_LIB_DIR     "@PACKAGE_WHISPER_LIB_INSTALL_DIR@")
+set_and_check(WHISPER_BIN_DIR     "@PACKAGE_WHISPER_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (GGML_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (GGML_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (GGML_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+find_library(whisper_LIBRARY whisper
+    REQUIRED
+    HINTS ${WHISPER_LIB_DIR})
+
+set(_whisper_link_deps "Threads::Threads" "@WHISPER_EXTRA_LIBS@")
+set(_whisper_transient_defines "@WHISPER_TRANSIENT_DEFINES@")
+
+add_library(whisper UNKNOWN IMPORTED)
+
+set_target_properties(whisper
+    PROPERTIES
+    INTERFACE_INCLUDE_DIRECTORIES "${WHISPER_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_whisper_link_deps}"
+        INTERFACE_COMPILE_DEFINITIONS "${_whisper_transient_defines}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${whisper_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(whisper)

cmake/whisper.pc.in

@@ -0,0 +1,10 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=@CMAKE_INSTALL_FULL_LIBDIR@
+includedir=${prefix}/include
+
+Name: whisper
+Description: Port of OpenAI's Whisper model in C/C++
+Version: @PROJECT_VERSION@
+Libs: -L${libdir} -lwhisper
+Cflags: -I${includedir}

examples/CMakeLists.txt

@@ -11,7 +11,7 @@ if (WHISPER_SDL2)
     string(STRIP "${SDL2_LIBRARIES}" SDL2_LIBRARIES)
 
     message(STATUS "SDL2_INCLUDE_DIRS = ${SDL2_INCLUDE_DIRS}")
-    message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
+    message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
 endif()
 
 if (WHISPER_CLBLAST)
@@ -22,6 +22,35 @@ endif()
 
 set(TARGET common)
 
+unset(COMMON_EXTRA_LIBS)
+
+if (WHISPER_FFMPEG)
+    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
+    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
+    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
+    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
+    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
+    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
+    find_package(FFmpeg REQUIRED)
+
+    if (NOT ${FFMPEG_FOUND})
+        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
+    endif()
+
+    message(STATUS "Found ffmpeg libs:       ${FFMPEG_LIBRARIES}")
+    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
+    message(STATUS "ffmpeg definitions:      ${FFMPEG_DEFINITIONS}")
+    message(STATUS "Found avformat           ${AVFORMAT_VERSION}")
+
+    include_directories(${FFMPEG_INCLUDE_DIRS})
+    add_compile_definitions(WHISPER_FFMPEG)
+
+    list(APPEND COMMON_EXTRA_LIBS ${FFMPEG_LIBRARIES})
+
+    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
+endif()
+
+
 add_library(${TARGET} STATIC
     common.h
     common.cpp
@@ -29,11 +58,12 @@ add_library(${TARGET} STATIC
     common-ggml.cpp
     grammar-parser.h
     grammar-parser.cpp
+    ${COMMON_SOURCES_FFMPEG}
     )
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper)
+target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
@@ -50,16 +80,16 @@ if (WHISPER_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
+    target_include_directories(${TARGET} PUBLIC  ${SDL2_INCLUDE_DIRS})
+    target_link_libraries     (${TARGET} PRIVATE ${SDL2_LIBRARIES})
 
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
 # add json lib
-add_library(json_cpp INTERFACE json.hpp)
-set_target_properties(json_cpp PROPERTIES FOLDER "libs")
+add_library(json_cpp INTERFACE)
+target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
 
 # examples
 
@@ -72,8 +102,8 @@ if (EMSCRIPTEN)
     set_target_properties(libstream PROPERTIES FOLDER "libs")
     add_subdirectory(command.wasm)
     set_target_properties(libcommand PROPERTIES FOLDER "libs")
-    add_subdirectory(talk.wasm)
-    set_target_properties(libtalk PROPERTIES FOLDER "libs")
+    #add_subdirectory(talk.wasm)
+    #set_target_properties(libtalk PROPERTIES FOLDER "libs")
     add_subdirectory(bench.wasm)
     set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
@@ -97,15 +127,17 @@ endif (WHISPER_SDL2)
     add_subdirectory(quantize)
     set_target_properties(quantize PROPERTIES FOLDER "examples")
 if (WHISPER_SDL2)
-    add_subdirectory(talk)
-    set_target_properties(talk PROPERTIES FOLDER "examples")
+    # TODO: disabled until update
+    #       https://github.com/ggerganov/whisper.cpp/issues/1818
+    #add_subdirectory(talk)
+    #set_target_properties(talk PROPERTIES FOLDER "examples")
     add_subdirectory(talk-llama)
     set_target_properties(talk-llama PROPERTIES FOLDER "examples")
     add_subdirectory(lsp)
     set_target_properties(lsp PROPERTIES FOLDER "examples")
-    if (LLAMA_SYCL)
+    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/addon.node/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET whisper-addon)
+set(TARGET addon.node)
 
 # Base settings
 #==================================================================

examples/addon.node/README.md

@@ -14,14 +14,14 @@ npm install
 Make sure it is in the project root directory and compiled with make-js.
 
 ```shell
-npx cmake-js compile -T whisper-addon -B Release
+npx cmake-js compile -T addon.node -B Release
 ```
 
 For Electron addon and cmake-js options, you can see [cmake-js](https://github.com/cmake-js/cmake-js) and make very few configuration changes.
 
 > Such as appointing special cmake path:
 > ```shell
-> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -B Release
+> npx cmake-js compile -c 'xxx/cmake' -T addon.node -B Release
 > ```
 
 ## Run

examples/addon.node/__test__/whisper.spec.js

@@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
   __dirname,
-  "../../../build/Release/whisper-addon"
+  "../../../build/Release/addon.node"
 ));
 const { promisify } = require("util");
 
@@ -12,6 +12,12 @@ const whisperParamsMock = {
   model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
   use_gpu: true,
+  flash_attn: false,
+  no_prints: true,
+  comma_in_time: false,
+  translate: true,
+  no_timestamps: false,
+  audio_ctx: 0,
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -19,12 +19,12 @@ struct whisper_params {
     int32_t max_len      = 0;
     int32_t best_of      = 5;
     int32_t beam_size    = -1;
+    int32_t audio_ctx    = 0;
 
     float word_thold    = 0.01f;
     float entropy_thold = 2.4f;
     float logprob_thold = -1.0f;
 
-    bool speed_up       = false;
     bool translate      = false;
     bool diarize        = false;
     bool output_txt     = false;
@@ -36,7 +36,10 @@ struct whisper_params {
     bool print_colors   = false;
     bool print_progress = false;
     bool no_timestamps  = false;
+    bool no_prints      = false;
     bool use_gpu        = true;
+    bool flash_attn     = false;
+    bool comma_in_time  = true;
 
     std::string language = "en";
     std::string prompt;
@@ -44,6 +47,8 @@ struct whisper_params {
 
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
+
+    std::vector<float> pcmf32 = {}; // mono-channel F32 PCM
 };
 
 struct whisper_print_user_data {
@@ -120,9 +125,15 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
+void cb_log_disable(enum ggml_log_level, const char *, void *) {}
+
 int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
-    if (params.fname_inp.empty()) {
-        fprintf(stderr, "error: no input files specified\n");
+    if (params.no_prints) {
+        whisper_log_set(cb_log_disable, NULL);
+    }
+
+    if (params.fname_inp.empty() && params.pcmf32.empty()) {
+        fprintf(stderr, "error: no input files or audio buffer specified\n");
         return 2;
     }
 
@@ -135,6 +146,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     if (ctx == nullptr) {
@@ -142,6 +154,14 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         return 3;
     }
 
+    // if params.pcmf32 is provided, set params.fname_inp to "buffer"
+    // this is simpler than further modifications in the code
+    if (!params.pcmf32.empty()) {
+        fprintf(stderr, "info: using audio buffer as input\n");
+        params.fname_inp.clear();
+        params.fname_inp.emplace_back("buffer");
+    }
+
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
         const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -149,20 +169,25 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         std::vector<float> pcmf32; // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
-            continue;
+        // read the input audio file if params.pcmf32 is not provided
+        if (params.pcmf32.empty()) {
+            if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+                fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+                continue;
+            }
+        } else {
+            pcmf32 = params.pcmf32;
         }
 
         // print system information
-        {
+        if (!params.no_prints) {
             fprintf(stderr, "\n");
             fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
                     params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
         }
 
         // print some info about the processing
-        {
+        if (!params.no_prints) {
             fprintf(stderr, "\n");
             if (!whisper_is_multilingual(ctx)) {
                 if (params.language != "en" || params.translate) {
@@ -171,12 +196,13 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
                     fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                 }
             }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d, audio_ctx = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                     params.n_threads, params.n_processors,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
-                    params.no_timestamps ? 0 : 1);
+                    params.no_timestamps ? 0 : 1,
+                    params.audio_ctx);
 
             fprintf(stderr, "\n");
         }
@@ -203,8 +229,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
-
-            wparams.speed_up         = params.speed_up;
+            wparams.audio_ctx        = params.audio_ctx;
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
@@ -248,8 +273,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-        result[i].emplace_back(to_timestamp(t0, true));
-        result[i].emplace_back(to_timestamp(t1, true));
+        result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
+        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
         result[i].emplace_back(text);
     }
 
@@ -300,13 +325,33 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   std::string model = whisper_params.Get("model").As<Napi::String>();
   std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
   bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
+  bool flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
+  bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
   bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
+  int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
+  bool comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
+
+  Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
+  std::vector<float> pcmf32_vec;
+  if (pcmf32Value.IsTypedArray()) {
+    Napi::Float32Array pcmf32 = pcmf32Value.As<Napi::Float32Array>();
+    size_t length = pcmf32.ElementLength();
+    pcmf32_vec.reserve(length);
+    for (size_t i = 0; i < length; i++) {
+      pcmf32_vec.push_back(pcmf32[i]);
+    }
+  }
 
   params.language = language;
   params.model = model;
   params.fname_inp.emplace_back(input);
   params.use_gpu = use_gpu;
+  params.flash_attn = flash_attn;
+  params.no_prints = no_prints;
   params.no_timestamps = no_timestamps;
+  params.audio_ctx = audio_ctx;
+  params.pcmf32 = pcmf32_vec;
+  params.comma_in_time = comma_in_time;
 
   Napi::Function callback = info[1].As<Napi::Function>();
   Worker* worker = new Worker(callback, params);

examples/addon.node/index.js

@@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
   __dirname,
-  "../../build/Release/whisper-addon"
+  "../../build/Release/addon.node"
 ));
 const { promisify } = require("util");
 
@@ -10,15 +10,27 @@ const whisperAsync = promisify(whisper);
 const whisperParams = {
   language: "en",
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
-  fname_inp: "../../samples/jfk.wav",
+  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
   use_gpu: true,
+  flash_attn: false,
+  no_prints: true,
+  comma_in_time: false,
+  translate: true,
+  no_timestamps: false,
+  audio_ctx: 0,
 };
 
 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
   arguments.reduce((pre, item) => {
     if (item.startsWith("--")) {
-      return [...pre, item.slice(2).split("=")];
+      const [key, value] = item.slice(2).split("=");
+      if (key === "audio_ctx") {
+        whisperParams[key] = parseInt(value);
+      } else {
+        whisperParams[key] = value;
+      }
+      return pre;
     }
     return pre;
   }, [])
@@ -33,5 +45,6 @@ for (const key in params) {
 console.log("whisperParams =", whisperParams);
 
 whisperAsync(whisperParams).then((result) => {
-  console.log(`Result from whisper: ${result}`);
+  console.log();
+  console.log(result);
 });

examples/addon.node/package.json

@@ -1,5 +1,5 @@
 {
-  "name": "whisper-addon",
+  "name": "addon.node",
   "version": "0.0.0",
   "description": "",
   "main": "index.js",

examples/bench/bench.cpp

@@ -12,12 +12,13 @@ struct whisper_params {
 
     std::string model = "models/ggml-base.en.bin";
 
-    bool use_gpu = true;
+    bool use_gpu    = true;
+    bool flash_attn = false;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -25,10 +26,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"  || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
-        else if (arg == "-m"  || arg == "--model")   { params.model     = argv[++i]; }
-        else if (arg == "-w"  || arg == "--what")    { params.what      = atoi(argv[++i]); }
-        else if (arg == "-ng" || arg == "--no-gpu")  { params.use_gpu   = false; }
+        else if (arg == "-t"  || arg == "--threads")    { params.n_threads  = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")     { params.use_gpu    = false; }
+        else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -49,17 +51,20 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
     fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
     fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int whisper_bench_full(const whisper_params & params) {
+static int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 

examples/command/command.cpp

@@ -38,12 +38,12 @@ struct whisper_params {
 
     grammar_parser::parse_state grammar_parsed;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
     bool no_timestamps = true;
     bool use_gpu       = true;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -59,7 +59,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -75,11 +75,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
         else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
@@ -113,11 +113,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
     fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@@ -130,7 +130,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(
+static std::string transcribe(
                  whisper_context * ctx,
             const whisper_params & params,
         const std::vector<float> & pcmf32,
@@ -162,7 +162,6 @@ std::string transcribe(
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx = params.audio_ctx;
-    wparams.speed_up  = params.speed_up;
 
     wparams.temperature     = 0.4f;
     wparams.temperature_inc = 1.0f;
@@ -217,7 +216,7 @@ std::string transcribe(
     return result;
 }
 
-std::vector<std::string> read_allowed_commands(const std::string & fname) {
+static std::vector<std::string> read_allowed_commands(const std::string & fname) {
     std::vector<std::string> allowed_commands;
 
     std::ifstream ifs(fname);
@@ -239,7 +238,7 @@ std::vector<std::string> read_allowed_commands(const std::string & fname) {
     return allowed_commands;
 }
 
-std::vector<std::string> get_words(const std::string &txt) {
+static std::vector<std::string> get_words(const std::string &txt) {
     std::vector<std::string> words;
 
     std::istringstream iss(txt);
@@ -253,7 +252,7 @@ std::vector<std::string> get_words(const std::string &txt) {
 
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: guided mode\n", __func__);
 
@@ -368,7 +367,6 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
             wparams.n_threads        = params.n_threads;
 
             wparams.audio_ctx        = params.audio_ctx;
-            wparams.speed_up         = params.speed_up;
 
             wparams.prompt_tokens    = k_tokens.data();
             wparams.prompt_n_tokens  = k_tokens.size();
@@ -465,7 +463,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
 
 // always-prompt mode
 // transcribe the voice into text after valid prompt
-int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running = true;
     bool ask_prompt = true;
 
@@ -545,7 +543,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
 
 // general-purpose mode
 // freely transcribe the voice into text
-int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running  = true;
     bool have_prompt = false;
     bool ask_prompt  = true;
@@ -696,7 +694,9 @@ int main(int argc, char ** argv) {
     // whisper init
 
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 

examples/common-ggml.cpp

@@ -71,6 +71,10 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_NL:
         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;
@@ -207,6 +211,12 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_IQ4_NL:
                 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:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/common-sdl.cpp

@@ -219,7 +219,7 @@ bool sdl_poll_events() {
             case SDL_QUIT:
                 {
                     return false;
-                } break;
+                }
             default:
                 break;
         }

examples/common.cpp

@@ -24,8 +24,13 @@
 #include <io.h>
 #endif
 
+#ifdef WHISPER_FFMPEG
+// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
+extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
+#endif
+
 // Function to check if the next argument exists
-std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
         return argv[++i];
     } else {
@@ -142,7 +147,6 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
         case 7: return "He";
         case 8: return "She";
         case 9: return "They";
-        default: return "To";
     }
 
     return "The";
@@ -341,7 +345,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
-std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
     std::vector<gpt_vocab::id> output;
     std::stringstream ss(input);
     std::string token;
@@ -353,7 +357,7 @@ std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, ch
     return output;
 }
 
-std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
     if (fpath_test.empty()){
         fprintf(stderr, "%s : No test file found.\n", __func__);
         return std::map<std::string, std::vector<gpt_vocab::id>>();
@@ -637,7 +641,7 @@ bool is_wav_buffer(const std::string buf) {
 
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
-    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin or ffmpeg decoding output
 
     if (fname == "-") {
         {
@@ -670,8 +674,19 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
         }
     }
     else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+#if defined(WHISPER_FFMPEG)
+        if (ffmpeg_decode_audio(fname, wav_data) != 0) {
+            fprintf(stderr, "error: failed to ffmpeg decode '%s' \n", fname.c_str());
+            return false;
+        }
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to read wav data as wav \n");
+            return false;
+        }
+#else
         fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
         return false;
+#endif
     }
 
     if (wav.channels != 1 && wav.channels != 2) {

examples/common.h

@@ -9,6 +9,7 @@
 #include <thread>
 #include <ctime>
 #include <fstream>
+#include <sstream>
 
 #define COMMON_SAMPLE_RATE 16000
 
@@ -21,7 +22,7 @@ struct gpt_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
     int32_t n_predict    = 200;  // new tokens to predict
     int32_t n_parallel   = 1;    // number of parallel streams
-    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_batch      = 32;   // batch size for prompt processing
     int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
     int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
 
@@ -185,7 +186,7 @@ private:
     // It is assumed that PCM data is normalized to a range from -1 to 1
     bool write_audio(const float * data, size_t length) {
         for (size_t i = 0; i < length; ++i) {
-            const int16_t intSample = data[i] * 32767;
+            const int16_t intSample = int16_t(data[i] * 32767);
             file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
             dataSize += sizeof(int16_t);
         }
@@ -286,12 +287,43 @@ void sam_print_usage(int argc, char ** argv, const sam_params & params);
 // Terminal utils
 //
 
+#define SQR(X)    ((X) * (X))
+#define UNCUBE(x) x < 48 ? 0 : x < 115 ? 1 : (x - 35) / 40
 
-// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
-// Lowest is red, middle is yellow, highest is green.
+/**
+ * Quantizes 24-bit RGB to xterm256 code range [16,256).
+ */
+static int rgb2xterm256(int r, int g, int b) {
+    unsigned char cube[] = {0, 0137, 0207, 0257, 0327, 0377};
+    int av, ir, ig, ib, il, qr, qg, qb, ql;
+    av = r * .299 + g * .587 + b * .114 + .5;
+    ql = (il = av > 238 ? 23 : (av - 3) / 10) * 10 + 8;
+    qr = cube[(ir = UNCUBE(r))];
+    qg = cube[(ig = UNCUBE(g))];
+    qb = cube[(ib = UNCUBE(b))];
+    if (SQR(qr - r) + SQR(qg - g) + SQR(qb - b) <=
+        SQR(ql - r) + SQR(ql - g) + SQR(ql - b))
+        return ir * 36 + ig * 6 + ib + 020;
+    return il + 0350;
+}
+
+static std::string set_xterm256_foreground(int r, int g, int b) {
+    int x = rgb2xterm256(r, g, b);
+    std::ostringstream oss;
+    oss << "\033[38;5;" << x << "m";
+    return oss.str();
+}
+
+// Lowest is red, middle is yellow, highest is green. Color scheme from
+// Paul Tol; it is colorblind friendly https://personal.sron.nl/~pault/
 const std::vector<std::string> k_colors = {
-    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
-    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
+    set_xterm256_foreground(220,   5,  12),
+    set_xterm256_foreground(232,  96,  28),
+    set_xterm256_foreground(241, 147,  45),
+    set_xterm256_foreground(246, 193,  65),
+    set_xterm256_foreground(247, 240,  86),
+    set_xterm256_foreground(144, 201, 135),
+    set_xterm256_foreground( 78, 178, 101),
 };
 
 //

examples/ffmpeg-transcode.cpp

@@ -0,0 +1,348 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+/*
+ * transcode.c - convert audio file to WAVE
+ *
+ * Copyright (C) 2019		Andrew Clayton <andrew@digital-domain.net>
+ * Copyright (C) 2024       William Tambellini <william.tambellini@gmail.com>
+ */
+
+// Just for conveninent C++ API
+#include <vector>
+#include <string>
+
+// C
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <sys/mman.h>
+
+extern "C" {
+#include <libavutil/opt.h>
+#include <libavcodec/avcodec.h>
+#include <libavformat/avformat.h>
+#include <libswresample/swresample.h>
+}
+
+typedef uint64_t u64;
+typedef int64_t  s64;
+typedef uint32_t u32;
+typedef int32_t  s32;
+typedef uint16_t u16;
+typedef int16_t  s16;
+typedef uint8_t   u8;
+typedef int8_t    s8;
+
+#define WAVE_SAMPLE_RATE	16000
+#define AVIO_CTX_BUF_SZ		 4096
+
+static const char* ffmpegLog = getenv("FFMPEG_LOG");
+// Todo: add __FILE__ __LINE__
+#define LOG(...) \
+  do { if (ffmpegLog) fprintf(stderr, __VA_ARGS__); } while(0) // C99
+
+/*
+ * WAVE file header based on definition from
+ * https://gist.github.com/Jon-Schneider/8b7c53d27a7a13346a643dac9c19d34f
+ *
+ * We must ensure this structure doesn't have any holes or
+ * padding so we can just map it straight to the WAVE data.
+ */
+struct wave_hdr {
+	/* RIFF Header: "RIFF" */
+	char riff_header[4];
+	/* size of audio data + sizeof(struct wave_hdr) - 8 */
+	int wav_size;
+	/* "WAVE" */
+	char wav_header[4];
+
+	/* Format Header */
+	/* "fmt " (includes trailing space) */
+	char fmt_header[4];
+	/* Should be 16 for PCM */
+	int fmt_chunk_size;
+	/* Should be 1 for PCM. 3 for IEEE Float */
+	s16 audio_format;
+	s16 num_channels;
+	int sample_rate;
+	/*
+	 * Number of bytes per second
+	 * sample_rate * num_channels * bit_depth/8
+	 */
+	int byte_rate;
+	/* num_channels * bytes per sample */
+	s16 sample_alignment;
+	/* bits per sample */
+	s16 bit_depth;
+
+	/* Data Header */
+	/* "data" */
+	char data_header[4];
+	/*
+	 * size of audio
+	 * number of samples * num_channels * bit_depth/8
+	 */
+	int data_bytes;
+} __attribute__((__packed__));
+
+struct audio_buffer {
+	u8 *ptr;
+	int size; /* size left in the buffer */
+};
+
+static void set_wave_hdr(wave_hdr& wh, size_t size) {
+    memcpy(&wh.riff_header, "RIFF", 4);
+    wh.wav_size = size + sizeof(struct wave_hdr) - 8;
+    memcpy(&wh.wav_header, "WAVE", 4);
+    memcpy(&wh.fmt_header, "fmt ", 4);
+    wh.fmt_chunk_size = 16;
+    wh.audio_format = 1;
+    wh.num_channels = 1;
+    wh.sample_rate = WAVE_SAMPLE_RATE;
+    wh.sample_alignment = 2;
+    wh.bit_depth = 16;
+    wh.byte_rate = wh.sample_rate * wh.sample_alignment;
+    memcpy(&wh.data_header, "data", 4);
+    wh.data_bytes = size;
+}
+
+static void write_wave_hdr(int fd, size_t size) {
+	struct wave_hdr wh;
+    set_wave_hdr(wh, size);
+	write(fd, &wh, sizeof(struct wave_hdr));
+}
+
+static int map_file(int fd, u8 **ptr, size_t *size)
+{
+	struct stat sb;
+
+	fstat(fd, &sb);
+	*size = sb.st_size;
+
+    *ptr = (u8*)mmap(NULL, *size, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0);
+	if (*ptr == MAP_FAILED) {
+		perror("mmap");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int read_packet(void *opaque, u8 *buf, int buf_size)
+{
+    struct audio_buffer *audio_buf = (audio_buffer*)opaque;
+
+	buf_size = FFMIN(buf_size, audio_buf->size);
+
+	/* copy internal buffer data to buf */
+	memcpy(buf, audio_buf->ptr, buf_size);
+	audio_buf->ptr += buf_size;
+	audio_buf->size -= buf_size;
+
+	return buf_size;
+}
+
+static void convert_frame(struct SwrContext *swr, AVCodecContext *codec,
+			  AVFrame *frame, s16 **data, int *size, bool flush)
+{
+	int nr_samples;
+	s64 delay;
+	u8 *buffer;
+
+	delay = swr_get_delay(swr, codec->sample_rate);
+	nr_samples = av_rescale_rnd(delay + frame->nb_samples,
+				    WAVE_SAMPLE_RATE, codec->sample_rate,
+				    AV_ROUND_UP);
+	av_samples_alloc(&buffer, NULL, 1, nr_samples, AV_SAMPLE_FMT_S16, 0);
+
+	/*
+	 * !flush is used to check if we are flushing any remaining
+	 * conversion buffers...
+	 */
+	nr_samples = swr_convert(swr, &buffer, nr_samples,
+				 !flush ? (const u8 **)frame->data : NULL,
+				 !flush ? frame->nb_samples : 0);
+
+    *data = (s16*)realloc(*data, (*size + nr_samples) * sizeof(s16));
+	memcpy(*data + *size, buffer, nr_samples * sizeof(s16));
+	*size += nr_samples;
+	av_freep(&buffer);
+}
+
+static bool is_audio_stream(const AVStream *stream)
+{
+	if (stream->codecpar->codec_type == AVMEDIA_TYPE_AUDIO)
+		return true;
+
+	return false;
+}
+
+// Return non zero on error, 0 on success
+// audio_buffer: input memory
+// data: decoded output audio data (wav file)
+// size: size of output data
+static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
+{
+    LOG("decode_audio: input size: %d\n", audio_buf->size);
+	AVFormatContext *fmt_ctx;
+	AVIOContext *avio_ctx;
+	AVStream *stream;
+	AVCodecContext *codec;
+	AVPacket packet;
+	AVFrame *frame;
+	struct SwrContext *swr;
+	u8 *avio_ctx_buffer;
+	unsigned int i;
+	int stream_index = -1;
+	int err;
+    const size_t errbuffsize = 1024;
+    char errbuff[errbuffsize];
+
+    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);
+    avio_ctx = avio_alloc_context(avio_ctx_buffer, AVIO_CTX_BUF_SZ, 0, audio_buf, &read_packet, NULL, NULL);
+	fmt_ctx->pb = avio_ctx;
+
+    // open the input stream and read header
+	err = avformat_open_input(&fmt_ctx, NULL, NULL, NULL);
+	if (err) {
+        LOG("Could not read audio buffer: %d: %s\n", err, av_make_error_string(errbuff, errbuffsize, err));
+        return err;
+	}
+
+	err = avformat_find_stream_info(fmt_ctx, NULL);
+	if (err < 0) {
+        LOG("Could not retrieve stream info from audio buffer: %d\n", err);
+        return err;
+	}
+
+	for (i = 0; i < fmt_ctx->nb_streams; i++) {
+		if (is_audio_stream(fmt_ctx->streams[i])) {
+			stream_index = i;
+			break;
+		}
+	}
+
+	if (stream_index == -1) {
+        LOG("Could not retrieve audio stream from buffer\n");
+		return -1;
+	}
+
+	stream = fmt_ctx->streams[stream_index];
+	codec = avcodec_alloc_context3(
+			avcodec_find_decoder(stream->codecpar->codec_id));
+	avcodec_parameters_to_context(codec, stream->codecpar);
+	err = avcodec_open2(codec, avcodec_find_decoder(codec->codec_id),
+							NULL);
+	if (err) {
+        LOG("Failed to open decoder for stream #%d in audio buffer\n", stream_index);
+        return err;
+	}
+
+	/* prepare resampler */
+	swr = swr_alloc();
+
+	av_opt_set_int(swr, "in_channel_count", codec->channels, 0);
+	av_opt_set_int(swr, "out_channel_count", 1, 0);
+	av_opt_set_int(swr, "in_channel_layout", codec->channel_layout, 0);
+	av_opt_set_int(swr, "out_channel_layout", AV_CH_LAYOUT_MONO, 0);
+	av_opt_set_int(swr, "in_sample_rate", codec->sample_rate, 0);
+	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
+	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
+	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
+
+	swr_init(swr);
+	if (!swr_is_initialized(swr)) {
+        LOG("Resampler has not been properly initialized\n");
+		return -1;
+	}
+
+	av_init_packet(&packet);
+	frame = av_frame_alloc();
+	if (!frame) {
+        LOG("Error allocating the frame\n");
+		return -1;
+	}
+
+	/* iterate through frames */
+	*data = NULL;
+	*size = 0;
+	while (av_read_frame(fmt_ctx, &packet) >= 0) {
+		avcodec_send_packet(codec, &packet);
+
+		err = avcodec_receive_frame(codec, frame);
+		if (err == AVERROR(EAGAIN))
+			continue;
+
+		convert_frame(swr, codec, frame, data, size, false);
+	}
+	/* Flush any remaining conversion buffers... */
+	convert_frame(swr, codec, frame, data, size, true);
+
+	av_frame_free(&frame);
+	swr_free(&swr);
+    //avio_context_free(); // todo?
+	avcodec_close(codec);
+	avformat_close_input(&fmt_ctx);
+	avformat_free_context(fmt_ctx);
+
+	if (avio_ctx) {
+		av_freep(&avio_ctx->buffer);
+		av_freep(&avio_ctx);
+	}
+
+	return 0;
+}
+
+// in mem decoding/conversion/resampling:
+// ifname: input file path
+// owav_data: in mem wav file. Can be forwarded as it to whisper/drwav
+// return 0 on success
+int ffmpeg_decode_audio(const std::string &ifname, std::vector<uint8_t>& owav_data) {
+    LOG("ffmpeg_decode_audio: %s\n", ifname.c_str());
+    int ifd = open(ifname.c_str(), O_RDONLY);
+    if (ifd == -1) {
+        fprintf(stderr, "Couldn't open input file %s\n", ifname.c_str());
+        return -1;
+    }
+    u8 *ibuf = NULL;
+    size_t ibuf_size;
+    int err = map_file(ifd, &ibuf, &ibuf_size);
+    if (err) {
+        LOG("Couldn't map input file %s\n", ifname.c_str());
+        return err;
+    }
+    LOG("Mapped input file: %s size: %d\n", ibuf, (int) ibuf_size);
+    struct audio_buffer inaudio_buf;
+    inaudio_buf.ptr = ibuf;
+    inaudio_buf.size = ibuf_size;
+
+    s16 *odata=NULL;
+    int osize=0;
+
+    err = decode_audio(&inaudio_buf, &odata, &osize);
+    LOG("decode_audio returned %d \n", err);
+    if (err != 0) {
+        LOG("decode_audio failed\n");
+        return err;
+    }
+    LOG("decode_audio output size: %d\n", osize);
+
+    wave_hdr wh;
+    const size_t outdatasize = osize * sizeof(s16);
+    set_wave_hdr(wh, outdatasize);
+    owav_data.resize(sizeof(wave_hdr) + outdatasize);
+    // header:
+    memcpy(owav_data.data(), &wh, sizeof(wave_hdr));
+    // the data:
+    memcpy(owav_data.data() + sizeof(wave_hdr), odata, osize* sizeof(s16));
+
+    return 0;
+}

examples/grammar-parser.cpp

@@ -9,7 +9,7 @@
 namespace grammar_parser {
     // NOTE: assumes valid utf8 (but checks for overrun)
     // copied from whisper.cpp
-    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
         static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
         uint8_t  first_byte = static_cast<uint8_t>(*src);
         uint8_t  highbits   = first_byte >> 4;
@@ -24,19 +24,19 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
         return result.first->second;
     }
 
-    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
         return next_id;
     }
 
-    void add_rule(
+    static void add_rule(
             parse_state & state,
             uint32_t      rule_id,
             const std::vector<whisper_grammar_element> & rule) {
@@ -46,11 +46,11 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
-    bool is_word_char(char c) {
+    static bool is_word_char(char c) {
         return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
     }
 
-    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
         const char * pos   = src;
         const char * end   = src + size;
         uint32_t     value = 0;
@@ -73,7 +73,7 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    const char * parse_space(const char * src, bool newline_ok) {
+    static const char * parse_space(const char * src, bool newline_ok) {
         const char * pos = src;
         while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                 (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@@ -88,7 +88,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_name(const char * src) {
+    static const char * parse_name(const char * src) {
         const char * pos = src;
         while (is_word_char(*pos)) {
             pos++;
@@ -99,7 +99,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    std::pair<uint32_t, const char *> parse_char(const char * src) {
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
                 case 'x': return parse_hex(src + 2, 2);
@@ -122,14 +122,14 @@ namespace grammar_parser {
         throw std::runtime_error("unexpected end of input");
     }
 
-    const char * parse_alternates(
+    static const char * parse_alternates(
             parse_state       & state,
             const char        * src,
             const std::string & rule_name,
             uint32_t            rule_id,
             bool                is_nested);
 
-    const char * parse_sequence(
+    static const char * parse_sequence(
             parse_state                        & state,
             const char                         * src,
             const std::string                  & rule_name,
@@ -229,7 +229,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_alternates(
+    static const char * parse_alternates(
             parse_state       & state,
             const char        * src,
             const std::string & rule_name,
@@ -247,7 +247,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_rule(parse_state & state, const char * src) {
+    static const char * parse_rule(parse_state & state, const char * src) {
         const char * name_end = parse_name(src);
         const char * pos      = parse_space(name_end, false);
         size_t       name_len = name_end - src;
@@ -285,7 +285,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_grammar_char(FILE * file, uint32_t c) {
+    static void print_grammar_char(FILE * file, uint32_t c) {
         if (0x20 <= c && c <= 0x7f) {
             fprintf(file, "%c", static_cast<char>(c));
         } else {
@@ -294,7 +294,7 @@ namespace grammar_parser {
         }
     }
 
-    bool is_char_element(whisper_grammar_element elem) {
+    static bool is_char_element(whisper_grammar_element elem) {
         switch (elem.type) {
             case WHISPER_GRETYPE_CHAR:           return true;
             case WHISPER_GRETYPE_CHAR_NOT:       return true;
@@ -304,7 +304,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+    static void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
         for (auto elem : rule) {
             switch (elem.type) {
                 case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
@@ -334,7 +334,7 @@ namespace grammar_parser {
         fprintf(file, "\n");
     }
 
-    void print_rule(
+    static void print_rule(
             FILE     * file,
             uint32_t   rule_id,
             const std::vector<whisper_grammar_element> & rule,
@@ -413,7 +413,7 @@ namespace grammar_parser {
         }
     }
 
-    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const {
         std::vector<const whisper_grammar_element *> ret;
         for (const auto & rule : rules) {
             ret.push_back(rule.data());

examples/helpers.js

@@ -34,9 +34,6 @@ async function fetchRemote(url, cbProgress, cbPrint) {
         url,
         {
             method: 'GET',
-            headers: {
-                'Content-Type': 'application/octet-stream',
-            },
         }
     );
 

examples/livestream.sh

@@ -48,7 +48,7 @@ if [ -n "$3" ]; then
 fi
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" "large-v3-turbo" )
 
 # list available models
 function list_models {

examples/lsp/lsp.cpp

@@ -26,11 +26,11 @@ struct whisper_params {
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
     bool use_gpu       = true;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -53,7 +53,7 @@ struct commandset {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -69,11 +69,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
         else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else {
@@ -100,16 +100,16 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
     fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "\n");
 }
-uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
+static uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
     using namespace std::chrono;
     uint64_t time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
     uint64_t start_time = time_now;
@@ -153,7 +153,7 @@ uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params &
     return time_now;
 }
 
-json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
+static json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
     std::vector<whisper_token> prompt_tokens;
     std::vector<float> pcmf32;
     uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 10000U, pcmf32);
@@ -181,7 +181,6 @@ json unguided_transcription(struct whisper_context * ctx, audio_async &audio, js
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
     wparams.suppress_non_speech_tokens = true;
     // run the transformer and a single decoding pass
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
@@ -200,7 +199,7 @@ json unguided_transcription(struct whisper_context * ctx, audio_async &audio, js
 
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
+static json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
     struct commandset cs = commandset_list[jparams.value("commandset_index", commandset_list.size()-1)];
     std::vector<float> pcmf32;
     uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 2000U, pcmf32);
@@ -220,7 +219,6 @@ json guided_transcription(struct whisper_context * ctx, audio_async &audio, cons
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
 
     // TODO: Do some time testing. Does an overly long prompt slow down processing?
     // Set up command sets/precompute prompts
@@ -287,7 +285,7 @@ json guided_transcription(struct whisper_context * ctx, audio_async &audio, cons
     }
 }
 
-json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
+static json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
     // TODO: check for token collision
     struct commandset cs;
 
@@ -327,7 +325,8 @@ json register_commandset(struct whisper_context * ctx, json jparams, std::vector
     commandset_list.push_back(cs);
     return json{{"index",index}};
 }
-json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
+
+static json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
     // whisper_state has the pertinent offsets, but there also seem to be a large
     // number of scratch buffers that would prevent rewinding context in a manner similar to llama
     // I'll give this a another pass once everything else is implemented,
@@ -337,7 +336,8 @@ json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*para
             {"message", "Seeking is not yet supported."}
     };
 }
-json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
+
+static json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
     // See: https://www.jsonrpc.org/specification
     json id = body.at("id");
     try {
@@ -377,7 +377,7 @@ json parse_job(const json &body, struct whisper_context * ctx, audio_async &audi
     }
 }
 
-void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+static void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
     std::deque<json> jobqueue;
     std::vector<struct commandset> commandset_list;
     while (true) {
@@ -436,7 +436,10 @@ int main(int argc, char ** argv) {
 
     // whisper init
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
     // init audio
 

examples/main/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

examples/main/main.cpp

@@ -17,7 +17,7 @@
 #endif
 
 // helper function to replace substrings
-void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     for (size_t pos = 0; ; pos += replace.length()) {
         pos = s.find(search, pos);
         if (pos == std::string::npos) break;
@@ -44,8 +44,9 @@ struct whisper_params {
     float entropy_thold   =  2.40f;
     float logprob_thold   = -1.00f;
     float grammar_penalty = 100.0f;
+    float temperature     = 0.0f;
+    float temperature_inc = 0.2f;
 
-    bool speed_up        = false;
     bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
@@ -68,6 +69,7 @@ struct whisper_params {
     bool no_timestamps   = false;
     bool log_score       = false;
     bool use_gpu         = true;
+    bool flash_attn      = false;
 
     std::string language  = "en";
     std::string prompt;
@@ -92,17 +94,17 @@ struct whisper_params {
     grammar_parser::parse_state grammar_parsed;
 };
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+static void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-char* whisper_param_turn_lowercase(char* in){
+static char * whisper_param_turn_lowercase(char * in){
     int string_len = strlen(in);
-    for(int i = 0; i < string_len; i++){
+    for (int i = 0; i < string_len; i++){
         *(in+i) = tolower((unsigned char)*(in+i));
     }
     return in;
 }
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -133,7 +135,8 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
+        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
+        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@@ -164,7 +167,8 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
         else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
         else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        else if (                  arg == "--suppress-regex")  { params.suppress_regex = argv[++i]; }
+        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = argv[++i]; }
         else if (                  arg == "--grammar")         { params.grammar         = argv[++i]; }
         else if (                  arg == "--grammar-rule")    { params.grammar_rule    = argv[++i]; }
         else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
@@ -178,7 +182,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
     fprintf(stderr, "\n");
@@ -198,7 +202,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
+    fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@@ -228,6 +233,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
     fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
     fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
     fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
     fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
     fprintf(stderr, "  --grammar-rule RULE            [%-7s] top-level GBNF grammar rule name\n",               params.grammar_rule.c_str());
@@ -242,7 +248,7 @@ struct whisper_print_user_data {
     int progress_prev;
 };
 
-std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+static std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
 
@@ -274,7 +280,8 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
 
     return speaker;
 }
-void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+
+static void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
     int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
     int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
     if (progress >= *progress_prev + progress_step) {
@@ -283,7 +290,7 @@ void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct wh
     }
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+static void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -352,7 +359,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
-bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -379,7 +386,7 @@ bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -411,7 +418,7 @@ bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -440,7 +447,7 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-char *escape_double_quotes_and_backslashes(const char *str) {
+static char * escape_double_quotes_and_backslashes(const char * str) {
     if (str == NULL) {
         return NULL;
     }
@@ -453,7 +460,7 @@ char *escape_double_quotes_and_backslashes(const char *str) {
         }
     }
 
-    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    char * escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
     if (escaped == NULL) {
         return NULL;
     }
@@ -471,7 +478,39 @@ char *escape_double_quotes_and_backslashes(const char *str) {
     return escaped;
 }
 
-bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+// double quote should be escaped by another double quote. (rfc4180)
+static char * escape_double_quotes_in_csv(const char * str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped[pos++] = '"';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+static bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -492,7 +531,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-        char * text_escaped = escape_double_quotes_and_backslashes(text);
+        char * text_escaped = escape_double_quotes_in_csv(text);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
         fout << 10 * t0 << "," << 10 * t1 << ",";
@@ -506,7 +545,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+static bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
     std::ofstream fout(fname);
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
@@ -525,7 +564,7 @@ bool output_score(struct whisper_context * ctx, const char * fname, const whispe
     return true;
 }
 
-bool output_json(
+static bool output_json(
              struct whisper_context * ctx,
                          const char * fname,
                const whisper_params & params,
@@ -696,7 +735,7 @@ bool output_json(
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
+static bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@@ -821,7 +860,7 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
     return true;
 }
 
-bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -862,7 +901,7 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
 }
 
 
-void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
 
 int main(int argc, char ** argv) {
     whisper_params params;
@@ -939,7 +978,9 @@ int main(int argc, char ** argv) {
     // whisper init
 
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
 
     if (!params.dtw.empty()) {
         cparams.dtw_token_timestamps = true;
@@ -956,6 +997,7 @@ int main(int argc, char ** argv) {
         if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
         if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
         if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
+        if (params.dtw == "large.v3.turbo")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3_TURBO;
 
         if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
             fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
@@ -1063,19 +1105,20 @@ int main(int argc, char ** argv) {
             wparams.split_on_word    = params.split_on_word;
             wparams.audio_ctx        = params.audio_ctx;
 
-            wparams.speed_up         = params.speed_up;
             wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 
-            wparams.suppress_regex   = params.suppress_regex.c_str();
+            wparams.suppress_regex   = params.suppress_regex.empty() ? nullptr : params.suppress_regex.c_str();
 
             wparams.initial_prompt   = params.prompt.c_str();
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : params.temperature_inc;
+            wparams.temperature      = params.temperature;
+
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
@@ -1193,7 +1236,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    whisper_print_timings(ctx);
+    if (!params.no_prints) {
+        whisper_print_timings(ctx);
+    }
     whisper_free(ctx);
 
     return 0;

examples/python/whisper_processor.py

@@ -21,7 +21,7 @@ def process_audio(wav_file, model_name="base.en"):
     if not os.path.exists(wav_file):
         raise FileNotFoundError(f"WAV file not found: {wav_file}")
 
-    full_command = f"./main -m {model} -f {wav_file} -np -nt"
+    full_command = f"./main -m {model} -f {wav_file} -nt"
 
     # Execute the command
     process = subprocess.Popen(full_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)

examples/quantize/quantize.cpp

@@ -36,7 +36,7 @@ struct whisper_filters {
 };
 
 // quantize a model
-bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+static bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
     gpt_vocab vocab;
 
     printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());

examples/server/server.cpp

@@ -34,6 +34,7 @@ struct server_params
     std::string hostname = "127.0.0.1";
     std::string public_path = "examples/server/public";
     std::string request_path = "";
+    std::string inference_path = "/inference";
 
     int32_t port          = 8080;
     int32_t read_timeout  = 600;
@@ -61,7 +62,6 @@ struct whisper_params {
     float temperature     =  0.00f;
     float temperature_inc =  0.20f;
 
-    bool speed_up        = false;
     bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
@@ -75,6 +75,7 @@ struct whisper_params {
     bool print_progress  = false;
     bool no_timestamps   = false;
     bool use_gpu         = true;
+    bool flash_attn      = false;
 
     std::string language        = "en";
     std::string prompt          = "";
@@ -111,7 +112,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@@ -133,6 +133,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
     fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
     fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
+    fprintf(stderr, "  --inference-path PATH,         [%-7s] Inference path for all requests\n", sparams.inference_path.c_str());
     fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
     fprintf(stderr, "\n");
 }
@@ -158,7 +159,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@@ -178,11 +178,13 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
         else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
         else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
+        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
         // server params
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
         else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
         else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
         else if (                  arg == "--request-path")    { sparams.request_path = argv[++i]; }
+        else if (                  arg == "--inference-path")  { sparams.inference_path = argv[++i]; }
         else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -217,7 +219,7 @@ void check_ffmpeg_availibility() {
 bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
     std::ostringstream cmd_stream;
     std::string converted_filename_temp = temp_filename + "_temp.wav";
-    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
+    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -y -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
     std::string cmd = cmd_stream.str();
 
     int status = std::system(cmd.c_str());
@@ -502,7 +504,10 @@ int main(int argc, char ** argv) {
     }
     // whisper init
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
     if (!params.dtw.empty()) {
         cparams.dtw_token_timestamps = true;
         cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
@@ -642,10 +647,10 @@ int main(int argc, char ** argv) {
         return false;
     });
 
-    svr.Options(sparams.request_path + "/inference", [&](const Request &, Response &){
+    svr.Options(sparams.request_path + sparams.inference_path, [&](const Request &, Response &){
     });
 
-    svr.Post(sparams.request_path + "/inference", [&](const Request &req, Response &res){
+    svr.Post(sparams.request_path + sparams.inference_path, [&](const Request &req, Response &res){
         // acquire whisper model mutex lock
         std::lock_guard<std::mutex> lock(whisper_mutex);
 
@@ -672,7 +677,8 @@ int main(int argc, char ** argv) {
         if (sparams.ffmpeg_converter) {
             // if file is not wav, convert to wav
             // write to temporary file
-            const std::string temp_filename = "whisper_server_temp_file.wav";
+            const std::string temp_filename_base = std::tmpnam(nullptr);
+            const std::string temp_filename = temp_filename_base + ".wav";
             std::ofstream temp_file{temp_filename, std::ios::binary};
             temp_file << audio_file.content;
             temp_file.close();
@@ -763,7 +769,6 @@ int main(int argc, char ** argv) {
             wparams.split_on_word    = params.split_on_word;
             wparams.audio_ctx        = params.audio_ctx;
 
-            wparams.speed_up         = params.speed_up;
             wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
@@ -831,7 +836,7 @@ int main(int argc, char ** argv) {
         if (params.response_format == text_format)
         {
             std::string results = output_str(ctx, params, pcmf32s);
-            res.set_content(results.c_str(), "text/html");
+            res.set_content(results.c_str(), "text/html; charset=utf-8");
         }
         else if (params.response_format == srt_format)
         {
@@ -942,7 +947,7 @@ int main(int argc, char ** argv) {
                             "application/json");
         }
 
-        // reset params to thier defaults
+        // reset params to their defaults
         params = default_params;
     });
     svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){

examples/stream/stream.cpp

@@ -27,7 +27,6 @@ struct whisper_params {
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool no_fallback   = false;
     bool print_special = false;
@@ -36,6 +35,7 @@ struct whisper_params {
     bool tinydiarize   = false;
     bool save_audio    = false; // save audio to wav file
     bool use_gpu       = true;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -44,7 +44,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -61,7 +61,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
         else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
@@ -72,6 +71,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-tdrz" || arg == "--tinydiarize")   { params.tinydiarize   = true; }
         else if (arg == "-sa"   || arg == "--save-audio")    { params.save_audio    = true; }
         else if (arg == "-ng"   || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"   || arg == "--flash-attn")    { params.flash_attn    = true; }
 
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -98,7 +98,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
     fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
@@ -109,6 +108,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
     fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
     fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention during inference\n",               params.flash_attn ? "true" : "false");
     fprintf(stderr, "\n");
 }
 
@@ -153,7 +153,9 @@ int main(int argc, char ** argv) {
     }
 
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
@@ -309,7 +311,6 @@ int main(int argc, char ** argv) {
             wparams.n_threads        = params.n_threads;
 
             wparams.audio_ctx        = params.audio_ctx;
-            wparams.speed_up         = params.speed_up;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 

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/CMakeLists.txt

@@ -1,7 +1,13 @@
 if (WHISPER_SDL2)
     # talk-llama
     set(TARGET talk-llama)
-    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp unicode-data.cpp)
+    add_executable(${TARGET} talk-llama.cpp
+        llama.cpp
+        llama-vocab.cpp
+        llama-grammar.cpp
+        llama-sampling.cpp
+        unicode.cpp
+        unicode-data.cpp)
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
 
     if (WHISPER_CLBLAST)

examples/talk-llama/llama-grammar.h

@@ -0,0 +1,144 @@
+#pragma once
+
+#include "llama-impl.h"
+
+#include <map>
+
+struct llama_vocab;
+
+// grammar element type
+enum llama_gretype {
+    // end of rule definition
+    LLAMA_GRETYPE_END            = 0,
+
+    // start of alternate definition for rule
+    LLAMA_GRETYPE_ALT            = 1,
+
+    // non-terminal element: reference to rule
+    LLAMA_GRETYPE_RULE_REF       = 2,
+
+    // terminal element: character (code point)
+    LLAMA_GRETYPE_CHAR           = 3,
+
+    // inverse char(s) ([^a], [^a-b] [^abc])
+    LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+    // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+    // be an inclusive range ([a-z])
+    LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+    // modifies a preceding LLAMA_GRETYPE_CHAR or
+    // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+    LLAMA_GRETYPE_CHAR_ALT       = 6,
+
+    // any character (.)
+    LLAMA_GRETYPE_CHAR_ANY       = 7,
+};
+
+typedef struct llama_grammar_element {
+    enum llama_gretype type;
+    uint32_t           value; // Unicode code point or rule ID
+} llama_grammar_element;
+
+struct llama_partial_utf8 {
+    uint32_t value;    // bit value so far (unshifted)
+    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
+};
+
+struct llama_grammar_candidate {
+    size_t               index;
+    const uint32_t     * code_points;
+    llama_partial_utf8   partial_utf8;
+};
+
+using llama_grammar_rule  = std::vector<      llama_grammar_element>;
+using llama_grammar_stack = std::vector<const llama_grammar_element *>;
+
+using llama_grammar_rules      = std::vector<llama_grammar_rule>;
+using llama_grammar_stacks     = std::vector<llama_grammar_stack>;
+using llama_grammar_candidates = std::vector<llama_grammar_candidate>;
+
+const llama_grammar_rules  & llama_grammar_get_rules (const struct llama_grammar * grammar);
+      llama_grammar_stacks & llama_grammar_get_stacks(      struct llama_grammar * grammar);
+
+// takes a set of possible pushdown stacks on a grammar, which are required to
+// be positioned at a character range (see `llama_grammar_advance_stack`), and
+// produces the N possible stacks if the given char is accepted at those
+// positions
+void llama_grammar_accept(
+        const llama_grammar_rules  & rules,
+        const llama_grammar_stacks & stacks,
+                          uint32_t   chr,
+              llama_grammar_stacks & stacks_new);
+
+std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
+        const llama_grammar_rules      & rules,
+        const llama_grammar_stack      & stack,
+        const llama_grammar_candidates & candidates);
+
+struct llama_grammar_parser {
+    std::map<std::string, uint32_t> symbol_ids;
+
+    llama_grammar_rules rules;
+
+    llama_grammar_stack c_rules() const;
+
+    uint32_t get_symbol_id(const char * src, size_t len);
+    uint32_t generate_symbol_id(const std::string & base_name);
+
+    void add_rule(uint32_t rule_id, const llama_grammar_rule & rule);
+
+    const char * parse_alternates(
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            const char         * src,
+            const std::string  & rule_name,
+            llama_grammar_rule & rule,
+            bool               is_nested);
+
+    const char * parse_rule(const char * src);
+
+    bool parse(const char * src);
+    void print(FILE * file);
+};
+
+struct llama_grammar {
+    // note: allow null vocab for testing (not great)
+    const llama_vocab * vocab;
+
+    const llama_grammar_rules  rules;  // TODO: shared ptr
+          llama_grammar_stacks stacks;
+
+    // buffer for partially generated UTF-8 sequence from accepted tokens
+    llama_partial_utf8 partial_utf8;
+};
+
+//
+// internal API
+//
+
+// note: needed for tests (not great)
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+        const llama_grammar_element ** rules,
+        size_t n_rules,
+        size_t start_rule_index);
+
+struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root);
+
+void llama_grammar_free_impl(struct llama_grammar * grammar);
+
+struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar);
+
+// TODO: move the API below as member functions of llama_grammar
+void llama_grammar_apply_impl(
+        const struct llama_grammar & grammar,
+            llama_token_data_array * cur_p);
+
+void llama_grammar_accept_impl(
+              struct llama_grammar & grammar,
+                       llama_token   token);

examples/talk-llama/llama-impl.h

@@ -0,0 +1,181 @@
+#pragma once
+
+#include "llama.h"
+
+#include <string>
+#include <vector>
+#include <stdexcept>
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_ATTRIBUTE_FORMAT(...)
+#endif
+
+//
+// logging
+//
+
+LLAMA_ATTRIBUTE_FORMAT(2, 3)
+void llama_log_internal        (ggml_log_level level, const char * format, ...);
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
+
+#define LLAMA_LOG(...)       llama_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
+#define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LLAMA_LOG_DEBUG(...) llama_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LLAMA_LOG_CONT(...)  llama_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__)
+
+//
+// helpers
+//
+
+struct time_meas {
+    time_meas(int64_t & t_acc, bool disable = false) : t_start_us(disable ? -1 : ggml_time_us()), t_acc(t_acc) {}
+
+    ~time_meas() {
+        if (t_start_us >= 0) {
+            t_acc += ggml_time_us() - t_start_us;
+        }
+    }
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);
+
+// the ring buffer works similarly to std::deque, but with a fixed capacity
+template<typename T>
+struct ring_buffer {
+    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
+
+    T & front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    const T & front() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[first];
+    }
+
+    T & back() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    const T & back() const {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        return data[pos];
+    }
+
+    void push_back(const T & value) {
+        if (capacity == 0) {
+            throw std::runtime_error("ring buffer: capacity is zero");
+        }
+
+        if (sz == capacity) {
+            // advance the start when buffer is full
+            first = (first + 1) % capacity;
+        } else {
+            sz++;
+        }
+        data[pos] = value;
+        pos = (pos + 1) % capacity;
+    }
+
+    T pop_front() {
+        if (sz == 0) {
+            throw std::runtime_error("ring buffer is empty");
+        }
+        T value = data[first];
+        first = (first + 1) % capacity;
+        sz--;
+        return value;
+    }
+
+    //T & operator[](size_t i) {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    //const T & at(size_t i) const {
+    //    if (i >= sz) {
+    //        throw std::runtime_error("ring buffer: index out of bounds");
+    //    }
+    //    return data[(first + i) % capacity];
+    //}
+
+    const T & rat(size_t i) const {
+        if (i >= sz) {
+            throw std::runtime_error("ring buffer: index out of bounds");
+        }
+        return data[(first + sz - i - 1) % capacity];
+    }
+
+    std::vector<T> to_vector() const {
+        std::vector<T> result;
+        result.reserve(sz);
+        for (size_t i = 0; i < sz; i++) {
+            result.push_back(data[(first + i) % capacity]);
+        }
+        return result;
+    }
+
+    void clear() {
+        // here only reset the status of the buffer
+        sz = 0;
+        first = 0;
+        pos = 0;
+    }
+
+    bool empty() const {
+        return sz == 0;
+    }
+
+    size_t size() const {
+        return sz;
+    }
+
+    size_t capacity = 0;
+    size_t sz = 0;
+    size_t first = 0;
+    size_t pos = 0;
+    std::vector<T> data;
+};

examples/talk-llama/llama-sampling.h

@@ -0,0 +1,48 @@
+#pragma once
+
+// TODO: rename llama-sampling.h/.cpp to llama-sampler.h/.cpp ?
+
+#include "llama-grammar.h"
+
+struct llama_vocab;
+struct llama_grammar;
+
+// sampler chain
+
+struct llama_sampler_chain {
+    llama_sampler_chain_params params;
+
+    std::vector<struct llama_sampler *> samplers;
+
+    // timing
+
+    mutable int64_t t_sample_us;
+
+    mutable int32_t n_sample;
+};
+
+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.h

@@ -0,0 +1,170 @@
+#pragma once
+
+#include "llama-impl.h"
+
+#include <string>
+#include <vector>
+#include <unordered_map>
+#include <map>
+#include <set>
+
+struct llm_tokenizer;
+
+struct llama_vocab {
+    using id    = llama_token;
+    using token = std::string;
+    using tattr = llama_token_attr;
+
+    struct token_data {
+        token text;
+        float score;
+        tattr attr;
+    };
+
+    uint32_t n_vocab = 0; // TODO: not great because has to keep in sync with hparams.n_vocab
+
+    enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
+    enum llama_vocab_pre_type type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
+
+    int max_token_len = 0; // used for optimizing longest token search
+
+    std::unordered_map<token, id> token_to_id;
+    std::vector<token_data>       id_to_token;
+
+    std::vector<id>    cache_special_tokens;
+    std::vector<token> cache_token_to_piece; // llama_token_to_piece(special = true);
+
+    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  = 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;
+
+    // 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;
+
+    // tokenizer flags
+    bool tokenizer_add_space_prefix           = false;
+    bool tokenizer_add_bos                    = false;
+    bool tokenizer_add_eos                    = false;
+    bool tokenizer_ignore_merges              = false;
+    bool tokenizer_clean_spaces               = false;  // clean_up_tokenization_spaces
+    bool tokenizer_remove_extra_whitespaces   = false;
+    bool tokenizer_escape_whitespaces         = true;
+    bool tokenizer_treat_whitespace_as_suffix = false;
+
+    std::vector<char> precompiled_charsmap;
+
+    llm_tokenizer * tokenizer = nullptr;
+
+    llama_vocab() = default;
+    ~llama_vocab();
+
+    int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+
+    void init_tokenizer();
+};
+
+//
+// internal API
+//
+
+// TODO: rename to llama_tokenize_impl
+// TODO: This should probably be in llama.h
+std::vector<llama_vocab::id> llama_tokenize_internal(
+        const llama_vocab & vocab,
+        std::string raw_text,
+        bool add_special,
+        bool parse_special = false);
+
+// TODO: move the API below as member functions of llama_vocab
+llama_token llama_byte_to_token_impl(const llama_vocab & vocab, uint8_t ch);
+
+const char * llama_token_get_text_impl(const struct llama_vocab & vocab, llama_token token);
+
+float llama_token_get_score_impl(const struct llama_vocab & vocab, llama_token token);
+
+llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, llama_token token);
+
+bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token);
+
+bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token token);
+
+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);
+
+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_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,
+                      const char * text,
+                         int32_t   text_len,
+                     llama_token * tokens,
+                         int32_t   n_tokens_max,
+                            bool   add_special,
+                            bool   parse_special);
+
+// does not write null-terminator to buf
+int32_t llama_token_to_piece_impl(
+        const struct llama_vocab & vocab,
+                     llama_token   token,
+                            char * buf,
+                         int32_t   length,
+                         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,
+                         int32_t   n_tokens,
+                            char * text,
+                         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/talk-llama.cpp

@@ -16,7 +16,7 @@
 #include <regex>
 #include <sstream>
 
-std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+static std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
     auto * model = llama_get_model(ctx);
 
     // upper limit for the number of tokens
@@ -33,12 +33,12 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
     return result;
 }
 
-std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), 0, false);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), 0, false);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -59,13 +59,13 @@ struct whisper_params {
     float vad_thold  = 0.6f;
     float freq_thold = 100.0f;
 
-    bool speed_up       = false;
     bool translate      = false;
     bool print_special  = false;
     bool print_energy   = false;
     bool no_timestamps  = true;
     bool verbose_prompt = false;
     bool use_gpu        = true;
+    bool flash_attn     = false;
 
     std::string person      = "Georgi";
     std::string bot_name    = "LLaMA";
@@ -83,7 +83,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -99,12 +99,12 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ngl" || arg == "--n-gpu-layers")   { params.n_gpu_layers   = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")      { params.vad_thold      = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")     { params.freq_thold     = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")       { params.speed_up       = true; }
         else if (arg == "-tr"  || arg == "--translate")      { params.translate      = true; }
         else if (arg == "-ps"  || arg == "--print-special")  { params.print_special  = true; }
         else if (arg == "-pe"  || arg == "--print-energy")   { params.print_energy   = true; }
         else if (arg == "-vp"  || arg == "--verbose-prompt") { params.verbose_prompt = true; }
         else if (arg == "-ng"  || arg == "--no-gpu")         { params.use_gpu        = false; }
+        else if (arg == "-fa"  || arg == "--flash-attn")     { params.flash_attn     = true; }
         else if (arg == "-p"   || arg == "--person")         { params.person         = argv[++i]; }
         else if (arg == "-bn"   || arg == "--bot-name")      { params.bot_name       = argv[++i]; }
         else if (arg == "--session")                         { params.path_session   = argv[++i]; }
@@ -123,7 +123,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             }
         }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
-        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -148,12 +147,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ngl N,   --n-gpu-layers N [%-7d] number of layers to store in VRAM\n",           params.n_gpu_layers);
     fprintf(stderr, "  -vth N,   --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,      --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,      --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
     fprintf(stderr, "  -vp,      --verbose-prompt [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
     fprintf(stderr, "  -ng,      --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
     fprintf(stderr, "  -p NAME,  --person NAME    [%-7s] person name (for prompt selection)\n",          params.person.c_str());
     fprintf(stderr, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
     fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
@@ -169,7 +168,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(
+static std::string transcribe(
         whisper_context * ctx,
         const whisper_params & params,
         const std::vector<float> & pcmf32,
@@ -203,7 +202,6 @@ std::string transcribe(
     wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
@@ -237,7 +235,7 @@ std::string transcribe(
     return result;
 }
 
-std::vector<std::string> get_words(const std::string &txt) {
+static std::vector<std::string> get_words(const std::string &txt) {
     std::vector<std::string> words;
 
     std::istringstream iss(txt);
@@ -285,9 +283,15 @@ int main(int argc, char ** argv) {
     // whisper init
 
     struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
 
     struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
+    if (!ctx_wsp) {
+        fprintf(stderr, "No whisper.cpp model specified. Please provide using -mw <modelfile>\n");
+        return 1;
+    }
 
     // llama init
 
@@ -301,13 +305,17 @@ int main(int argc, char ** argv) {
     }
 
     struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
+    if (!model_llama) {
+        fprintf(stderr, "No llama.cpp model specified. Please provide using -ml <modelfile>\n");
+        return 1;
+    }
 
     llama_context_params lcparams = llama_context_default_params();
 
     // tune these to your liking
     lcparams.n_ctx      = 2048;
-    lcparams.seed       = 1;
     lcparams.n_threads  = params.n_threads;
+    lcparams.flash_attn = params.flash_attn;
 
     struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
 
@@ -393,6 +401,26 @@ int main(int argc, char ** argv) {
 
     llama_batch batch = llama_batch_init(llama_n_ctx(ctx_llama), 0, 1);
 
+    // init sampler
+    const float top_k = 5;
+    const float top_p = 0.80f;
+    const float temp  = 0.30f;
+
+    const int seed = 0;
+
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    if (temp > 0.0f) {
+        llama_sampler_chain_add(smpl, llama_sampler_init_top_k(top_k));
+        llama_sampler_chain_add(smpl, llama_sampler_init_top_p(top_p, 1));
+        llama_sampler_chain_add(smpl, llama_sampler_init_temp (temp));
+        llama_sampler_chain_add(smpl, llama_sampler_init_dist (seed));
+    } else {
+        llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+    }
+
     // init session
     std::string path_session = params.path_session;
     std::vector<llama_token> session_tokens;
@@ -408,7 +436,7 @@ int main(int argc, char ** argv) {
 
             session_tokens.resize(llama_n_ctx(ctx_llama));
             size_t n_token_count_out = 0;
-            if (!llama_load_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
+            if (!llama_state_load_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
                 return 1;
             }
@@ -691,54 +719,13 @@ int main(int argc, char ** argv) {
 
                     {
                         // out of user input, sample next token
-                        const float top_k          = 5;
-                        const float top_p          = 0.80f;
-                        const float temp           = 0.30f;
-                        const float repeat_penalty = 1.1764f;
-
-                        const int repeat_last_n    = 256;
 
                         if (!path_session.empty() && need_to_save_session) {
                             need_to_save_session = false;
-                            llama_save_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.size());
+                            llama_state_save_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.size());
                         }
 
-                        llama_token id = 0;
-
-                        {
-                            auto logits = llama_get_logits(ctx_llama);
-                            auto n_vocab = llama_n_vocab(model_llama);
-
-                            logits[llama_token_eos(model_llama)] = 0;
-
-                            std::vector<llama_token_data> candidates;
-                            candidates.reserve(n_vocab);
-                            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-                            }
-
-                            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-                            // apply repeat penalty
-                            const float nl_logit = logits[llama_token_nl(model_llama)];
-
-                            llama_sample_repetition_penalties(ctx_llama, &candidates_p,
-                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, repeat_penalty, 0.0, 0.0f);
-
-                            logits[llama_token_nl(model_llama)] = nl_logit;
-
-                            if (temp <= 0) {
-                                // Greedy sampling
-                                id = llama_sample_token_greedy(ctx_llama, &candidates_p);
-                            } else {
-                                // Temperature sampling
-                                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
-                                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                                llama_sample_temp (ctx_llama, &candidates_p, temp);
-                                id = llama_sample_token(ctx_llama, &candidates_p);
-                            }
-                        }
+                        const llama_token id = llama_sampler_sample(smpl, ctx_llama, -1);
 
                         if (id != llama_token_eos(model_llama)) {
                             // add it to the context
@@ -788,8 +775,14 @@ int main(int argc, char ** argv) {
     whisper_print_timings(ctx_wsp);
     whisper_free(ctx_wsp);
 
-    llama_print_timings(ctx_llama);
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx_llama);
+
+    llama_sampler_free(smpl);
+    llama_batch_free(batch);
     llama_free(ctx_llama);
 
+    llama_backend_free();
+
     return 0;
 }