models : update readme

.devops/main-cuda.Dockerfile

@@ -1,40 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.3.1
-# Target the CUDA build image
-ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-# Target the CUDA runtime image
-ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
-WORKDIR /app
-
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
-# Set nvcc architecture
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV WHISPER_CUBLAS=1
-
-RUN apt-get update && \
-    apt-get install -y build-essential \
-    && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-# Ref: https://stackoverflow.com/a/53464012
-ENV CUDA_MAIN_VERSION=12.3
-ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
-
-COPY .. .
-RUN make
-
-FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
-ENV CUDA_MAIN_VERSION=12.3
-ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY --from=build /app /app
-ENTRYPOINT [ "bash", "-c" ]

.devops/main.Dockerfile

@@ -1,19 +0,0 @@
-FROM ubuntu:22.04 AS build
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y build-essential \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY .. .
-RUN make
-
-FROM ubuntu:22.04 AS runtime
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY --from=build /app /app
-ENTRYPOINT [ "bash", "-c" ]

.github/workflows/build.yml

@@ -15,17 +15,16 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2
 
       - name: Build ${{ matrix.arch }}
         run: |
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
             apt update
             apt install -y build-essential libsdl2-dev
             make
@@ -36,7 +35,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Dependencies
         run: |
@@ -53,10 +52,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Build
-        uses: cross-platform-actions/action@v0.24.0
+        uses: cross-platform-actions/action@v0.15.0
         with:
           operating_system: freebsd
           version: '13.2'
@@ -77,17 +76,16 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2
 
       - name: Build ${{ matrix.arch }}
         run: |
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
             apt update
             apt install -y build-essential cmake libsdl2-dev
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
@@ -105,19 +103,18 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2
 
       - name: Build ${{ matrix.arch }}
         run: |
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
             apt update
-            apt install -y clang build-essential cmake libsdl2-dev
+            apt install -y build-essential cmake libsdl2-dev
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
             make
             ctest -L gh --output-on-failure'
@@ -133,181 +130,22 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2
 
       - name: Build ${{ matrix.arch }}
         run: |
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
             apt update
             apt install -y build-essential cmake
             cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
             make
             ctest -L gh --output-on-failure'
 
-  ubuntu-22-cmake-sycl:
-    runs-on: ubuntu-22.04
-
-    strategy:
-      fail-fast: false
-      matrix:
-        dwhisper_sycl: [ON]
-        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          mkdir build
-          cd build
-          cmake -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
-          cmake --build . --config Release -j $(nproc)
-
-  ubuntu-22-cmake-sycl-fp16:
-    runs-on: ubuntu-22.04
-
-    strategy:
-      fail-fast: false
-      matrix:
-        dwhisper_sycl: [ON]
-        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          mkdir build
-          cd build
-          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
-          cmake --build . --config Release -j $(nproc)
-
-  windows-msys2:
-    runs-on: windows-latest
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
-          - { sys: CLANG64, env: clang-x86_64, build: Release }
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Setup ${{ matrix.sys }}
-        uses: msys2/setup-msys2@v2
-        with:
-          update: true
-          msystem: ${{matrix.sys}}
-          install: >-
-            base-devel
-            mingw-w64-${{matrix.env}}-toolchain
-            mingw-w64-${{matrix.env}}-cmake
-            mingw-w64-${{matrix.env}}-SDL2
-            mingw-w64-${{matrix.env}}-openblas
-
-      - name: Build using make
-        shell: msys2 {0}
-        run: |
-            make -j $(nproc)
-
-      - name: Clean after building using make
-        shell: msys2 {0}
-        run: |
-            make clean
-
-      - name: Build using make w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            make WHISPER_OPENBLAS=1 -j $(nproc)
-
-      - name: Build using CMake
-        shell: msys2 {0}
-        run: |
-            cmake -B build
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
-      - name: Clean after building using CMake
-        shell: msys2 {0}
-        run: |
-            rm -rf build
-
-      - name: Build using CMake w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            cmake -B build -DWHISPER_OPENBLAS=ON
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
   windows:
     runs-on: windows-latest
 
@@ -324,14 +162,14 @@ jobs:
             s2arc: x64
             jnaPath: win32-x86-64
           - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -356,14 +194,14 @@ jobs:
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
       - name: Upload dll
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
         with:
           name: ${{ matrix.jnaPath }}_whisper.dll
           path: build/bin/${{ matrix.build }}/whisper.dll
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
         with:
           name: whisper-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
@@ -379,23 +217,20 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-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
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-x64.zip
             s2arc: x64
-            clblast: ON
-            clver: 1.6.1
           - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1
 
       - name: Fetch OpenBLAS
         if: matrix.blas == 'ON'
@@ -413,18 +248,6 @@ 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 ${{ matrix.arch }}
@@ -432,7 +255,6 @@ jobs:
           -DWHISPER_OPENBLAS=${{ matrix.blas }}
           -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
-          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
@@ -447,19 +269,15 @@ jobs:
         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@v4
+        uses: actions/upload-artifact@v1
         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-2019
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -467,25 +285,22 @@ jobs:
         arch: [x64]
         cublas: [ON]
         sdl2: [ON]
-        cuda-toolkit: [12.2.0, 11.8.0]
         include:
           - arch: x64
             s2arc: x64
           - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1
 
       - name: Install CUDA Toolkit
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
-        with:
-          cuda: '${{ matrix.cuda-toolkit }}'
+        uses: Jimver/cuda-toolkit@v0.2.10
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -498,20 +313,12 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUDA=${{ matrix.cublas }}
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CUBLAS=1
 
-      - name: Build ${{ matrix.cuda-toolkit }}
+      - name: Build
         run: |
           cd ./build
-          cmake --build . --config ${{ matrix.build }}
-
-      - name: Copy CUDA DLLs
-        run: >
-          Copy-Item -PassThru
-          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
-          -Include cudart64_*,cublas64_*,cublasLt64_*
-          -Destination build/bin/${{ matrix.build }}
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
@@ -519,9 +326,9 @@ jobs:
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
         with:
-          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
+          name: whisper-cublas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   emscripten:
@@ -533,10 +340,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Setup emsdk
-        uses: mymindstorm/setup-emsdk@v14
+        uses: mymindstorm/setup-emsdk@v12
 
       - name: Verify
         run: emcc -v
@@ -555,7 +362,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Configure
         run: |
@@ -573,75 +380,35 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
-        with:
-          path: whisper
-
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          repository: ggerganov/ggml
-          path: ggml
+        uses: actions/checkout@v3
 
       - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v3
         with:
           distribution: zulu
-          java-version: 21
+          java-version: 17
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
+        uses: android-actions/setup-android@v2
 
       - name: Build
         run: |
-          cd whisper/examples/whisper.android
+          cd examples/whisper.android
           ./gradlew assembleRelease --no-daemon
 
-      - name: Build with external ggml
-        run: |
-          export PATH_TO_GGML=$PWD/ggml
-          cd whisper/examples/whisper.android
-          ./gradlew assembleRelease --no-daemon -PGGML_HOME=$PATH_TO_GGML
-
-  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
-
   java:
     needs: [ 'windows' ]
     runs-on: windows-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3
 
       - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v1
         with:
-          distribution: zulu
-          java-version: 20
+          java-version: 17
 
       - name: Download Windows lib
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v3
         with:
           name: win32-x86-64_whisper.dll
           path: bindings/java/build/generated/resources/main/win32-x86-64
@@ -654,7 +421,7 @@ jobs:
           ./gradlew build
 
       - name: Upload jar
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
         with:
           name: whispercpp.jar
           path: bindings/java/build/libs/whispercpp-*.jar
@@ -676,7 +443,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Test quantize
         run: |

.github/workflows/docker.yml

@@ -1,57 +0,0 @@
-name: Publish Docker image
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-jobs:
-  push_to_registry:
-    name: Push Docker image to Docker Hub
-    if: github.event.pull_request.draft == false
-
-    runs-on: ubuntu-latest
-    env:
-      COMMIT_SHA: ${{ github.sha }}
-    strategy:
-      matrix:
-        config:
-          - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64,linux/arm64" }
-          - { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
-
-    steps:
-      - name: Check out the repo
-        uses: actions/checkout@v3
-
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Log in to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.repository_owner }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Build and push Docker image (versioned)
-        if: github.event_name == 'push'
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
-          file: ${{ matrix.config.dockerfile }}
-
-      - name: Build and push Docker image (tagged)
-        uses: docker/build-push-action@v4
-        with:
-          context: .
-          push: ${{ github.event_name == 'push' }}
-          platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ 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 addon.node -B Release
+        run: npx cmake-js compile -T whisper-addon -B Release
 
       - name: Download test model
         run: |

.gitignore

@@ -6,11 +6,8 @@
 .vs/
 .vscode/
 .DS_Store
-.vimspector.json
-/CMakeSettings.json
 
 build/
-build-coreml/
 build-em/
 build-debug/
 build-release/
@@ -21,11 +18,6 @@ build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
-# SPM
-.build/
-.swiftpm
-*.metallib
-
 /main
 /stream
 /command
@@ -33,7 +25,6 @@ build-sanitize-thread/
 /talk-llama
 /bench
 /quantize
-/server
 /lsp
 
 arm_neon.h
@@ -57,7 +48,3 @@ bindings/java/.idea/
 .idea/
 
 benchmark_results.csv
-cmake-build-debug/
-.cxx/
-.gradle/
-local.properties

AUTHORS

@@ -1,301 +0,0 @@
-# date: Tue Apr  9 20:27:03 EEST 2024
-# this file is auto-generated by scripts/gen-authors.sh
-
-0/0 <zero@imaskeleton.me>
-0cc4m <picard12@live.de>
-0xsourcecode <134374803+0xsourcecode@users.noreply.github.com>
-AT <manyoso@users.noreply.github.com>
-Aarni Koskela <akx@iki.fi>
-Aaron Pham <29749331+aarnphm@users.noreply.github.com>
-Aaron Taylor <aaron@exphat.com>
-Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
-Abitofevrything <54505189+abitofevrything@users.noreply.github.com>
-AfryMask <AfryMask@163.com>
-Ahmad Bilal <ahmad.bilal@empglabs.com>
-AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
-Akash Mahajan <akash7190@gmail.com>
-Akash Mahajan <akashmjn@stanford.edu>
-Al Hoang <3811822-hoanga@users.noreply.gitlab.com>
-Alan <unknown>
-Aleksander Andrzejewski <18704749+aleksanderandrzejewski@users.noreply.github.com>
-Alex Azarov <alex@azarov.by>
-Alex Bacart <13940752+alex-bacart@users.noreply.github.com>
-Alex Evgrashin <aevgrashin@yandex.ru>
-Alexandr Graschenkov <alexandr.graschenkov91@gmail.com>
-Alexandru Mariuti <alex@mariuti.com>
-Alexey Kharlamov <alexey@kharlamov.biz>
-Alfredo Montesinos <alfredo.montesinos@g.austincc.edu>
-Ali Alameh <ali.alameh@isae.edu.lb>
-Ananta Bastola <anantarajbastola@gmail.com>
-Andreu Huguet <andreuhuguet@gmail.com>
-Andrew Huynh <a5thuynh@gmail.com>
-Andrew S <andrews54757@gmail.com>
-Andy Maloney <asmaloney@gmail.com>
-Anton Kostin <masguit42@users.noreply.github.com>
-Artyom Mezin <psycho.fading@gmail.com>
-Asad Memon <asad.lionpk@gmail.com>
-Ashraful Islam <ashraful.meche@gmail.com>
-AsukaMinato <asukaminato@nyan.eu.org>
-AustinMroz <austinmroz@utexas.edu>
-Avik Sengupta <avik@sengupta.net>
-Bader-eddine Ouaich <49657842+baderouaich@users.noreply.github.com>
-Baffin Lee <baffinlee@gmail.com>
-Ben Nortier <bjnortier@gmail.com>
-Benjamin Heiniger <benjamin.heiniger@bluewin.ch>
-Bo-Yi Wu <appleboy.tw@gmail.com>
-Boris Bliznioukov <blib@mail.com>
-Borislav Stanimirov <b.stanimirov@abv.bg>
-Brad Murray <59848399+bradmurray-dt@users.noreply.github.com>
-Brian Murray <brian@bmurray.ca>
-CRD716 <crd716@gmail.com>
-Canis Lupus <Canis-UK@users.noreply.github.com>
-Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
-ChangSeok Oh <shivamidow@users.noreply.github.com>
-Chaoqun <27287694+OpenWaygate@users.noreply.github.com>
-Chia-Hsiang Cheng <88014292+garychia@users.noreply.github.com>
-Chidi Williams <williamschidi1@gmail.com>
-Christian <12550267+iceychris@users.noreply.github.com>
-Clifford Heath <clifford.heath@gmail.com>
-Colin <github@whoisc.cc>
-DGdev91 <DGdev91@users.noreply.github.com>
-Damian Czaja <trojan295@protonmail.com>
-Daniel Bevenius <daniel.bevenius@gmail.com>
-David <dnhkng@gmail.com>
-David Thorpe <djt@mutablelogic.com>
-Davidson Francis <davidsondfgl@gmail.com>
-Dener Stassun <denerstassun@gmail.com>
-Didzis Gosko <didzis@users.noreply.github.com>
-Digipom <admin@digipom.com>
-Dimo <dimo@ieee.org>
-Dody Suria Wijaya <dodysw@gmail.com>
-Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
-Duncan McConnell <ddmcconnell4@gmail.com>
-Egor Egorov <me@egorfine.com>
-Elkana Bardugo <ttv200@gmail.com>
-Emmanuel Schmidbauer <eschmidbauer@gmail.com>
-Engininja2 <139037756+Engininja2@users.noreply.github.com>
-Eric Swanson <eswanson@alloscomp.com>
-Eric Tendian <erictendian@gmail.com>
-Erik Scholz <Green-Sky@users.noreply.github.com>
-Evan Jones <evan.q.jones@gmail.com>
-Evan Martin <evan.martin@gmail.com>
-Eve <139727413+netrunnereve@users.noreply.github.com>
-Evgeny Kuznetsov <evgeny@kuznetsov.md>
-F1L1P <78918286+F1L1Pv2@users.noreply.github.com>
-Fangjun Kuang <csukuangfj@gmail.com>
-Felix <stenbackfelix@gmail.com>
-Finn Voorhees <finnvoorhees@gmail.com>
-FlippFuzz <41221030+FlippFuzz@users.noreply.github.com>
-Gang Chen <goncha@gmail.com>
-Gavin Cai <gavin1818@hotmail.com>
-George Hindle <george@georgehindle.com>
-Georgi Gerganov <ggerganov@gmail.com>
-GitAritron <103900385+GitAritron@users.noreply.github.com>
-GiviMAD <GiviMAD@users.noreply.github.com>
-Gleicon Moraes <gleicon@gmail.com>
-Gregor Jasny <gjasny@googlemail.com>
-Guillaume Wenzek <gwenzek@users.noreply.github.com>
-HY. Kelvin Lee <34256578+hykelvinlee42@users.noreply.github.com>
-Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
-Hang <bebound@gmail.com>
-Herman Semenov <GermanAizek@yandex.ru>
-Hrishikesh Barman <geekodour@users.noreply.github.com>
-Ian Bicking <ian@ianbicking.org>
-Ian Bull <irbull@eclipsesource.com>
-Ikko Ashimine <eltociear@gmail.com>
-InconsolableCellist <23345188+InconsolableCellist@users.noreply.github.com>
-Ismatulla Mansurov <47342870+sapoepsilon@users.noreply.github.com>
-Ivan Gorin <ivangorin21@gmail.com>
-JJ <103335846+computerscienceiscool@users.noreply.github.com>
-Jack Mousseau <jmousseau@users.noreply.github.com>
-JacobLinCool <jacoblincool@gmail.com>
-Jakub Ráček <blizzcz@gmail.com>
-Jared Van Bortel <jared@nomic.ai>
-Jay Binks <jaybinks@gmail.com>
-Jhen-Jie Hong <developer@jhen.me>
-Jhen-Jie Hong <iainst0409@gmail.com>
-JidongZhang-THU <1119708529@qq.com>
-Jo Liss <joliss42@gmail.com>
-Johan <jr.raffin@gmail.com>
-Johannes Gäßler <johannesg@5d6.de>
-John Balis <phobossystems@gmail.com>
-Jonathan Soo <jcsoo@agora.com>
-Jonno <1160532+razodactyl@users.noreply.github.com>
-Joonas Pihlajamaa <joonas.pihlajamaa@iki.fi>
-Jose <34888496+Jerry-Master@users.noreply.github.com>
-Josh Bleecher Snyder <josharian@gmail.com>
-Judd <foldl@users.noreply.github.com>
-Jumper775 <78500318+jumpers775@users.noreply.github.com>
-Justine Tunney <jtunney@gmail.com>
-KP Kaiser <kirk@zothcorp.com>
-Kamilake <exjang0@gmail.com>
-Kartik Saranathan <278928+Kartiku@users.noreply.github.com>
-Kasumi <90275229+kasumi-1@users.noreply.github.com>
-Kawrakow <48489457+ikawrakow@users.noreply.github.com>
-Kevin Brothaler <admin@digipom.com>
-Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
-Kreijstal <rainb@tfwno.gf>
-Kylin <56434533+KyL0N@users.noreply.github.com>
-LBlue <153975653+lbluep@users.noreply.github.com>
-Larry Battle <larry.battle.tech@gmail.com>
-Laytan Laats <laytanlaats@hotmail.com>
-Leo Moll <leo.moll@yeasoft.com>
-Lexevolution <31176843+Lexevolution@users.noreply.github.com>
-LittleLoli <26589867+WhichWho@users.noreply.github.com>
-Lucas Zanek <57494138+LucasZNK@users.noreply.github.com>
-Luis Herrera <herrera-luis@users.noreply.github.com>
-Lukas Rist <glaslos@gmail.com>
-M. A. Ali <73258591+MightyStud@users.noreply.github.com>
-M. Eren Akbiyik <erenakbiyik@gmail.com>
-Maciek <maciek.mab122@gmail.com>
-Marcin Mielniczuk <marmistrz.dev@zoho.eu>
-Martin Warnaar <martinwarnaar@gmail.com>
-Matheus de Sousa <23645013+keyehzy@users.noreply.github.com>
-Mathijs de Bruin <mathijs@mathijsfietst.nl>
-Matija Pevec <mightymatth@users.noreply.github.com>
-Maximiliano Levi <8160966+maxilevi@users.noreply.github.com>
-Meng, Hengyu <hengyu.meng@intel.com>
-Michael Podvitskiy <podvitskiymichael@gmail.com>
-Michael Rienstra <mrienstra@gmail.com>
-Mikhail Grigorev <sleuthhound@gmail.com>
-Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
-Mohit Agarwal <mohit@sdf.org>
-Murilo Santana <mvrilo@gmail.com>
-Neil Chudleigh <nchudleigh@users.noreply.github.com>
-Neo Zhang Jianyu <jianyu.zhang@intel.com>
-Neuman Vong <neuman.vong@gmail.com>
-Nicholas Albion <nalbion@yahoo.com>
-Niels Mayer <Niels.Mayer@gmail.com>
-Okabintaro <103938900+Okabintaro@users.noreply.github.com>
-Oleg Sidorov <me@whitebox.io>
-Oleg Sidorov <oleg@sidorov.nl>
-Ondrej Kokes <ondrej.kokes@gmail.com>
-Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
-Paul Tsochantaris <ptsochantaris@icloud.com>
-Philipp Zabel <philipp.zabel@gmail.com>
-Philippe Normand <phil@base-art.net>
-Przemysław Pawełczyk <przemoc@gmail.com>
-Qianhe Chen <54462604+chenqianhe@users.noreply.github.com>
-Radosław Gryta <radek.gryta@gmail.com>
-Reinforce-II <fate@eastal.com>
-Reinis Muiznieks <muiznieks.reinis@gmail.com>
-RelatedTitle <r3latedtitle@gmail.com>
-RhinoDevel <RhinoDevel@users.noreply.github.com>
-Rich Jones <miserlou@gmail.com>
-Robin <robin.xw@hotmail.com>
-Roddur Dasgupta <roddurd@gmail.com>
-Roland Rabien <figbug@gmail.com>
-Rotem Dan <rotemdan@gmail.com>
-Ryan Hitchman <hitchmanr@gmail.com>
-Ryan Metcalfe <107415876+RyanMetcalfeInt8@users.noreply.github.com>
-RyanChang <ftes90015@gmail.com>
-Sam <49637763+Onlyartist9@users.noreply.github.com>
-Sam Pullara <spullara@gmail.com>
-Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
-Sergio López <slp@sinrega.org>
-Siddharth Ramakrishnan <srr2141@columbia.edu>
-Simon Moisselin <simon.moisstoll@gmail.com>
-Sindre Sorhus <sindresorhus@gmail.com>
-Slava Primenko <primenko.s@gmail.com>
-Syahmi Azhar <prsyahmi@gmail.com>
-Syed Jafri <syedjafri97@gmail.com>
-Sơn Phan Trung <phantrungson17@gmail.com>
-Taisei Mima <bhbstar.me@gmail.com>
-Takeshi Inoue <inoue.takeshi@gmail.com>
-Tamotsu Takahashi <ttakah+github@gmail.com>
-Taras Glek <taras@thegp.com>
-Tauseef Mohiuddin <35351464+tauseefmohammed2@users.noreply.github.com>
-Thijs Raymakers <thijs@raymakers.nl>
-Thomas Fitzsimmons <fitzsim@fitzsim.org>
-Tiago Fassoni <tiagofassoni@users.noreply.github.com>
-Tienshiao Ma <tienshiao@tienshiao.org>
-Timothy Cronin <40186632+4imothy@users.noreply.github.com>
-Tobrun <tobrun.van.nuland@gmail.com>
-Todd <taf2@users.noreply.github.com>
-Tong Li <31761981+litongjava@users.noreply.github.com>
-Topping1 <78745143+Topping1@users.noreply.github.com>
-Travis Cline <travis.cline@gmail.com>
-UEXTM.com <84163508+uextm@users.noreply.github.com>
-Vadim Peretokin <vperetokin@hey.com>
-Valentin Gosu <1454649+valenting@users.noreply.github.com>
-Vulcan <93451215+trholding@users.noreply.github.com>
-WhiteOlivierus <36532695+WhiteOlivierus@users.noreply.github.com>
-Xiang (Kevin) Li <kevinli020508@gmail.com>
-Xiao-Yong Jin <jinxiaoyong@gmail.com>
-XiaotaoChen <chenxiaotao1234@gmail.com>
-Yajing Tang <phillis@google.com>
-Yang Shen <aplshenyang@gmail.com>
-Yunès <jean.baptiste.yunes@free.fr>
-ZaBlazzingZephyrus <119159668+blazingzephyr@users.noreply.github.com>
-Zigfrid Zvezdin <ziggerZZ@gmail.com>
-Zollner <24618122+Zolliner@users.noreply.github.com>
-ai-at-home <149282006+ai-at-home@users.noreply.github.com>
-alonfaraj <alonfaraj@gmail.com>
-andypayne <apayne@gmail.com>
-ardfork <134447697+ardfork@users.noreply.github.com>
-automaticcat <daogiatuank54@gmail.com>
-be-next <jerome.ramette@gmail.com>
-bert hubert <bert@hubertnet.nl>
-bmwl <brian.marshall@tolko.com>
-bobqianic <129547291+bobqianic@users.noreply.github.com>
-bocytko <bocytko+github@gmail.com>
-boolemancer <48014766+boolemancer@users.noreply.github.com>
-boolemancer <boolemancer@gmail.com>
-bradmit <151883577+bradmit@users.noreply.github.com>
-brunofaustino <b.fa.amorim@gmail.com>
-bssrdf <merlintiger@hotmail.com>
-byte-6174 <88070277+byte-6174@users.noreply.github.com>
-cdosoftei <ciprian.dosoftei@gmail.com>
-clach04 <Chris.Clark@actian.com>
-compilade <113953597+compilade@users.noreply.github.com>
-conradg <conradjgodfrey@gmail.com>
-ddpasa <112642920+ddpasa@users.noreply.github.com>
-denersc <denerstassun@gmail.com>
-dscripka <dscripka@users.noreply.github.com>
-duthils <duthils@duthils.net>
-ecneladis <ecneladis@users.noreply.github.com>
-faker <nspyia2002@gmail.com>
-fitzsim <fitzsim@fitzsim.org>
-fraxy-v <65565042+fraxy-v@users.noreply.github.com>
-genevera (she/her) <genevera@users.noreply.github.com>
-geniusnut <geniusnut@gmail.com>
-greeshmay <greeshmay@gmail.com>
-hydai <z54981220@gmail.com>
-iamthad <thadeus.j.fleming@gmail.com>
-james wolf <contractorwolf@hotmail.com>
-joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
-jorismertz <35079666+jorismertz@users.noreply.github.com>
-junkfood <69683722+JunkFood02@users.noreply.github.com>
-jwijffels <jwijffels@bnosac.be>
-kamranjon <kamranjon@gmail.com>
-katsu560 <katsu560oo-@docomo.ne.jp>
-kennethge <57784063+kenneth-ge@users.noreply.github.com>
-keyehzy <msamuel@aluno.puc-rio.br>
-leejet <leejet714@gmail.com>
-litong <31761981+litongjava@users.noreply.github.com>
-lnyan <lkwq007@gmail.com>
-m.bell <m.bell@techsmith.com>
-mkiol <mkiol@users.noreply.github.com>
-novag <7754358+novag@users.noreply.github.com>
-pajowu <pajowu@pajowu.de>
-polarmoon <90010972+polarmoon@users.noreply.github.com>
-rlapray <lapray.romain@gmail.com>
-sandrohanea <40202887+sandrohanea@users.noreply.github.com>
-semiformal-net <84111142+semiformal-net@users.noreply.github.com>
-shibukazu <61775791+shibukazu@users.noreply.github.com>
-shikokuchuo <53399081+shikokuchuo@users.noreply.github.com>
-slaren <slarengh@gmail.com>
-slashlib <slashlib@users.noreply.github.com>
-snadampal <87143774+snadampal@users.noreply.github.com>
-st-gr <38470677+st-gr@users.noreply.github.com>
-texmex76 <40733439+texmex76@users.noreply.github.com>
-thefinaldegree <thefinaldegree@gmail.com>
-trixirt <trix@redhat.com>
-ulatekh <ulatekh@yahoo.com>
-undef <undefdev@gmail.com>
-venkr <venkateshrameshkumar+1@gmail.com>
-vicalloy <zbirder@gmail.com>
-xdrudis <xavierdrudis@yahoo.es>
-zhouwg <6889919+zhouwg@users.noreply.github.com>
-布客飞龙 <562826179@qq.com>
-Артём Земляк <azemlyak@smart-consulting.ru>

CMakeLists.txt

@@ -1,10 +1,6 @@
 cmake_minimum_required (VERSION 3.5)
 
-# Allow for the creation of solution folders.
-set_property(GLOBAL PROPERTY USE_FOLDERS ON)
-
-project(whisper.cpp VERSION 1.6.2)
-set(SOVERSION 1)
+project(whisper.cpp VERSION 1.4.2)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -59,17 +55,10 @@ option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDA
 
 option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
 
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    option(WHISPER_FFMPEG                 "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
-endif()
-
-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_NO_AVX                 "whisper: disable AVX"  OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)
 
 option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
 
@@ -79,19 +68,13 @@ if (APPLE)
     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)
+    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_CUBLAS                "whisper: support for cuBLAS"  OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS" OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
 endif()
 
 option(WHISPER_PERF "whisper: enable perf timings" OFF)
@@ -122,33 +105,6 @@ endif()
 
 find_package(Threads REQUIRED)
 
-#compile flag sycl
-if (WHISPER_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
-
-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)
-    set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS} ${FFMPEG_LIBRARIES})
-endif()
-
 # on APPLE
 if (APPLE)
     # include Accelerate framework
@@ -159,7 +115,7 @@ if (APPLE)
             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)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
         else()
             message(FATAL_ERROR "Accelerate framework not found")
         endif()
@@ -189,42 +145,8 @@ if (APPLE)
 
         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)
+        # copy ggml-metal.metal to bin directory
         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_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-            set(COMMON_HEADER   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-
-            file(MAKE_DIRECTORY "${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)
@@ -248,82 +170,30 @@ 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})
+    if (WIN32)
+        if(DEFINED ENV{OPENBLAS_PATH})
+            set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories($ENV{OPENBLAS_PATH}/include)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
         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 ${CMAKE_SHARED_LIBRARY_SUFFIX})
-            endif ()
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
         endif ()
-        # OpenBLAS prebuilt libraries hardcode "lib" prefix in filename even on Windows
-        if (WHISPER_OPENBLAS)
-            set(WHISPER_BLAS_LIB_PREFIX "lib")
-        endif ()
-        message(STATUS "BLAS compatible library path provided")
-        set(BLAS_LIBRARIES "$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_STATIC 1)
         set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-        if (WHISPER_OPENBLAS_INTERFACE64)
-            set(BLA_SIZEOF_INTEGER 8)
-        else ()
-            set(BLA_SIZEOF_INTEGER 4)
-        endif()
+        set(BLA_SIZEOF_INTEGER 8)
         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}")
+            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
             include_directories(${BLAS_INCLUDE_DIRS})
             set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
@@ -333,19 +203,7 @@ if (WHISPER_BLAS)
     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)
@@ -355,24 +213,16 @@ if (WHISPER_CUDA)
 
         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)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
-        add_compile_definitions(GGML_USE_CUDA)
+        add_compile_definitions(GGML_USE_CUBLAS)
 
         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 CUDA::cufft)
-            else ()
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static CUDA::cufft_static)
-            endif()
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
         else()
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt CUDA::cufft)
+            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()
@@ -394,18 +244,16 @@ if (WHISPER_HIPBLAS)
 
     if (${hipblas_FOUND} AND ${hip_FOUND})
         message(STATUS "HIP and hipBLAS found")
-        set(GGML_HEADERS_ROCM "ggml-cuda.h")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        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)
 
-        file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
-
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
         if (WHISPER_STATIC)
             message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
         endif()
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+        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()
@@ -430,30 +278,6 @@ 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)
@@ -485,8 +309,7 @@ if (WHISPER_ALL_WARNINGS)
 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} -Werror=vla")
     #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
 endif()
 
@@ -502,35 +325,21 @@ else()
         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)
+        if(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")
+        else()
+            if(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 "${CMAKE_C_FLAGS} /arch:AVX")
+            endif()
         endif()
     else()
         if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
         else()
             if(NOT WHISPER_NO_AVX)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
@@ -538,15 +347,6 @@ else()
             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")
-                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()
-            endif()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
@@ -633,7 +433,6 @@ if (WHISPER_COREML)
     set_target_properties(${TARGET} PROPERTIES
         COMPILE_FLAGS "-fobjc-arc"
         )
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
 if (WHISPER_OPENVINO)
@@ -652,7 +451,6 @@ if (WHISPER_OPENVINO)
     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()
 
 #
@@ -673,25 +471,10 @@ add_library(${TARGET}
     ${GGML_SOURCES_METAL}
     ${GGML_SOURCES_CUDA}
     ${GGML_SOURCES_OPENCL}
-    ${GGML_SOURCES_SYCL}        ${GGML_HEADERS_SYCL}
-    ${GGML_SOURCES_ROCM}        ${GGML_HEADERS_ROCM}
     whisper.h
     whisper.cpp
     )
 
-if (WHISPER_CUDA)
-    target_sources(${TARGET} PRIVATE whisper-mel-cuda.cu)
-endif()
-
-include_directories (
-    .
-)
-# Set the version numbers
-set_target_properties(whisper PROPERTIES
-    VERSION ${PROJECT_VERSION}
-    SOVERSION ${SOVERSION}
-)
-
 include(DefaultTargetOptions)
 
 target_include_directories(${TARGET} PUBLIC
@@ -706,10 +489,6 @@ if (WHISPER_OPENVINO)
     target_link_libraries(${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})
 
@@ -719,7 +498,6 @@ else()
 endif()
 
 if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_link_libraries(${TARGET} PUBLIC
         ${CMAKE_DL_LIBS}
         )
@@ -743,13 +521,7 @@ endif()
 
 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_ARCHITECTURES OFF)
     set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
 endif()
 
@@ -761,8 +533,7 @@ 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")
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
 
 include(GNUInstallDirs)
 

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023-2024 The ggml authors
+Copyright (c) 2023 Georgi Gerganov
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -1,4 +1,4 @@
-default: main bench quantize server
+default: main bench quantize
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -18,17 +18,6 @@ ifndef NVCC_VERSION
 	endif
 endif
 
-# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
-# of non-gcc compilers don't have to provide g++ alias or wrapper.
-DEFCC  := cc
-DEFCXX := c++
-ifeq ($(origin CC),default)
-CC  := $(DEFCC)
-endif
-ifeq ($(origin CXX),default)
-CXX := $(DEFCXX)
-endif
-
 CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)
 
@@ -53,12 +42,6 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
-ifdef MACOSX_DEPLOYMENT_TARGET
-	CFLAGS   += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-	CXXFLAGS += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-	LDFLAGS  += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-endif
-
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
@@ -116,16 +99,6 @@ ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$
 	CXXFLAGS += -pthread
 endif
 
-# detect Windows
-ifneq ($(findstring _NT,$(UNAME_S)),)
-	_WIN32 := 1
-endif
-
-# Windows Sockets 2 (Winsock) for network-capable apps
-ifeq ($(_WIN32),1)
-	LWINSOCK2 := -lws2_32
-endif
-
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -134,7 +107,7 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
 	else ifeq ($(UNAME_S),Linux)
 		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter MINGW32_NT% MINGW64_NT% MSYS_NT%,$(UNAME_S)))
+	else ifneq (,$(filter MINGW32_NT% MINGW64_NT%,$(UNAME_S)))
 		CPUINFO_CMD := cat /proc/cpuinfo
 	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
 		CPUINFO_CMD := grep Features /var/run/dmesg.boot
@@ -142,68 +115,41 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysinfo -cpu
 	endif
 
-	# x86 ISA extensions (chronological order)
 	ifdef CPUINFO_CMD
-		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
-		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
 		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
-		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
-		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
-		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
-		AVX512F_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512F')
-		AVX512VBMI_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512VBMI')
-		AVX512VNNI_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX512_VNNI|AVX512VNNI')
-
-		# AVX-512 has many subsets, so let's make it easy to disable them all
-		ifneq ($(filter-out 0,$(WHISPER_NO_AVX512)),)
-			AVX512F_M :=
-			AVX512VBMI_M :=
-			AVX512VNNI_M :=
-		endif
-
-		ifneq (,$(SSE3_M))
-			CFLAGS   += -msse3
-			CXXFLAGS += -msse3
-		endif
-
-		ifneq (,$(SSSE3_M))
-			CFLAGS   += -mssse3
-			CXXFLAGS += -mssse3
-		endif
-
 		ifneq (,$(AVX_M))
 			CFLAGS   += -mavx
 			CXXFLAGS += -mavx
 		endif
 
-		ifneq (,$(F16C_M))
-			CFLAGS   += -mf16c
-			CXXFLAGS += -mf16c
-		endif
-
-		ifneq (,$(FMA_M))
-			CFLAGS   += -mfma
-			CXXFLAGS += -mfma
-		endif
-
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
 		ifneq (,$(AVX2_M))
 			CFLAGS   += -mavx2
 			CXXFLAGS += -mavx2
 		endif
 
-		ifneq (,$(AVX512F_M))
-			CFLAGS   += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
-			CXXFLAGS += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
 		endif
 
-		ifneq (,$(AVX512VBMI_M))
-			CFLAGS   += -mavx512vbmi
-			CXXFLAGS += -mavx512vbmi
+		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
+		ifneq (,$(F16C_M))
+			CFLAGS   += -mf16c
+			CXXFLAGS += -mf16c
 		endif
 
-		ifneq (,$(AVX512VNNI_M))
-			CFLAGS   += -mavx512vnni
-			CXXFLAGS += -mavx512vnni
+		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
+		ifneq (,$(SSE3_M))
+			CFLAGS   += -msse3
+			CXXFLAGS += -msse3
+		endif
+
+		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
+		ifneq (,$(SSSE3_M))
+			CFLAGS   += -mssse3
+			CXXFLAGS += -mssse3
 		endif
 	endif
 endif
@@ -223,8 +169,6 @@ ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
 		CFLAGS  += -DGGML_USE_ACCELERATE
-		CFLAGS  += -DACCELERATE_NEW_LAPACK
-		CFLAGS  += -DACCELERATE_LAPACK_ILP64
 		LDFLAGS += -framework Accelerate
 	endif
 endif
@@ -248,76 +192,41 @@ ifndef WHISPER_NO_METAL
 	endif
 endif
 
-ifneq ($(filter-out 0,$(WHISPER_OPENBLAS)),) # OpenBLAS
-	WHISPER_OPENBLAS_INTERFACE64 ?= 0 # use 32-bit interface by default
-	ifneq ($(filter-out 0,$(WHISPER_OPENBLAS_INTERFACE64)),)
-		WHISPER_BLAS_LIB := openblas64
-	else
-		WHISPER_BLAS_LIB := openblas
-	endif
-	ifneq ($(OPENBLAS_PATH),)
-		WHISPER_BLAS_CFLAGS  := -I$(OPENBLAS_PATH)/include
-		WHISPER_BLAS_LDFLAGS := -L$(OPENBLAS_PATH)/lib -l$(WHISPER_BLAS_LIB)
-	else
-		WHISPER_BLAS_LIB_PC_EXISTS := $(shell pkg-config --exists $(WHISPER_BLAS_LIB) && echo 1)
-		ifneq ($(filter-out 0,$(WHISPER_BLAS_LIB_PC_EXISTS)),)
-			WHISPER_BLAS_CFLAGS  := $(shell pkg-config --cflags $(WHISPER_BLAS_LIB))
-			WHISPER_BLAS_LDFLAGS := $(shell pkg-config --libs   $(WHISPER_BLAS_LIB))
-		else
-			WHISPER_BLAS_CFLAGS  := -I/usr/include/openblas
-			WHISPER_BLAS_LDFLAGS := -l$(WHISPER_BLAS_LIB)
-		endif
-	endif
-	CFLAGS  += $(WHISPER_BLAS_CFLAGS) -DGGML_USE_OPENBLAS
-	LDFLAGS += $(WHISPER_BLAS_LDFLAGS)
+ifdef WHISPER_OPENBLAS
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
+	LDFLAGS += -lopenblas
 endif
 
 ifdef WHISPER_CUBLAS
-# WHISPER_CUBLAS is deprecated and will be removed in the future
-	WHISPER_CUDA := 1
-endif
-
-ifdef WHISPER_CUDA
 	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
-		CUDA_ARCH_FLAG ?= native
+		CUDA_ARCH_FLAG=native
 	else
-		CUDA_ARCH_FLAG ?= all
+		CUDA_ARCH_FLAG=all
 	endif
 
-	CFLAGS      += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	CXXFLAGS    += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lcufft -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
-	WHISPER_OBJ += ggml-cuda.o whisper-mel-cuda.o
-	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
+	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -c $< -o $@
-
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 
-whisper-mel-cuda.o: whisper-mel-cuda.cu whisper.h ggml.h ggml-backend.h whisper-mel.hpp whisper-mel-cuda.hpp
-	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
-
 ifdef WHISPER_HIPBLAS
 	ROCM_PATH   ?= /opt/rocm
 	HIPCC       ?= $(ROCM_PATH)/bin/hipcc
 	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
-	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
 	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	LDFLAGS     += -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
 	WHISPER_OBJ += ggml-cuda.o
-	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 
-ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 endif
 
@@ -382,13 +291,6 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-ifdef WHISPER_CUBLAS
-$(info !!!!)
-$(info WHISPER_CUBLAS is deprecated and will be removed in the future. Use WHISPER_CUDA instead.)
-$(info !!!!)
-$(info )
-endif
-
 #
 # Build library
 #
@@ -405,9 +307,9 @@ ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-WHISPER_OBJ += ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o
+WHISPER_OBJ += ggml-alloc.o ggml-backend.o ggml-quants.o
 
-whisper.o: whisper.cpp whisper.h whisper-mel.hpp ggml.h ggml-cuda.h
+whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ifndef WHISPER_COREML
@@ -427,36 +329,16 @@ ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 WHISPER_OBJ += ggml-metal.o
-
-ifdef WHISPER_METAL_EMBED_LIBRARY
-CFLAGS += -DGGML_METAL_EMBED_LIBRARY
-
-ggml-metal-embed.o: ggml-metal.metal ggml-common.h
-	@echo "Embedding Metal library"
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	$(eval TEMP_METALLIB=$(shell mktemp))
-	@sed "/^#include \"ggml-common.h\"/{r ggml-common.h"$$'\n'"d;}" ggml-metal.metal > $(TEMP_METALLIB)
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$(TEMP_METALLIB)\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f $(TEMP_ASSEMBLY) $(TEMP_METALLIB)
-
-WHISPER_OBJ += ggml-metal-embed.o
-endif
 endif
 
-libwhisper.a: $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a $(WHISPER_OBJ)
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
 
-libwhisper.so: $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so $(WHISPER_OBJ) $(LDFLAGS)
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk talk-llama bench quantize server lsp libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize lsp libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -464,36 +346,33 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.cpp
+SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-bench: examples/bench/bench.cpp $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp $(WHISPER_OBJ) -o bench $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
-	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
-server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS) $(LWINSOCK2)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
 
-lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
-
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples
@@ -539,9 +418,9 @@ samples:
 .PHONY: medium
 .PHONY: large-v1
 .PHONY: large-v2
-.PHONY: large-v3
+.PHONY: large
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="

Package.swift

@@ -1,61 +0,0 @@
-// swift-tools-version:5.5
-
-import PackageDescription
-
-let package = Package(
-    name: "whisper",
-    platforms: [
-        .macOS(.v12),
-        .iOS(.v14),
-        .watchOS(.v4),
-        .tvOS(.v14)
-    ],
-    products: [
-        .library(name: "whisper", targets: ["whisper"]),
-    ],
-    targets: [
-        .target(
-            name: "whisper",
-            path: ".",
-            exclude: [
-               "bindings",
-               "cmake",
-               "coreml",
-               "examples",
-               "extra",
-               "models",
-               "samples",
-               "tests",
-               "CMakeLists.txt",
-               "ggml-cuda.cu",
-               "ggml-cuda.h",
-               "Makefile"
-            ],
-            sources: [
-                "ggml.c",
-                "whisper.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m"
-            ],
-            resources: [.process("ggml-metal.metal")],
-            publicHeadersPath: "spm-headers",
-            cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL")
-                // NOTE: NEW_LAPACK will required iOS version 16.4+
-                // We should consider add this in the future when we drop support for iOS 14
-                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-                // .define("ACCELERATE_NEW_LAPACK"),
-                // .define("ACCELERATE_LAPACK_ILP64")
-            ],
-            linkerSettings: [
-                .linkedFramework("Accelerate")
-            ]
-        )
-    ],
-    cxxLanguageStandard: .cxx11
-)

README.md

@@ -4,10 +4,9 @@
 
 [![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: [v1.6.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Beta: [v1.4.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.2) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [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:
 
@@ -17,10 +16,12 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - 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)
+- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
 - Support for CPU-only inference
-- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- [Partial GPU support for NVIDIA via cuBLAS](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)
+- [BLAS CPU support via OpenBLAS](https://github.com/ggerganov/whisper.cpp#blas-cpu-support-via-openblas)
 - [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
@@ -34,10 +35,11 @@ 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)
 
-The entire high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
-The rest of the code is part of the [`ggml`](https://github.com/ggerganov/ggml) machine learning library.
+The entire implementation of the model is contained in 2 source files:
+
+- Tensor operations: [ggml.h](ggml.h) / [ggml.c](ggml.c)
+- Transformer inference: [whisper.h](whisper.h) / [whisper.cpp](whisper.cpp)
 
 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
 As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device: [whisper.objc](examples/whisper.objc)
@@ -62,22 +64,22 @@ Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 - 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
 
-The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
+The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD
+intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
+the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
 
 ## Quick start
 
-First clone the repository:
+First clone the repository.
 
-```bash
-git clone https://github.com/ggerganov/whisper.cpp.git
-```
-
-Then, download one of the Whisper [models](models/README.md) converted in [`ggml` format](#ggml-format). For example:
+Then, download one of the Whisper models converted in [ggml format](models). For example:
 
 ```bash
 bash ./models/download-ggml-model.sh base.en
 ```
 
+If you wish to convert the Whisper models to ggml format yourself, instructions are in [models/README.md](models/README.md).
+
 Now build the [main](examples/main) example and transcribe an audio file like this:
 
 ```bash
@@ -92,7 +94,7 @@ make
 
 For a quick demo, simply run `make base.en`:
 
-```text
+```java
 $ make base.en
 
 cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
@@ -112,8 +114,8 @@ options:
   -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
   -ml N,     --max-len N         [0      ] maximum segment length in characters
   -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [5      ] beam size for beam search
+  -bo N,     --best-of N         [2      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
   -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
   -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
@@ -130,7 +132,6 @@ options:
   -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
   -oj,       --output-json       [false  ] output result in a JSON file
-  -ojf,      --output-json-full  [false  ] include more information in the JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
@@ -142,8 +143,7 @@ options:
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
   -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of tokens
-  -ng,       --no-gpu            [false  ] disable GPU
+  -ls,       --log-score         [false  ] log best decoder scores of token
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -208,7 +208,7 @@ For detailed usage instructions, run: `./main -h`
 Note that the [main](examples/main) example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.
 For example, you can use `ffmpeg` like this:
 
-```bash
+```java
 ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
 ```
 
@@ -235,18 +235,18 @@ make medium.en
 make medium
 make large-v1
 make large-v2
-make large-v3
+make large
 ```
 
 ## Memory usage
 
-| Model  | Disk    | Mem     |
-| ------ | ------- | ------- |
-| tiny   | 75 MiB  | ~273 MB |
-| base   | 142 MiB | ~388 MB |
-| small  | 466 MiB | ~852 MB |
-| medium | 1.5 GiB | ~2.1 GB |
-| large  | 2.9 GiB | ~3.9 GB |
+| Model  | Disk   | Mem     | SHA                                        |
+| ---    | ---    | ---     | ---                                        |
+| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
+| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
+| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
+| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
+| large  | 2.9 GB | ~3.3 GB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
 
 ## Quantization
 
@@ -279,8 +279,7 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
   - To ensure `coremltools` operates correctly, please confirm that [Xcode](https://developer.apple.com/xcode/) is installed and execute `xcode-select --install` to install the command-line tools.
   - Python 3.10 is recommended.
-  - MacOS Sonoma (version 14) or newer is recommended, as older versions of MacOS might experience issues with transcription hallucination.
-  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for this step:
+  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html)  for this step:
     - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
     - To activate the environment, use: `conda activate py310-whisper`
 
@@ -306,8 +305,8 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
 - Run the examples as usual. For example:
 
-  ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -335,23 +334,21 @@ This can result in significant speedup in encoder performance. Here are the inst
 - First, setup python virtual env. and install python dependencies. Python 3.10 is recommended.
 
   Windows:
-
-  ```powershell
+  ```
   cd models
   python -m venv openvino_conv_env
   openvino_conv_env\Scripts\activate
   python -m pip install --upgrade pip
-  pip install -r requirements-openvino.txt
+  pip install -r openvino-conversion-requirements.txt
   ```
 
   Linux and macOS:
-
-  ```bash
+  ```
   cd models
   python3 -m venv openvino_conv_env
   source openvino_conv_env/bin/activate
   python -m pip install --upgrade pip
-  pip install -r requirements-openvino.txt
+  pip install -r openvino-conversion-requirements.txt
   ```
 
 - Generate an OpenVINO encoder model. For example, to generate a `base.en` model, use:
@@ -360,7 +357,7 @@ This can result in significant speedup in encoder performance. Here are the inst
   python convert-whisper-to-openvino.py --model base.en
   ```
 
-  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as `ggml` models, as that
+  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as ggml models, as that
   is the default location that the OpenVINO extension will search at runtime.
 
 - Build `whisper.cpp` with OpenVINO support:
@@ -370,28 +367,24 @@ This can result in significant speedup in encoder performance. Here are the inst
   After downloading & extracting package onto your development system, set up required environment by sourcing setupvars script. For example:
 
   Linux:
-
   ```bash
   source /path/to/l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64/setupvars.sh
   ```
 
   Windows (cmd):
-
-  ```powershell
+  ```
   C:\Path\To\w_openvino_toolkit_windows_2023.0.0.10926.b4452d56304_x86_64\setupvars.bat
   ```
 
   And then build the project using cmake:
-
   ```bash
   cmake -B build -DWHISPER_OPENVINO=1
   cmake --build build -j --config Release
   ```
 
 - Run the examples as usual. For example:
-
-  ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -407,19 +400,19 @@ This can result in significant speedup in encoder performance. Here are the inst
 
   The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
   cached for the next run.
-
+  
 For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
 
-## NVIDIA GPU support
+## NVIDIA GPU support via cuBLAS
 
-With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
+With NVIDIA cards the Encoder processing can to a large extent be offloaded to the GPU through cuBLAS.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads
 
-Now build `whisper.cpp` with CUDA support:
+Now build `whisper.cpp` with cuBLAS support:
 
 ```
 make clean
-WHISPER_CUDA=1 make -j
+WHISPER_CUBLAS=1 make -j
 ```
 
 ## OpenCL GPU support via CLBlast
@@ -442,6 +435,7 @@ cmake -B build -DWHISPER_CLBLAST=ON
 cmake --build build -j --config Release
 ```
 
+
 Run all the examples as usual.
 
 ## BLAS CPU support via OpenBLAS
@@ -456,63 +450,6 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```
 
-## BLAS CPU support via Intel MKL
-
-Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
-First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
-
-Now build `whisper.cpp` with Intel MKL BLAS support:
-
-```
-source /opt/intel/oneapi/setvars.sh
-mkdir build
-cd build
-cmake -DWHISPER_MKL=ON ..
-WHISPER_MKL=1 make -j
-```
-
-## Docker
-
-### Prerequisites
-
-- Docker must be installed and running on your system.
-- Create a folder to store big models & intermediate files (ex. /whisper/models)
-
-### Images
-
-We have two Docker images available for this project:
-
-1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
-
-### Usage
-
-```shell
-# download model and persist it in a local folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  whisper.cpp:main "./models/download-ggml-model.sh base /models"
-# transcribe an audio file
-docker run -it --rm \
-  -v path/to/models:/models \
-  -v path/to/audios:/audios \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
-# transcribe an audio file in samples folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  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
@@ -525,7 +462,7 @@ in about half a minute on a MacBook M1 Pro, using `medium.en` model:
 <details>
   <summary>Expand to see the result</summary>
 
-```text
+```java
 $ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8
 
 whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
@@ -597,7 +534,6 @@ whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per
 whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
 whisper_print_timings:    total time = 32733.52 ms
 ```
-
 </details>
 
 ## Real-time audio input example
@@ -606,7 +542,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
-```bash
+```java
 make stream
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
@@ -618,7 +554,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:
 
-```bash
+```java
 ./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
 ```
 
@@ -628,8 +564,8 @@ to highlight words with high or low confidence:
 
 For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:
 
-```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
+```java
+./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -652,8 +588,8 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr
 
 The `--max-len` argument can be used to obtain word-level timestamps. Simply use `-ml 1`:
 
-```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
+```java
+./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -721,9 +657,9 @@ 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
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
 source ./samples/jfk.wav.wts
 ffplay ./samples/jfk.wav.mp4
@@ -733,7 +669,7 @@ https://user-images.githubusercontent.com/1991296/199337465-dbee4b5e-9aeb-48a3-b
 
 ---
 
-```bash
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owts
 source ./samples/mm0.wav.wts
 ffplay ./samples/mm0.wav.mp4
@@ -743,7 +679,7 @@ https://user-images.githubusercontent.com/1991296/199337504-cc8fd233-0cb7-4920-9
 
 ---
 
-```bash
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owts
 source ./samples/gb0.wav.wts
 ffplay ./samples/gb0.wav.mp4
@@ -755,10 +691,10 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ## Video comparison of different models
 
-Use the [scripts/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/scripts/bench-wts.sh) script to generate a video in the following format:
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
 
-```bash
-./scripts/bench-wts.sh samples/jfk.wav
+```java
+./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```
 
@@ -779,14 +715,15 @@ Additionally a script to run whisper.cpp with different models and audio files i
 You can run it with the following command, by default it will run against any standard model in the models folder.
 
 ```bash
-python3 scripts/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
 ```
 
 It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
 
 It outputs a csv file with the results of the benchmarking.
 
-## `ggml` format
+
+## ggml format
 
 The original models are converted to a custom binary format. This allows to pack everything needed into a single file:
 
@@ -801,51 +738,49 @@ or manually from here:
 - https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
-For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or [models/README.md](models/README.md).
+For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
+in [models](models).
 
 ## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
-- [x] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
-- [x] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
+- [X] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
   - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
-- [x] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
-- [x] Java:
+- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [X] Java:
   - [GiviMAD/whisper-jni](https://github.com/GiviMAD/whisper-jni)
-- [x] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
-- [x] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
+- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
   - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
-- [x] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
   - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
-- [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [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)
-- [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
-- [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
 There are various examples of using the library for different projects in the [examples](examples) folder.
 Some of the examples are even ported to run in the browser using WebAssembly. Check them out!
 
-| Example                                             | Web                                   | Description                                                                                                                     |
-| --------------------------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
-| [main](examples/main)                               | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
-| [bench](examples/bench)                             | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
-| [stream](examples/stream)                           | [stream.wasm](examples/stream.wasm)   | Real-time transcription of raw microphone capture                                                                               |
-| [command](examples/command)                         | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic                                                         |
-| [wchess](examples/wchess)                           | [wchess.wasm](examples/wchess)        | Voice-controlled chess                                                                                                          |
-| [talk](examples/talk)                               | [talk.wasm](examples/talk.wasm)       | Talk with a GPT-2 bot                                                                                                           |
-| [talk-llama](examples/talk-llama)                   |                                       | Talk with a LLaMA bot                                                                                                           |
-| [whisper.objc](examples/whisper.objc)               |                                       | iOS mobile application using whisper.cpp                                                                                        |
-| [whisper.swiftui](examples/whisper.swiftui)         |                                       | SwiftUI iOS / macOS application using whisper.cpp                                                                               |
-| [whisper.android](examples/whisper.android)         |                                       | Android mobile application using whisper.cpp                                                                                    |
-| [whisper.nvim](examples/whisper.nvim)               |                                       | Speech-to-text plugin for Neovim                                                                                                |
-| [generate-karaoke.sh](examples/generate-karaoke.sh) |                                       | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture                           |
-| [livestream.sh](examples/livestream.sh)             |                                       | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185)                                           |
-| [yt-wsp.sh](examples/yt-wsp.sh)                     |                                       | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
-| [server](examples/server)                           |                                       | HTTP transcription server with OAI-like API                                                                                     |
+| Example | Web | Description |
+| ---     | --- | ---         |
+| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
+| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
+| [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
+| [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
+| [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
+| [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
+| [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
+| [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
+| [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
+| [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
+| [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
+| [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
 
 ## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)
 

README_sycl.md

@@ -1,249 +0,0 @@
-# whisper.cpp for SYCL
-
-[Background](#background)
-
-[OS](#os)
-
-[Intel GPU](#intel-gpu)
-
-[Linux](#linux)
-
-[Environment Variable](#environment-variable)
-
-[Known Issue](#known-issue)
-
-[Todo](#todo)
-
-## Background
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators<72>such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
-
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
-
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
-
-To avoid  re-inventing the wheel, this code refers other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel<EFBFBD> DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
-
-The whisper.cpp for SYCL is used to support Intel GPUs.
-
-For Intel CPU, recommend to use whisper.cpp for X86 (Intel MKL build).
-
-## OS
-
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04|
-|Windows|Ongoing| |
-
-
-## Intel GPU
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
-
-
-## Linux
-
-### Setup Environment
-
-1. Install Intel GPU driver.
-
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
-
-Note: for iGPU, please install the client GPU driver.
-
-b. Add user to group: video, render.
-
-```
-sudo usermod -aG render username
-sudo usermod -aG video username
-```
-
-Note: re-login to enable it.
-
-c. Check
-
-```
-sudo apt install clinfo
-sudo clinfo -l
-```
-
-Output (example):
-
-```
-Platform #0: Intel(R) OpenCL Graphics
- `-- Device #0: Intel(R) Arc(TM) A770 Graphics
-
-
-Platform #0: Intel(R) OpenCL HD Graphics
- `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
-```
-
-2. Install Intel<65> oneAPI Base toolkit.
-
-
-a. Please follow the procedure in [Get the Intel<65> oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
-
-Recommend to install to default folder: **/opt/intel/oneapi**.
-
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
-
-b. Check
-
-```
-source /opt/intel/oneapi/setvars.sh
-
-sycl-ls
-```
-
-There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
-
-Output (example):
-```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
-
-```
-
-2. Build locally:
-
-```
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON 
-
-#for FP32
-cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v
-
-```
-
-or
-
-```
-./examples/sycl/build.sh
-```
-
-Note:
-
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
-
-### Run
-
-1. Put model file to folder **models**
-
-2. Enable oneAPI running environment
-
-```
-source /opt/intel/oneapi/setvars.sh
-```
-
-3. List device ID
-
-Run without parameter:
-
-```
-./build/bin/ls-sycl-device
-
-or
-
-./build/bin/main
-```
-
-Check the ID in startup log, like:
-
-```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
-```
-
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
-
-4. Set device ID and execute whisper.cpp
-
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
-
-```
-GGML_SYCL_DEVICE=0 ./build/bin/main -m models/ggml-base.en.bin -f samples/jfk.wav
-```
-or run by script:
-
-```
-./examples/sycl/run_whisper.sh
-```
-
-
-
-5. Check the device ID in output
-
-Like:
-```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
-```
-
-
-## Environment Variable
-
-#### Build
-
-|Name|Value|Function|
-|-|-|-|
-|WHISPER_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, WHISPER_SYCL=ON is mandatory.|
-|WHISPER_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path.For FP32, do not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
-
-#### Running
-
-
-|Name|Value|Function|
-|-|-|-|
-|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
-|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-
-## Known Issue
-
-- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
-
-  Miss to enable oneAPI running environment.
-
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
-
-
-- Hang during startup
-
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
-
-  Solution: add **--no-mmap**.
-
-## Todo
-
-- Support to build in Windows.
-
-- Support multiple cards.

bindings/go/Makefile

@@ -1,26 +1,9 @@
-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
-GGML_METAL_PATH_RESOURCES := $(abspath ../..)
 BUILD_DIR := build
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
 INCLUDE_PATH := $(abspath ../..)
 LIBRARY_PATH := $(abspath ../..)
 
-ifeq ($(UNAME_S),Darwin)
-	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit
-endif
-
 all: clean whisper examples
 
 whisper: mkdir
@@ -28,13 +11,8 @@ whisper: mkdir
 	@${MAKE} -C ../.. libwhisper.a
 
 test: model-small whisper modtidy
-ifeq ($(UNAME_S),Darwin)
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v .
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v ./pkg/whisper/...
-else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v .
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v ./pkg/whisper/...
-endif
 
 examples: $(EXAMPLES_DIR)
 
@@ -43,11 +21,7 @@ model-small: mkdir examples/go-model-download
 
 $(EXAMPLES_DIR): mkdir whisper modtidy
 	@echo Build example $(notdir $@)
-ifeq ($(UNAME_S),Darwin)
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go build ${BUILD_FLAGS} -ldflags "-extldflags '$(EXT_LDFLAGS)'" -o ${BUILD_DIR}/$(notdir $@) ./$@
-else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
-endif
 
 mkdir:
 	@echo Mkdir ${BUILD_DIR}

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"}
 )
 
 var (

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

@@ -68,6 +68,10 @@ 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"
 }
@@ -107,6 +111,10 @@ 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)
@@ -138,6 +146,7 @@ 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/params.go

@@ -47,6 +47,10 @@ 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 {
@@ -119,11 +123,6 @@ func (p *Params) SetAudioCtx(n int) {
 	p.audio_ctx = C.int(n)
 }
 
-// Set initial prompt
-func (p *Params) SetInitialPrompt(prompt string) {
-	p.initial_prompt = C.CString(prompt)
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS
 
@@ -148,7 +147,6 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	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))
 	if p.translate {
 		str += " translate"
 	}
@@ -173,6 +171,9 @@ 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

@@ -76,6 +76,11 @@ 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)
 }
@@ -125,11 +130,6 @@ func (context *context) SetAudioCtx(n uint) {
 	context.params.SetAudioCtx(int(n))
 }
 
-// Set initial prompt
-func (context *context) SetInitialPrompt(prompt string) {
-	context.params.SetInitialPrompt(prompt)
-}
-
 // ResetTimings resets the mode timings. Should be called before processing
 func (context *context) ResetTimings() {
 	context.model.ctx.Whisper_reset_timings()

bindings/go/pkg/whisper/interface.go

@@ -38,17 +38,17 @@ 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
-	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
+	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
 
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the

bindings/go/whisper.go

@@ -10,7 +10,7 @@ import (
 
 /*
 #cgo LDFLAGS: -lwhisper -lm -lstdc++
-#cgo darwin LDFLAGS: -framework Accelerate -framework Metal -framework Foundation -framework CoreGraphics
+#cgo darwin LDFLAGS: -framework Accelerate
 #include <whisper.h>
 #include <stdlib.h>
 

bindings/java/build.gradle

@@ -9,7 +9,6 @@ archivesBaseName = 'whispercpp'
 group = 'io.github.ggerganov'
 version = '1.4.0'
 
-
 sourceCompatibility = 1.8
 targetCompatibility = 1.8
 

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

@@ -2,7 +2,6 @@ package io.github.ggerganov.whispercpp;
 
 import com.sun.jna.Native;
 import com.sun.jna.Pointer;
-import io.github.ggerganov.whispercpp.bean.WhisperSegment;
 import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
@@ -10,8 +9,6 @@ import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
 import java.io.File;
 import java.io.FileNotFoundException;
 import java.io.IOException;
-import java.util.ArrayList;
-import java.util.List;
 
 /**
  * Before calling most methods, you must call `initContext(modelPath)` to initialise the `ctx` Pointer.
@@ -163,28 +160,6 @@ public class WhisperCpp implements AutoCloseable {
 
         return str.toString().trim();
     }
-    public List<WhisperSegment> fullTranscribeWithTime(WhisperFullParams whisperParams, float[] audioData) throws IOException {
-        if (ctx == null) {
-            throw new IllegalStateException("Model not initialised");
-        }
-
-        if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
-            throw new IOException("Failed to process audio");
-        }
-
-        int nSegments = lib.whisper_full_n_segments(ctx);
-        List<WhisperSegment> segments= new ArrayList<>(nSegments);
-
-
-        for (int i = 0; i < nSegments; i++) {
-            long t0 = lib.whisper_full_get_segment_t0(ctx, i);
-            String text = lib.whisper_full_get_segment_text(ctx, i);
-            long t1 = lib.whisper_full_get_segment_t1(ctx, i);
-            segments.add(new WhisperSegment(t0,t1,text));
-        }
-
-        return segments;
-    }
 
 //    public int getTextSegmentCount(Pointer ctx) {
 //        return lib.whisper_full_n_segments(ctx);

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:
@@ -304,6 +304,14 @@ 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/bean/WhisperSegment.java

@@ -1,47 +0,0 @@
-package io.github.ggerganov.whispercpp.bean;
-
-/**
- * Created by litonglinux@qq.com on 10/21/2023_7:48 AM
- */
-public class WhisperSegment {
-  private long start, end;
-  private String sentence;
-
-  public WhisperSegment() {
-  }
-
-  public WhisperSegment(long start, long end, String sentence) {
-    this.start = start;
-    this.end = end;
-    this.sentence = sentence;
-  }
-
-  public long getStart() {
-    return start;
-  }
-
-  public long getEnd() {
-    return end;
-  }
-
-  public String getSentence() {
-    return sentence;
-  }
-
-  public void setStart(long start) {
-    this.start = start;
-  }
-
-  public void setEnd(long end) {
-    this.end = end;
-  }
-
-  public void setSentence(String sentence) {
-    this.sentence = sentence;
-  }
-
-  @Override
-  public String toString() {
-    return "[" + start + " --> " + end + "]:" + sentence;
-  }
-}

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

@@ -58,9 +58,6 @@ public class WhisperFullParams extends Structure {
         no_context = enable ? CBool.FALSE : CBool.TRUE;
     }
 
-    /** Generate timestamps or not? */
-    public CBool no_timestamps;
-
     /** Flag to force single segment output (useful for streaming). (default = false) */
     public CBool single_segment;
 
@@ -129,6 +126,14 @@ 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;
 
@@ -140,9 +145,6 @@ public class WhisperFullParams extends Structure {
         tdrz_enable = enable ? CBool.TRUE : CBool.FALSE;
     }
 
-    /** Regular expression matching tokens to suppress. */
-    public String suppress_regex;
-
     /** Tokens to provide to the whisper decoder as an initial prompt.
      * These are prepended to any existing text context from a previous call. */
     public String initial_prompt;
@@ -302,25 +304,18 @@ public class WhisperFullParams extends Structure {
         logits_filter_callback = CallbackReference.getFunctionPointer(callback);
     }
 
-    /** Grammar stuff */
-    public Pointer grammar_rules;
-    public long n_grammar_rules;
-    public long i_start_rule;
-    public float grammar_penalty;
-
     @Override
     protected List<String> getFieldOrder() {
         return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
-                "no_context", "single_segment", "no_timestamps",
+                "no_context", "single_segment",
                 "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
-                "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",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "tdrz_enable", "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",
                 "new_segment_callback", "new_segment_callback_user_data",
                 "progress_callback", "progress_callback_user_data",
                 "encoder_begin_callback", "encoder_begin_callback_user_data",
-                "logits_filter_callback", "logits_filter_callback_user_data",
-                "grammar_rules", "n_grammar_rules", "i_start_rule", "grammar_penalty");
+                "logits_filter_callback", "logits_filter_callback_user_data");
     }
 }

bindings/java/src/test/java/io/github/ggerganov/whispercpp/WhisperCppTest.java

@@ -2,7 +2,6 @@ package io.github.ggerganov.whispercpp;
 
 import static org.junit.jupiter.api.Assertions.*;
 
-import io.github.ggerganov.whispercpp.bean.WhisperSegment;
 import io.github.ggerganov.whispercpp.params.CBool;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
@@ -12,7 +11,6 @@ import javax.sound.sampled.AudioInputStream;
 import javax.sound.sampled.AudioSystem;
 import java.io.File;
 import java.io.FileNotFoundException;
-import java.util.List;
 
 class WhisperCppTest {
     private static WhisperCpp whisper = new WhisperCpp();
@@ -22,12 +20,11 @@ class WhisperCppTest {
     static void init() throws FileNotFoundException {
         // By default, models are loaded from ~/.cache/whisper/ and are usually named "ggml-${name}.bin"
         // or you can provide the absolute path to the model file.
-        //String modelName = "../../models/ggml-tiny.bin";
         String modelName = "../../models/ggml-tiny.en.bin";
         try {
             whisper.initContext(modelName);
-            //whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
-            //whisper.getJavaDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+//            whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+//            whisper.getJavaDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
             modelInitialised = true;
         } catch (FileNotFoundException ex) {
             System.out.println("Model " + modelName + " not found");
@@ -45,7 +42,7 @@ class WhisperCppTest {
         assertEquals(16384, params.n_max_text_ctx);
         assertFalse(params.translate);
         assertEquals(0.01f, params.thold_pt);
-        assertEquals(5, params.beam_search.beam_size);
+        assertEquals(2, params.beam_search.beam_size);
         assertEquals(-1.0f, params.beam_search.patience);
     }
 
@@ -58,7 +55,7 @@ class WhisperCppTest {
         assertEquals(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY.ordinal(), params.strategy);
         assertNotEquals(0, params.n_threads);
         assertEquals(16384, params.n_max_text_ctx);
-        assertEquals(5, params.greedy.best_of);
+        assertEquals(2, params.greedy.best_of);
     }
 
     @Test
@@ -75,11 +72,11 @@ class WhisperCppTest {
         byte[] b = new byte[audioInputStream.available()];
         float[] floats = new float[b.length / 2];
 
-        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+//        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
         WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
         params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
         params.print_progress = CBool.FALSE;
-        //params.initial_prompt = "and so my fellow Americans um, like";
+//        params.initial_prompt = "and so my fellow Americans um, like";
 
 
         try {
@@ -102,43 +99,4 @@ class WhisperCppTest {
             audioInputStream.close();
         }
     }
-
-    @Test
-    void testFullTranscribeWithTime() throws Exception {
-        if (!modelInitialised) {
-            System.out.println("Model not initialised, skipping test");
-            return;
-        }
-
-        // Given
-        File file = new File(System.getProperty("user.dir"), "../../samples/jfk.wav");
-        AudioInputStream audioInputStream = AudioSystem.getAudioInputStream(file);
-
-        byte[] b = new byte[audioInputStream.available()];
-        float[] floats = new float[b.length / 2];
-
-        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
-        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
-        params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
-        params.print_progress = CBool.FALSE;
-        //params.initial_prompt = "and so my fellow Americans um, like";
-
-        try {
-            audioInputStream.read(b);
-
-            for (int i = 0, j = 0; i < b.length; i += 2, j++) {
-                int intSample = (int) (b[i + 1]) << 8 | (int) (b[i]) & 0xFF;
-                floats[j] = intSample / 32767.0f;
-            }
-
-            List<WhisperSegment> segments = whisper.fullTranscribeWithTime(params, floats);
-            assertTrue(segments.size() > 0, "The size of segments should be greater than 0");
-            for (WhisperSegment segment : segments) {
-                System.out.println(segment);
-            }
-        } finally {
-            audioInputStream.close();
-        }
-    }
-
 }

bindings/javascript/README.md

@@ -41,7 +41,7 @@ make publish-npm
 
 ## Sample run
 
-```text
+```java
 $ node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js
 
 whisper_model_load: loading model from 'whisper.bin'
@@ -63,7 +63,7 @@ whisper_model_load: ggml ctx size =  140.60 MB
 whisper_model_load: memory size   =   22.83 MB
 whisper_model_load: model size    =  140.54 MB
 
-system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 |
+system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 | 
 
 operator(): processing 176000 samples, 11.0 sec, 8 threads, 1 processors, lang = en, task = transcribe ...
 

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.6.2",
+  "version": "1.4.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/Rakefile

@@ -1,12 +0,0 @@
-require 'rake/clean'
-  require 'rubygems/package'
-
-desc 'Build gem'
-task :package do
-  spec_source = File.read File.join(File.dirname(__FILE__),'whispercpp.gemspec')
-  spec = nil
-  # see: http://gist.github.com/16215
-  Thread.new { spec = eval("#{spec_source}") }.join
-  spec.validate
-  Gem::Package.build(spec)
-end

bindings/ruby/ext/extconf.rb

@@ -1,7 +1,6 @@
 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__),'..','..','..','whisper-mel.hpp')} .")
 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')} .")
@@ -10,7 +9,6 @@ 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')} .")

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

@@ -12,63 +12,31 @@ extern "C" {
     // Backend buffer
     //
 
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        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_buffer_type_t    buft;
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
         ggml_backend_buffer_context_t context;
+
         size_t size;
-        enum ggml_backend_buffer_usage usage;
     };
 
-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
+    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);
 
-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
     //
     // Backend
     //
@@ -76,66 +44,44 @@ extern "C" {
     typedef void * ggml_backend_context_t;
 
     struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+        const char * (*get_name)(ggml_backend_t backend);
 
-        void (*GGML_CALL free)(ggml_backend_t backend);
+        void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
 
-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);
 
-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(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);
 
-        // compute graph with a plan (not used currently)
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // (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
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        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
+        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
     };
 
     struct ggml_backend {
-        ggml_guid_t guid;
-
         struct ggml_backend_i iface;
+
         ggml_backend_context_t context;
     };
 
-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
 #ifdef  __cplusplus
 }
 #endif

bindings/ruby/ext/ggml-backend.h

@@ -7,123 +7,69 @@
 extern "C" {
 #endif
 
-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
     //
     // Backend buffer
     //
 
-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
 
-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-    };
-
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (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 GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+    // 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
     //
 
-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t 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_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
 
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
 
-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    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 ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    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 void 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);
 
-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
-
     //
     // CPU backend
     //
 
     GGML_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    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_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
 
-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
 
     //
     // Backend scheduler
@@ -137,96 +83,53 @@ extern "C" {
     /*
       Example usage:
 
-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        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
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
 
-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
+        // 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 assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+            // 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);
+        }
 
-        ggml_backend_sched_reserve(sched, reserve_graph);
+        // 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);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
     */
 
     struct ggml_backend_sched;
     typedef struct ggml_backend_sched * ggml_backend_sched_t;
 
-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+    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 bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
 
-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+    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 size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    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
 }

bindings/ruby/ext/ggml-cuda.h

@@ -1,43 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-impl.h

@@ -5,7 +5,6 @@
 // GGML internal header
 
 #include <assert.h>
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
@@ -19,7 +18,6 @@ extern "C" {
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef __cplusplus
 #ifndef static_assert
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
 #define static_assert(cond, msg) _Static_assert(cond, msg)
@@ -27,7 +25,6 @@ extern "C" {
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
 #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__))
@@ -37,18 +34,17 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#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
@@ -60,30 +56,14 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-typedef __fp16 ggml_fp16_internal_t;
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
 
-#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)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
 
 #else
 
-typedef uint16_t ggml_fp16_internal_t;
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -237,7 +217,8 @@ 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)
+#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));
@@ -245,23 +226,19 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
 #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)
 
-struct ggml_hash_set ggml_hash_set_new(size_t size);
-
 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_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// 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

bindings/ruby/ext/ggml-kompute.h

@@ -1,46 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-#ifdef __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-metal.h

@@ -1,66 +0,0 @@
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
-
-#ifdef __cplusplus
-}
-#endif
-

bindings/ruby/ext/ggml-opencl.h

@@ -1,36 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-GGML_API void ggml_cl_init(void);
-
-GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
-GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-// GGML_API void * ggml_cl_host_malloc(size_t size);
-// GGML_API void   ggml_cl_host_free(void * ptr);
-
-GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-// backend API
-
-// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
-
-// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
-// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-quants.h

@@ -1,133 +1,224 @@
 #pragma once
 
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
+#include "ggml-impl.h"
 
 // GGML internal header
 
-#ifdef __cplusplus
-extern "C" {
+#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 * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+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 * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t 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_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t 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_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t 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 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+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 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t 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 * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
-
-#ifdef __cplusplus
-}
-#endif
+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/ggml-sycl.h

@@ -1,49 +0,0 @@
-//
-//  MIT license
-//  Copyright (C) 2024 Intel Corporation
-//  SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
-
-// devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
-
-GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
-
-// SYCL doesn't support registering host memory, keep here for reference
-// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ggml-vulkan.h

@@ -1,29 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_VK_NAME "Vulkan"
-#define GGML_VK_MAX_DEVICES 16
-
-GGML_API void ggml_vk_instance_init(void);
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
-
-GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif

bindings/ruby/ext/ruby_whisper.cpp

@@ -311,6 +311,12 @@ static VALUE ruby_whisper_params_get_split_on_word(VALUE self) {
 static VALUE ruby_whisper_params_set_split_on_word(VALUE self, VALUE value) {
   BOOL_PARAMS_SETTER(self, split_on_word, value)
 }
+static VALUE ruby_whisper_params_get_speed_up(VALUE self) {
+  BOOL_PARAMS_GETTER(self, speed_up)
+}
+static VALUE ruby_whisper_params_set_speed_up(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, speed_up, value)
+}
 static VALUE ruby_whisper_params_get_diarize(VALUE self) {
   ruby_whisper_params *rwp;
   Data_Get_Struct(self, ruby_whisper_params, rwp);
@@ -402,6 +408,8 @@ void Init_whisper() {
   rb_define_method(cParams, "token_timestamps=", ruby_whisper_params_set_token_timestamps, 1);
   rb_define_method(cParams, "split_on_word", ruby_whisper_params_get_split_on_word, 0);
   rb_define_method(cParams, "split_on_word=", ruby_whisper_params_set_split_on_word, 1);
+  rb_define_method(cParams, "speed_up", ruby_whisper_params_get_speed_up, 0);
+  rb_define_method(cParams, "speed_up=", ruby_whisper_params_set_speed_up, 1);
   rb_define_method(cParams, "diarize", ruby_whisper_params_get_diarize, 0);
   rb_define_method(cParams, "diarize=", ruby_whisper_params_set_diarize, 1);
 

bindings/ruby/tests/test_whisper.rb

@@ -117,6 +117,13 @@ class TestWhisper < Test::Unit::TestCase
     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'))
     params  = Whisper::Params.new

bindings/ruby/whispercpp.gemspec

@@ -1,28 +0,0 @@
-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 = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"]
-
-  #### Load-time details
-  s.require_paths = ['lib','ext']
-  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/FindFFmpeg.cmake

@@ -1,163 +0,0 @@
-# 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})
-

coreml/whisper-encoder-impl.h

@@ -123,7 +123,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 
 /**
     Make a prediction using the convenience interface
-    @param logmel_data as 1 × n_mel × 3000 3-dimensional array of floats:
+    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
     @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
     @return the prediction as whisper_encoder_implOutput
 */

coreml/whisper-encoder.h

@@ -3,8 +3,6 @@
 // Code is derived from the work of Github user @wangchou
 // ref: https://github.com/wangchou/callCoreMLFromCpp
 
-#include <stdint.h>
-
 #if __cplusplus
 extern "C" {
 #endif
@@ -16,8 +14,6 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx);
 
 void whisper_coreml_encode(
         const whisper_coreml_context * ctx,
-                             int64_t   n_ctx,
-                             int64_t   n_mel,
                                float * mel,
                                float * out);
 

coreml/whisper-encoder.mm

@@ -24,9 +24,9 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     // select which device to run the Core ML model on
     MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
-    // config.computeUnits = MLComputeUnitsCPUAndGPU;
+    config.computeUnits = MLComputeUnitsCPUAndGPU;
     //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
-    config.computeUnits = MLComputeUnitsAll;
+    //config.computeUnits = MLComputeUnitsAll;
 
     const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 
@@ -48,15 +48,13 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx) {
 
 void whisper_coreml_encode(
         const whisper_coreml_context * ctx,
-                             int64_t   n_ctx,
-                             int64_t   n_mel,
                                float * mel,
                                float * out) {
     MLMultiArray * inMultiArray = [
         [MLMultiArray alloc] initWithDataPointer: mel
-                                           shape: @[@1, @(n_mel), @(n_ctx)]
+                                           shape: @[@1, @80, @3000]
                                         dataType: MLMultiArrayDataTypeFloat32
-                                         strides: @[@(n_ctx*n_mel), @(n_ctx), @1]
+                                         strides: @[@(240000), @(3000), @1]
                                      deallocator: nil
                                            error: nil
     ];

examples/CMakeLists.txt

@@ -14,26 +14,15 @@ if (WHISPER_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
-if (WHISPER_CLBLAST)
-    find_package(CLBlast REQUIRED)
-endif()
-
 # common
 
 set(TARGET common)
 
-if (WHISPER_FFMPEG)
-    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
-endif()
-
 add_library(${TARGET} STATIC
     common.h
     common.cpp
     common-ggml.h
     common-ggml.cpp
-    grammar-parser.h
-    grammar-parser.cpp
-    ${COMMON_SOURCES_FFMPEG}
     )
 
 include(DefaultTargetOptions)
@@ -41,7 +30,6 @@ include(DefaultTargetOptions)
 target_link_libraries(${TARGET} PRIVATE whisper)
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 
 if (WHISPER_SDL2)
     # common-sdl
@@ -59,63 +47,27 @@ if (WHISPER_SDL2)
     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)
-target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
-
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
     add_subdirectory(whisper.wasm)
-    set_target_properties(libmain PROPERTIES FOLDER "libs")
     add_subdirectory(stream.wasm)
-    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(bench.wasm)
-    set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
     add_subdirectory(addon.node)
-    set_target_properties(addon.node PROPERTIES FOLDER "examples")
 else()
     add_subdirectory(main)
-    set_target_properties(main PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
     add_subdirectory(stream)
-    set_target_properties(stream PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
-    add_subdirectory(server)
-    set_target_properties(server PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
     add_subdirectory(command)
-    set_target_properties(command PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
     add_subdirectory(bench)
-    set_target_properties(bench PROPERTIES FOLDER "examples")
     add_subdirectory(quantize)
-    set_target_properties(quantize PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
     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)
-        add_subdirectory(sycl)
-        set_target_properties(sycl PROPERTIES FOLDER "examples")
-    endif()
-endif (WHISPER_SDL2)
 endif()
-
-if (WHISPER_SDL2)
-    add_subdirectory(wchess)
-    set_target_properties(wchess PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)

examples/addon.node/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET addon.node)
+set(TARGET whisper-addon)
 
 # 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 addon.node -B Release
+npx cmake-js compile -T whisper-addon -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 addon.node -B Release
+> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -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/addon.node"
+  "../../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");
 
@@ -12,12 +12,6 @@ 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;
@@ -36,10 +36,7 @@ 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;
@@ -47,8 +44,6 @@ 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 {
@@ -57,6 +52,27 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
 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;
@@ -88,8 +104,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
         if (params.diarize && pcmf32s.size() == 2) {
             const int64_t n_samples = pcmf32s[0].size();
 
-            const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
-            const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+            const int64_t is0 = timestamp_to_sample(t0, n_samples);
+            const int64_t is1 = timestamp_to_sample(t1, n_samples);
 
             double energy0 = 0.0f;
             double energy1 = 0.0f;
@@ -125,15 +141,9 @@ 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.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");
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
         return 2;
     }
 
@@ -144,9 +154,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     // whisper init
 
-    struct whisper_context_params cparams = whisper_context_default_params();
+    struct whisper_context_params cparams;
     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) {
@@ -154,14 +163,6 @@ 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];
@@ -169,25 +170,20 @@ 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
 
-        // 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;
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
         }
 
         // 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) {
@@ -196,13 +192,12 @@ 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, audio_ctx = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %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.audio_ctx);
+                    params.no_timestamps ? 0 : 1);
 
             fprintf(stderr, "\n");
         }
@@ -229,15 +224,14 @@ 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.audio_ctx        = params.audio_ctx;
+
+            wparams.speed_up         = params.speed_up;
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
             wparams.initial_prompt   = params.prompt.c_str();
 
-            wparams.no_timestamps    = params.no_timestamps;
-
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
             // this callback is called on each new segment
@@ -273,8 +267,8 @@ 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, params.comma_in_time));
-        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
+        result[i].emplace_back(to_timestamp(t0, true));
+        result[i].emplace_back(to_timestamp(t1, true));
         result[i].emplace_back(text);
     }
 
@@ -325,33 +319,11 @@ 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/addon.node"
+  "../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");
 
@@ -10,27 +10,15 @@ const whisperAsync = promisify(whisper);
 const whisperParams = {
   language: "en",
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
-  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  fname_inp: "../../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("--")) {
-      const [key, value] = item.slice(2).split("=");
-      if (key === "audio_ctx") {
-        whisperParams[key] = parseInt(value);
-      } else {
-        whisperParams[key] = value;
-      }
-      return pre;
+      return [...pre, item.slice(2).split("=")];
     }
     return pre;
   }, [])
@@ -45,6 +33,5 @@ for (const key in params) {
 console.log("whisperParams =", whisperParams);
 
 whisperAsync(whisperParams).then((result) => {
-  console.log();
-  console.log(result);
+  console.log(`Result from whisper: ${result}`);
 });

examples/addon.node/package.json

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

examples/bench/bench.cpp

@@ -8,12 +8,11 @@
 // command-line parameters
 struct whisper_params {
     int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t what = 0; // what to benchmark: 0 - whisper encoder, 1 - memcpy, 2 - ggml_mul_mat
+    int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
 
     std::string model = "models/ggml-base.en.bin";
 
-    bool use_gpu    = true;
-    bool flash_attn = false;
+    bool use_gpu = true;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -26,11 +25,10 @@ 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 == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
+        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 {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -51,7 +49,6 @@ 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",                            "");
@@ -61,10 +58,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
 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.flash_attn = params.flash_attn;
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
@@ -86,7 +81,7 @@ int whisper_bench_full(const whisper_params & params) {
     }
     // heat encoder
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;
     }
 
@@ -95,13 +90,13 @@ int whisper_bench_full(const whisper_params & params) {
 
     // prompt heat
     if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;
     }
 
     // text-generation heat
     if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;
     }
 
@@ -109,30 +104,20 @@ int whisper_bench_full(const whisper_params & params) {
 
     // actual run
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
         return 4;
     }
 
-    // text-generation
-    for (int i = 0; i < 256; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
-            return 4;
-        }
-    }
-
-    // batched decoding
-    for (int i = 0; i < 64; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
-            return 4;
-        }
-    }
-
-    // prompt processing
     for (int i = 0; i < 16; i++) {
         if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+            return 4;
+        }
+    }
+
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to encode model: %d\n", ret);
             return 4;
         }
     }

examples/command/README.md

@@ -37,13 +37,9 @@ https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9
 The `command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev
 
-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2
 

examples/command/command.cpp

@@ -9,7 +9,6 @@
 #include "common-sdl.h"
 #include "common.h"
 #include "whisper.h"
-#include "grammar-parser.h"
 
 #include <sstream>
 #include <cassert>
@@ -31,30 +30,21 @@ struct whisper_params {
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
 
-    float vad_thold  = 0.6f;
-    float freq_thold = 100.0f;
-
-    float grammar_penalty = 100.0f;
-
-    grammar_parser::parse_state grammar_parsed;
+    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 use_gpu       = true;
-    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
     std::string fname_out;
     std::string commands;
     std::string prompt;
-    std::string context;
-    std::string grammar;
-
-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -75,20 +65,16 @@ 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]; }
         else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
         else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
-        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
-        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
-        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
-        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -113,41 +99,26 @@ 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());
     fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
     fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
-    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
-    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
-    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
-    fprintf(stderr, "  --suppress-regex REGEX       [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(
-                 whisper_context * ctx,
-            const whisper_params & params,
-        const std::vector<float> & pcmf32,
-               const std::string & grammar_rule,
-                           float & logprob_min,
-                           float & logprob_sum,
-                             int & n_tokens,
-                         int64_t & t_ms) {
+std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
-    logprob_min = 0.0f;
-    logprob_sum = 0.0f;
-    n_tokens    = 0;
+    prob = 0.0f;
     t_ms = 0;
 
-    //whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
     wparams.print_progress   = false;
     wparams.print_special    = params.print_special;
@@ -155,42 +126,19 @@ std::string transcribe(
     wparams.print_timestamps = !params.no_timestamps;
     wparams.translate        = params.translate;
     wparams.no_context       = true;
-    wparams.no_timestamps    = params.no_timestamps;
     wparams.single_segment   = true;
     wparams.max_tokens       = params.max_tokens;
     wparams.language         = params.language.c_str();
     wparams.n_threads        = params.n_threads;
 
-    wparams.audio_ctx = params.audio_ctx;
-
-    wparams.temperature     = 0.4f;
-    wparams.temperature_inc = 1.0f;
-    wparams.greedy.best_of  = 5;
-
-    wparams.beam_search.beam_size = 5;
-
-    wparams.initial_prompt = params.context.data();
-
-    wparams.suppress_regex = params.suppress_regex.c_str();
-
-    const auto & grammar_parsed = params.grammar_parsed;
-    auto grammar_rules = grammar_parsed.c_rules();
-
-    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
-        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
-            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, grammar_rule.c_str());
-        } else {
-            wparams.grammar_rules   = grammar_rules.data();
-            wparams.n_grammar_rules = grammar_rules.size();
-            wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
-            wparams.grammar_penalty = params.grammar_penalty;
-        }
-    }
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
     }
 
+    int prob_n = 0;
     std::string result;
 
     const int n_segments = whisper_full_n_segments(ctx);
@@ -199,17 +147,19 @@ std::string transcribe(
 
         result += text;
 
-        const int n = whisper_full_n_tokens(ctx, i);
-        for (int j = 0; j < n; ++j) {
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
             const auto token = whisper_full_get_token_data(ctx, i, j);
 
-            if(token.plog > 0.0f) exit(0);
-            logprob_min = std::min(logprob_min, token.plog);
-            logprob_sum += token.plog;
-            ++n_tokens;
+            prob += token.p;
+            ++prob_n;
         }
     }
 
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
     const auto t_end = std::chrono::high_resolution_clock::now();
     t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
 
@@ -300,7 +250,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
         fprintf(stderr, " ]\n");
     }
 
-    std::string k_prompt = "select one from the available words: ";
+    std::string  k_prompt = "select one from the available words: ";
     for (int i = 0; i < (int) allowed_commands.size(); ++i) {
         if (i > 0) {
             k_prompt += ", ";
@@ -367,6 +317,7 @@ 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();
@@ -467,9 +418,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
     bool is_running = true;
     bool ask_prompt = true;
 
-    float logprob_min = 0.0f;
-    float logprob_sum = 0.0f;
-    int   n_tokens    = 0;
+    float prob = 0.0f;
 
     std::vector<float> pcmf32_cur;
 
@@ -507,7 +456,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
                 // detect the commands
                 audio.get(params.command_ms, pcmf32_cur);
 
-                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "", logprob_min, logprob_sum, n_tokens, t_ms));
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
 
                 const auto words = get_words(txt);
 
@@ -543,27 +492,18 @@ 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) {
+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;
 
-    float logprob_min0 = 0.0f;
-    float logprob_min  = 0.0f;
-
-    float logprob_sum0 = 0.0f;
-    float logprob_sum  = 0.0f;
-
-    int n_tokens0 = 0;
-    int n_tokens  = 0;
+    float prob0 = 0.0f;
+    float prob  = 0.0f;
 
     std::vector<float> pcmf32_cur;
     std::vector<float> pcmf32_prompt;
 
-    std::string k_prompt = "Ok Whisper, start listening for commands.";
-    if (!params.prompt.empty()) {
-        k_prompt = params.prompt;
-    }
+    const std::string k_prompt = "Ok Whisper, start listening for commands.";
 
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: general-purpose mode\n", __func__);
@@ -596,11 +536,9 @@ int process_general_transcription(struct whisper_context * ctx, audio_async & au
                     // wait for activation phrase
                     audio.get(params.prompt_ms, pcmf32_cur);
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "prompt", logprob_min0, logprob_sum0, n_tokens0, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
 
-                    const float p = 100.0f * std::exp(logprob_min0);
-
-                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms, p = %.2f%%)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms, p);
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
 
                     const float sim = similarity(txt, k_prompt);
 
@@ -621,30 +559,19 @@ int process_general_transcription(struct whisper_context * ctx, audio_async & au
                     // we have heard the activation phrase, now detect the commands
                     audio.get(params.command_ms, pcmf32_cur);
 
-                    //printf("len prompt:  %.4f\n", pcmf32_prompt.size() / (float) WHISPER_SAMPLE_RATE);
-                    //printf("len command: %.4f\n", pcmf32_cur.size() / (float) WHISPER_SAMPLE_RATE);
-
-                    // prepend 3 second of silence
-                    pcmf32_cur.insert(pcmf32_cur.begin(), 3.0f*WHISPER_SAMPLE_RATE, 0.0f);
-
                     // prepend the prompt audio
                     pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "root", logprob_min, logprob_sum, n_tokens, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
 
-                    //const float p = 100.0f * std::exp((logprob - logprob0) / (n_tokens - n_tokens0));
-                    const float p = 100.0f * std::exp(logprob_min);
+                    prob = 100.0f*(prob - prob0);
 
                     //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
 
                     // find the prompt in the text
                     float best_sim = 0.0f;
                     size_t best_len = 0;
-                    for (size_t n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
-                        if (n >= txt.size()) {
-                            break;
-                        }
-
+                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
                         const auto prompt = txt.substr(0, n);
 
                         const float sim = similarity(prompt, k_prompt);
@@ -657,16 +584,9 @@ int process_general_transcription(struct whisper_context * ctx, audio_async & au
                         }
                     }
 
-                    fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
-                    if (best_len == 0) {
-                        fprintf(stdout, "%s: WARNING: command not recognized, try again\n", __func__);
-                    } else {
-                        // cut the prompt from the decoded text
-                        const std::string command = ::trim(txt.substr(best_len));
-
-                        fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
-                    }
+                    const std::string command = ::trim(txt.substr(best_len));
 
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
                     fprintf(stdout, "\n");
                 }
 
@@ -693,10 +613,8 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
@@ -736,36 +654,12 @@ int main(int argc, char ** argv) {
 
     int  ret_val = 0;
 
-    if (!params.grammar.empty()) {
-        auto & grammar = params.grammar_parsed;
-        if (is_file_exist(params.grammar.c_str())) {
-            // read grammar from file
-            std::ifstream ifs(params.grammar.c_str());
-            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
-            grammar = grammar_parser::parse(txt.c_str());
-        } else {
-            // read grammar from string
-            grammar = grammar_parser::parse(params.grammar.c_str());
-        }
-
-        // will be empty (default) if there are parse errors
-        if (grammar.rules.empty()) {
-            ret_val = 1;
-        } else {
-            fprintf(stderr, "%s: grammar:\n", __func__);
-            grammar_parser::print_grammar(stderr, grammar);
-            fprintf(stderr, "\n");
-        }
-    }
-
-    if (ret_val == 0) {
-        if (!params.commands.empty()) {
-            ret_val = process_command_list(ctx, audio, params);
-        } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
-            ret_val = always_prompt_transcription(ctx, audio, params);
-        } else {
-            ret_val = process_general_transcription(ctx, audio, params);
-        }
+    if (!params.commands.empty()) {
+        ret_val = process_command_list(ctx, audio, params);
+    } else if (!params.prompt.empty()) {
+        ret_val = always_prompt_transcription(ctx, audio, params);
+    } else {
+        ret_val = process_general_transcription(ctx, audio, params);
     }
 
     audio.pause();

examples/common-ggml.cpp

@@ -9,11 +9,6 @@ static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
     {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
     {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
     {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
-    {"q2_k", GGML_FTYPE_MOSTLY_Q2_K},
-    {"q3_k", GGML_FTYPE_MOSTLY_Q3_K},
-    {"q4_k", GGML_FTYPE_MOSTLY_Q4_K},
-    {"q5_k", GGML_FTYPE_MOSTLY_Q5_K},
-    {"q6_k", GGML_FTYPE_MOSTLY_Q6_K},
 };
 
 void ggml_print_ftypes(FILE * fp) {
@@ -53,25 +48,15 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
         case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
         case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
-        case GGML_FTYPE_MOSTLY_Q2_K: qtype = GGML_TYPE_Q2_K; break;
-        case GGML_FTYPE_MOSTLY_Q3_K: qtype = GGML_TYPE_Q3_K; break;
-        case GGML_FTYPE_MOSTLY_Q4_K: qtype = GGML_TYPE_Q4_K; break;
-        case GGML_FTYPE_MOSTLY_Q5_K: qtype = GGML_TYPE_Q5_K; break;
-        case GGML_FTYPE_MOSTLY_Q6_K: qtype = GGML_TYPE_Q6_K; break;
         case GGML_FTYPE_UNKNOWN:
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
-        case GGML_FTYPE_MOSTLY_IQ2_XXS:
-        case GGML_FTYPE_MOSTLY_IQ2_XS:
-        case GGML_FTYPE_MOSTLY_IQ2_S:
-        case GGML_FTYPE_MOSTLY_IQ3_XXS:
-        case GGML_FTYPE_MOSTLY_IQ3_S:
-        case GGML_FTYPE_MOSTLY_IQ1_S:
-        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_Q2_K:
+        case GGML_FTYPE_MOSTLY_Q3_K:
+        case GGML_FTYPE_MOSTLY_Q4_K:
+        case GGML_FTYPE_MOSTLY_Q5_K:
+        case GGML_FTYPE_MOSTLY_Q6_K:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -92,6 +77,8 @@ bool ggml_common_quantize_0(
     std::vector<ggml_fp16_t> data_f16;
     std::vector<float>       data_f32;
 
+    std::vector<int64_t> hist_all(1 << 4, 0);
+
     while (true) {
         int32_t n_dims;
         int32_t length;
@@ -176,39 +163,41 @@ bool ggml_common_quantize_0(
             work.resize(nelements); // for quantization
 
             size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q5_0:
-                case GGML_TYPE_Q5_1:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q2_K:
-                case GGML_TYPE_Q3_K:
-                case GGML_TYPE_Q4_K:
-                case GGML_TYPE_Q5_K:
-                case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], nullptr);
+                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_1:
+                    {
+                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_0:
+                    {
+                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_1:
+                    {
+                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q8_0:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_I8:
                 case GGML_TYPE_I16:
                 case GGML_TYPE_I32:
-                case GGML_TYPE_I64:
-                case GGML_TYPE_F64:
                 case GGML_TYPE_Q8_1:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
                 case GGML_TYPE_Q8_K:
-                case GGML_TYPE_IQ2_XXS:
-                case GGML_TYPE_IQ2_XS:
-                case GGML_TYPE_IQ2_S:
-                case GGML_TYPE_IQ3_XXS:
-                case GGML_TYPE_IQ3_S:
-                case GGML_TYPE_IQ1_S:
-                case GGML_TYPE_IQ4_NL:
-                case GGML_TYPE_IQ4_XS:
-                case GGML_TYPE_IQ1_M:
-                case GGML_TYPE_BF16:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
@@ -219,7 +208,15 @@ bool ggml_common_quantize_0(
             fout.write(reinterpret_cast<char *>(work.data()), cur_size);
             total_size_new += cur_size;
 
-            printf("size = %8.2f MB -> %8.2f MB\n", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
         } else {
             printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
             fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
@@ -232,5 +229,18 @@ bool ggml_common_quantize_0(
     printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
     printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
 
+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
     return true;
 }

examples/common-sdl.cpp

@@ -139,13 +139,10 @@ void audio_async::callback(uint8_t * stream, int len) {
         return;
     }
 
-    size_t n_samples = len / sizeof(float);
+    const size_t n_samples = len / sizeof(float);
 
-    if (n_samples > m_audio.size()) {
-        n_samples = m_audio.size();
-
-        stream += (len - (n_samples * sizeof(float)));
-    }
+    m_audio_new.resize(n_samples);
+    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
 
     //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
 
@@ -156,7 +153,7 @@ void audio_async::callback(uint8_t * stream, int len) {
             const size_t n0 = m_audio.size() - m_audio_pos;
 
             memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], stream + n0 * sizeof(float), (n_samples - n0) * sizeof(float));
+            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
 
             m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
             m_audio_len = m_audio.size();

examples/common-sdl.h

@@ -41,6 +41,7 @@ private:
     std::mutex       m_mutex;
 
     std::vector<float> m_audio;
+    std::vector<float> m_audio_new;
     size_t             m_audio_pos = 0;
     size_t             m_audio_len = 0;
 };

examples/common.cpp

@@ -19,16 +19,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#ifdef _WIN32
-#include <fcntl.h>
-#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) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
@@ -625,31 +615,12 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 
 }
 
-bool is_wav_buffer(const std::string buf) {
-    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
-    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
-    if (buf.size() < 12 || buf.substr(0, 4) != "RIFF" || buf.substr(8, 4) != "WAVE") {
-        return false;
-    }
-
-    uint32_t chunk_size = *reinterpret_cast<const uint32_t*>(buf.data() + 4);
-    if (chunk_size + 8 != buf.size()) {
-        return false;
-    }
-
-    return true;
-}
-
 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 or ffmpeg decoding output
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
 
     if (fname == "-") {
         {
-            #ifdef _WIN32
-            _setmode(_fileno(stdin), _O_BINARY);
-            #endif
-
             uint8_t buf[1024];
             while (true)
             {
@@ -668,49 +639,28 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
 
         fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
     }
-    else if (is_wav_buffer(fname)) {
-        if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
-            fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
-            return false;
-        }
-    }
     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) {
         fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
         return false;
     }
 
     if (stereo && wav.channels != 2) {
         fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.sampleRate != COMMON_SAMPLE_RATE) {
         fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
-        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.bitsPerSample != 16) {
         fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
         return false;
     }
 
@@ -865,48 +815,3 @@ void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
     fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
-
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma) {
-    int64_t msec = t * 10;
-    int64_t hr = msec / (1000 * 60 * 60);
-    msec = msec - hr * (1000 * 60 * 60);
-    int64_t min = msec / (1000 * 60);
-    msec = msec - min * (1000 * 60);
-    int64_t sec = msec / 1000;
-    msec = msec - sec * 1000;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
-
-    return std::string(buf);
-}
-
-int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate) {
-    return std::max(0, std::min((int) n_samples - 1, (int) ((t*whisper_sample_rate)/100)));
-}
-
-bool is_file_exist(const char *fileName)
-{
-    std::ifstream infile(fileName);
-    return infile.good();
-}
-
-bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id)
-{
-    std::ofstream speak_file(path.c_str());
-    if (speak_file.fail()) {
-        fprintf(stderr, "%s: failed to open speak_file\n", __func__);
-        return false;
-    } else {
-        speak_file.write(text.c_str(), text.size());
-        speak_file.close();
-        int ret = system((command + " " + std::to_string(voice_id) + " " + path).c_str());
-        if (ret != 0) {
-            fprintf(stderr, "%s: failed to speak\n", __func__);
-            return false;
-        }
-    }
-    return true;
-}

examples/common.h

@@ -135,11 +135,7 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 // Audio utils
 //
 
-// Check if a buffer is a WAV audio file
-bool is_wav_buffer(const std::string buf);
-
 // Read WAV audio file and store the PCM data into pcmf32
-// fname can be a buffer of WAV data instead of a filename
 // The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
 // If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
 bool read_wav(
@@ -185,7 +181,7 @@ 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 = int16_t(data[i] * 32767);
+            const auto intSample = static_cast<const int16_t>(data[i] * 32767);
             file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
             dataSize += sizeof(int16_t);
         }
@@ -281,31 +277,3 @@ struct sam_params {
 bool sam_params_parse(int argc, char ** argv, sam_params & params);
 
 void sam_print_usage(int argc, char ** argv, const sam_params & params);
-
-//
-// Terminal utils
-//
-
-
-// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
-// Lowest is red, middle is yellow, highest is green.
-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",
-};
-
-//
-// Other utils
-//
-
-// convert timestamp to string, 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma = false);
-
-// given a timestamp get the sample
-int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate);
-
-// check if file exists using ifstream
-bool is_file_exist(const char *fileName);
-
-// write text to file, and call system("command voice_id file")
-bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id);

examples/ffmpeg-transcode.cpp

@@ -1,350 +0,0 @@
-/* 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];
-
-    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
-
-    fmt_ctx = avformat_alloc_context();
-    avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
-    LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);
-    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: %x size: %d\n", ibuf, 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

@@ -1,423 +0,0 @@
-#include "grammar-parser.h"
-#include <cstdint>
-#include <cwchar>
-#include <string>
-#include <utility>
-#include <stdexcept>
-#include <exception>
-
-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 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;
-        int      len        = lookup[highbits];
-        uint8_t  mask       = (1 << (8 - len)) - 1;
-        uint32_t value      = first_byte & mask;
-        const char * end    = src + len; // may overrun!
-        const char * pos    = src + 1;
-        for ( ; pos < end && *pos; pos++) {
-            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
-        }
-        return std::make_pair(value, pos);
-    }
-
-    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) {
-        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(
-            parse_state & state,
-            uint32_t      rule_id,
-            const std::vector<whisper_grammar_element> & rule) {
-        if (state.rules.size() <= rule_id) {
-            state.rules.resize(rule_id + 1);
-        }
-        state.rules[rule_id] = rule;
-    }
-
-    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) {
-        const char * pos   = src;
-        const char * end   = src + size;
-        uint32_t     value = 0;
-        for ( ; pos < end && *pos; pos++) {
-            value <<= 4;
-            char c = *pos;
-            if ('a' <= c && c <= 'f') {
-                value += c - 'a' + 10;
-            } else if ('A' <= c && c <= 'F') {
-                value += c - 'A' + 10;
-            } else if ('0' <= c && c <= '9') {
-                value += c - '0';
-            } else {
-                break;
-            }
-        }
-        if (pos != end) {
-            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
-        }
-        return std::make_pair(value, pos);
-    }
-
-    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'))) {
-            if (*pos == '#') {
-                while (*pos && *pos != '\r' && *pos != '\n') {
-                    pos++;
-                }
-            } else {
-                pos++;
-            }
-        }
-        return pos;
-    }
-
-    const char * parse_name(const char * src) {
-        const char * pos = src;
-        while (is_word_char(*pos)) {
-            pos++;
-        }
-        if (pos == src) {
-            throw std::runtime_error(std::string("expecting name at ") + src);
-        }
-        return pos;
-    }
-
-    std::pair<uint32_t, const char *> parse_char(const char * src) {
-        if (*src == '\\') {
-            switch (src[1]) {
-                case 'x': return parse_hex(src + 2, 2);
-                case 'u': return parse_hex(src + 2, 4);
-                case 'U': return parse_hex(src + 2, 8);
-                case 't': return std::make_pair('\t', src + 2);
-                case 'r': return std::make_pair('\r', src + 2);
-                case 'n': return std::make_pair('\n', src + 2);
-                case '\\':
-                case '"':
-                case '[':
-                case ']':
-                    return std::make_pair(src[1], src + 2);
-                default:
-                    throw std::runtime_error(std::string("unknown escape at ") + src);
-            }
-        } else if (*src) {
-            return decode_utf8(src);
-        }
-        throw std::runtime_error("unexpected end of input");
-    }
-
-    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(
-            parse_state                        & state,
-            const char                         * src,
-            const std::string                  & rule_name,
-            std::vector<whisper_grammar_element> & out_elements,
-            bool                                 is_nested) {
-        size_t last_sym_start = out_elements.size();
-        const char * pos = src;
-        while (*pos) {
-            if (*pos == '"') { // literal string
-                pos++;
-                last_sym_start = out_elements.size();
-                while (*pos != '"') {
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    out_elements.push_back({WHISPER_GRETYPE_CHAR, char_pair.first});
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '[') { // char range(s)
-                pos++;
-                enum whisper_gretype start_type = WHISPER_GRETYPE_CHAR;
-                if (*pos == '^') {
-                    pos++;
-                    start_type = WHISPER_GRETYPE_CHAR_NOT;
-                }
-                last_sym_start = out_elements.size();
-                while (*pos != ']') {
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    enum whisper_gretype type = last_sym_start < out_elements.size()
-                        ? WHISPER_GRETYPE_CHAR_ALT
-                        : start_type;
-
-                    out_elements.push_back({type, char_pair.first});
-                    if (pos[0] == '-' && pos[1] != ']') {
-                        auto endchar_pair = parse_char(pos + 1);
-                             pos          = endchar_pair.second;
-                        out_elements.push_back({WHISPER_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
-                    }
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (is_word_char(*pos)) { // rule reference
-                const char * name_end    = parse_name(pos);
-                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
-                pos = parse_space(name_end, is_nested);
-                last_sym_start = out_elements.size();
-                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, ref_rule_id});
-            } else if (*pos == '(') { // grouping
-                // parse nested alternates into synthesized rule
-                pos = parse_space(pos + 1, true);
-                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
-                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
-                last_sym_start = out_elements.size();
-                // output reference to synthesized rule
-                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
-                if (*pos != ')') {
-                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
-                if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
-                }
-
-                // apply transformation to previous symbol (last_sym_start to end) according to
-                // rewrite rules:
-                // S* --> S' ::= S S' |
-                // S+ --> S' ::= S S' | S
-                // S? --> S' ::= S |
-                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
-                std::vector<whisper_grammar_element> sub_rule;
-                // add preceding symbol to generated rule
-                sub_rule.insert(
-                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
-                if (*pos == '*' || *pos == '+') {
-                    // cause generated rule to recurse
-                    sub_rule.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
-                }
-                // mark start of alternate def
-                sub_rule.push_back({WHISPER_GRETYPE_ALT, 0});
-                if (*pos == '+') {
-                    // add preceding symbol as alternate only for '+' (otherwise empty)
-                    sub_rule.insert(
-                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
-                }
-                sub_rule.push_back({WHISPER_GRETYPE_END, 0});
-                add_rule(state, sub_rule_id, sub_rule);
-
-                // in original rule, replace previous symbol with reference to generated rule
-                out_elements.resize(last_sym_start);
-                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
-
-                pos = parse_space(pos + 1, is_nested);
-            } else {
-                break;
-            }
-        }
-        return pos;
-    }
-
-    const char * parse_alternates(
-            parse_state       & state,
-            const char        * src,
-            const std::string & rule_name,
-            uint32_t            rule_id,
-            bool                is_nested) {
-        std::vector<whisper_grammar_element> rule;
-        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
-        while (*pos == '|') {
-            rule.push_back({WHISPER_GRETYPE_ALT, 0});
-            pos = parse_space(pos + 1, true);
-            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
-        }
-        rule.push_back({WHISPER_GRETYPE_END, 0});
-        add_rule(state, rule_id, rule);
-        return pos;
-    }
-
-    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;
-        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
-        const std::string name(src, name_len);
-
-        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
-            throw std::runtime_error(std::string("expecting ::= at ") + pos);
-        }
-        pos = parse_space(pos + 3, true);
-
-        pos = parse_alternates(state, pos, name, rule_id, false);
-
-        if (*pos == '\r') {
-            pos += pos[1] == '\n' ? 2 : 1;
-        } else if (*pos == '\n') {
-            pos++;
-        } else if (*pos) {
-            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
-        }
-        return parse_space(pos, true);
-    }
-
-    parse_state parse(const char * src) {
-        try {
-            parse_state state;
-            const char * pos = parse_space(src, true);
-            while (*pos) {
-                pos = parse_rule(state, pos);
-            }
-            return state;
-        } catch (const std::exception & err) {
-            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
-            return parse_state();
-        }
-    }
-
-    void print_grammar_char(FILE * file, uint32_t c) {
-        if (0x20 <= c && c <= 0x7f) {
-            fprintf(file, "%c", static_cast<char>(c));
-        } else {
-            // cop out of encoding UTF-8
-            fprintf(file, "<U+%04X>", c);
-        }
-    }
-
-    bool is_char_element(whisper_grammar_element elem) {
-        switch (elem.type) {
-            case WHISPER_GRETYPE_CHAR:           return true;
-            case WHISPER_GRETYPE_CHAR_NOT:       return true;
-            case WHISPER_GRETYPE_CHAR_ALT:       return true;
-            case WHISPER_GRETYPE_CHAR_RNG_UPPER: return true;
-            default:                           return false;
-        }
-    }
-
-    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;
-                case WHISPER_GRETYPE_ALT:            fprintf(file, "ALT");            break;
-                case WHISPER_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
-                case WHISPER_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
-                case WHISPER_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
-                case WHISPER_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
-                case WHISPER_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
-            }
-            switch (elem.type) {
-                case WHISPER_GRETYPE_END:
-                case WHISPER_GRETYPE_ALT:
-                case WHISPER_GRETYPE_RULE_REF:
-                    fprintf(file, "(%u) ", elem.value);
-                    break;
-                case WHISPER_GRETYPE_CHAR:
-                case WHISPER_GRETYPE_CHAR_NOT:
-                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
-                case WHISPER_GRETYPE_CHAR_ALT:
-                    fprintf(file, "(\"");
-                    print_grammar_char(file, elem.value);
-                    fprintf(file, "\") ");
-                    break;
-            }
-        }
-        fprintf(file, "\n");
-    }
-
-    void print_rule(
-            FILE     * file,
-            uint32_t   rule_id,
-            const std::vector<whisper_grammar_element> & rule,
-            const std::map<uint32_t, std::string>    & symbol_id_names) {
-        if (rule.empty() || rule.back().type != WHISPER_GRETYPE_END) {
-            throw std::runtime_error(
-                "malformed rule, does not end with WHISPER_GRETYPE_END: " + std::to_string(rule_id));
-        }
-        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
-        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
-            whisper_grammar_element elem = rule[i];
-            switch (elem.type) {
-                case WHISPER_GRETYPE_END:
-                    throw std::runtime_error(
-                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
-                        std::to_string(i));
-                case WHISPER_GRETYPE_ALT:
-                    fprintf(file, "| ");
-                    break;
-                case WHISPER_GRETYPE_RULE_REF:
-                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
-                    break;
-                case WHISPER_GRETYPE_CHAR:
-                    fprintf(file, "[");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case WHISPER_GRETYPE_CHAR_NOT:
-                    fprintf(file, "[^");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
-                    if (i == 0 || !is_char_element(rule[i - 1])) {
-                        throw std::runtime_error(
-                            "WHISPER_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
-                            std::to_string(rule_id) + "," + std::to_string(i));
-                    }
-                    fprintf(file, "-");
-                    print_grammar_char(file, elem.value);
-                    break;
-                case WHISPER_GRETYPE_CHAR_ALT:
-                    if (i == 0 || !is_char_element(rule[i - 1])) {
-                        throw std::runtime_error(
-                            "WHISPER_GRETYPE_CHAR_ALT without preceding char: " +
-                            std::to_string(rule_id) + "," + std::to_string(i));
-                    }
-                    print_grammar_char(file, elem.value);
-                    break;
-            }
-            if (is_char_element(elem)) {
-                switch (rule[i + 1].type) {
-                    case WHISPER_GRETYPE_CHAR_ALT:
-                    case WHISPER_GRETYPE_CHAR_RNG_UPPER:
-                        break;
-                    default:
-                        fprintf(file, "] ");
-                }
-            }
-        }
-        fprintf(file, "\n");
-    }
-
-    void print_grammar(FILE * file, const parse_state & state) {
-        try {
-            std::map<uint32_t, std::string> symbol_id_names;
-            for (auto kv : state.symbol_ids) {
-                symbol_id_names[kv.second] = kv.first;
-            }
-            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
-                // fprintf(file, "%zu: ", i);
-                // print_rule_binary(file, state.rules[i]);
-                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
-                // fprintf(file, "\n");
-            }
-        } catch (const std::exception & err) {
-            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
-        }
-    }
-
-    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());
-        }
-        return ret;
-    }
-}

examples/grammar-parser.h

@@ -1,29 +0,0 @@
-// Implements a parser for an extended Backus-Naur form (BNF), producing the
-// binary context-free grammar format specified by whisper.h. Supports character
-// ranges, grouping, and repetition operators. As an example, a grammar for
-// arithmetic might look like:
-//
-// root  ::= expr
-// expr  ::= term ([-+*/] term)*
-// term  ::= num | "(" space expr ")" space
-// num   ::= [0-9]+ space
-// space ::= [ \t\n]*
-
-#pragma once
-#include "whisper.h"
-#include <vector>
-#include <map>
-#include <cstdint>
-#include <string>
-
-namespace grammar_parser {
-    struct parse_state {
-        std::map<std::string, uint32_t>                   symbol_ids;
-        std::vector<std::vector<whisper_grammar_element>> rules;
-
-        std::vector<const whisper_grammar_element *>      c_rules() const;
-    };
-
-    parse_state parse(const char * src);
-    void print_grammar(FILE * file, const parse_state & state);
-}

examples/helpers.js

@@ -22,7 +22,6 @@ var printTextarea = (function() {
 async function clearCache() {
     if (confirm('Are you sure you want to clear the cache?\nAll the models will be downloaded again.')) {
         indexedDB.deleteDatabase(dbName);
-        location.reload();
     }
 }
 
@@ -34,6 +33,9 @@ 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" )
 
 # list available models
 function list_models {

examples/lsp/CMakeLists.txt

@@ -5,5 +5,5 @@ if (WHISPER_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_link_libraries(${TARGET} PRIVATE common json_cpp common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

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";
@@ -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 {
@@ -100,11 +100,11 @@ 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, "\n");
@@ -181,6 +181,7 @@ 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) {
@@ -219,6 +220,7 @@ 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
@@ -433,11 +435,8 @@ int main(int argc, char ** argv) {
     }
 
     // whisper init
-    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
     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 ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/main/README.md

@@ -17,37 +17,28 @@ options:
   -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
   -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
   -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
   -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [5      ] beam size for beam search
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
   -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
   -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
   -tr,       --translate         [false  ] translate from source language to english
   -di,       --diarize           [false  ] stereo audio diarization
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
   -otxt,     --output-txt        [false  ] output result in a text file
   -ovtt,     --output-vtt        [false  ] output result in a vtt file
   -osrt,     --output-srt        [false  ] output result in a srt file
-  -olrc,     --output-lrc        [false  ] output result in a lrc file
   -owts,     --output-words      [false  ] output script for generating karaoke video
-  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
   -oj,       --output-json       [false  ] output result in a JSON file
-  -ojf,      --output-json-full  [false  ] include more information in the JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
   -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -nt,       --no-timestamps     [true   ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
              --prompt PROMPT     [       ] initial prompt
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of tokens
-  -ng,       --no-gpu            [false  ] disable GPU
 ```

examples/main/main.cpp

@@ -1,12 +1,10 @@
 #include "common.h"
 
 #include "whisper.h"
-#include "grammar-parser.h"
 
 #include <cmath>
 #include <fstream>
 #include <cstdio>
-#include <regex>
 #include <string>
 #include <thread>
 #include <vector>
@@ -16,6 +14,34 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
+// Lowest is red, middle is yellow, highest is green.
+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",
+};
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
 // helper function to replace substrings
 void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     for (size_t pos = 0; ; pos += replace.length()) {
@@ -28,25 +54,22 @@ void replace_all(std::string & s, const std::string & search, const std::string
 
 // command-line parameters
 struct whisper_params {
-    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors  = 1;
-    int32_t offset_t_ms   = 0;
-    int32_t offset_n      = 0;
-    int32_t duration_ms   = 0;
-    int32_t progress_step = 5;
-    int32_t max_context   = -1;
-    int32_t max_len       = 0;
-    int32_t best_of       = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
-    int32_t beam_size     = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
-    int32_t audio_ctx     = 0;
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
+    int32_t max_context  = -1;
+    int32_t max_len      =  0;
+    int32_t best_of      =  2;
+    int32_t beam_size    = -1;
 
-    float word_thold      =  0.01f;
-    float entropy_thold   =  2.40f;
-    float logprob_thold   = -1.00f;
-    float grammar_penalty = 100.0f;
-    float temperature     = 0.0f;
-    float temperature_inc = 0.2f;
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;
 
+    bool speed_up        = false;
     bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
@@ -62,48 +85,29 @@ struct whisper_params {
     bool output_jsn      = false;
     bool output_jsn_full = false;
     bool output_lrc      = false;
-    bool no_prints       = false;
     bool print_special   = false;
     bool print_colors    = false;
     bool print_progress  = false;
     bool no_timestamps   = false;
     bool log_score       = false;
     bool use_gpu         = true;
-    bool flash_attn      = false;
 
     std::string language  = "en";
     std::string prompt;
     std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model     = "models/ggml-base.en.bin";
-    std::string grammar;
-    std::string grammar_rule;
 
     // [TDRZ] speaker turn string
     std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
 
-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
-
     std::string openvino_encode_device = "CPU";
 
-    std::string dtw = "";
-
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
-
-    grammar_parser::parse_state grammar_parsed;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-char* whisper_param_turn_lowercase(char* in){
-    int string_len = strlen(in);
-    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) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
@@ -131,12 +135,10 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
         else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
         else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
         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 == "-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 == "-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; }
@@ -153,25 +155,18 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
         else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
         else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
-        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
         else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
         else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
         else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
         else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = whisper_param_turn_lowercase(argv[++i]); }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
         else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
         else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
         else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
         else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
         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 == "-ls"   || arg == "--log-score")       { params.log_score       = 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 == "--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]); }
+        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score = true; }
+        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);
@@ -198,12 +193,10 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
     fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
     fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
-    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
     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, "  -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, "  -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");
@@ -219,25 +212,18 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
     fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
-    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
     fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
     fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
     fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
     fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt (max n_text_ctx/2 tokens)\n",       params.prompt.c_str());
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
     fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
     fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
-    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());
-    fprintf(stderr, "  --grammar-penalty N            [%-7.1f] scales down logits of nongrammar tokens\n",      params.grammar_penalty);
     fprintf(stderr, "\n");
 }
 
@@ -252,8 +238,8 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
 
-    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
-    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+    const int64_t is0 = timestamp_to_sample(t0, n_samples);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples);
 
     double energy0 = 0.0f;
     double energy1 = 0.0f;
@@ -477,38 +463,6 @@ char *escape_double_quotes_and_backslashes(const char *str) {
     return escaped;
 }
 
-// double quote should be escaped by another double quote. (rfc4180)
-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;
-}
-
 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()) {
@@ -530,7 +484,7 @@ 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_in_csv(text);
+        char * text_escaped = escape_double_quotes_and_backslashes(text);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
         fout << 10 * t0 << "," << 10 * t1 << ",";
@@ -709,8 +663,7 @@ bool output_json(
                                     times_o(token.t0, token.t1, false);
                                 }
                                 value_i("id", token.id, false);
-                                value_f("p", token.p, false);
-                                value_f("t_dtw", token.t_dtw, true);
+                                value_f("p", token.p, true);
                             end_obj(j == (n - 1));
                         }
                         end_arr(!params.diarize && !params.tinydiarize);
@@ -899,59 +852,14 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-
-void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
-
 int main(int argc, char ** argv) {
     whisper_params params;
 
-    // If the only argument starts with "@", read arguments line-by-line
-    // from the given file.
-    std::vector<std::string> vec_args;
-    if (argc == 2 && argv != nullptr && argv[1] != nullptr && argv[1][0] == '@') {
-        // Save the name of the executable.
-        vec_args.push_back(argv[0]);
-
-        // Open the response file.
-        char const * rspfile = argv[1] + sizeof(char);
-        std::ifstream fin(rspfile);
-        if (fin.is_open() == false) {
-            fprintf(stderr, "error: response file '%s' not found\n", rspfile);
-            return 1;
-        }
-
-        // Read the entire response file.
-        std::string line;
-        while (std::getline(fin, line)) {
-            vec_args.push_back(line);
-        }
-
-        // Use the contents of the response file as the command-line arguments.
-        argc = static_cast<int>(vec_args.size());
-        argv = static_cast<char **>(alloca(argc * sizeof (char *)));
-        for (int i = 0; i < argc; ++i) {
-            argv[i] = const_cast<char *>(vec_args[i].c_str());
-        }
-    }
-
     if (whisper_params_parse(argc, argv, params) == false) {
         whisper_print_usage(argc, argv, params);
         return 1;
     }
 
-    // remove non-existent files
-    for (auto it = params.fname_inp.begin(); it != params.fname_inp.end();) {
-        const auto fname_inp = it->c_str();
-
-        if (*it != "-" && !is_file_exist(fname_inp)) {
-            fprintf(stderr, "error: input file not found '%s'\n", fname_inp);
-            it = params.fname_inp.erase(it);
-            continue;
-        }
-
-        it++;
-    }
-
     if (params.fname_inp.empty()) {
         fprintf(stderr, "error: no input files specified\n");
         whisper_print_usage(argc, argv, params);
@@ -970,38 +878,10 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
-    if (params.no_prints) {
-        whisper_log_set(cb_log_disable, NULL);
-    }
-
     // whisper init
 
-    struct whisper_context_params cparams = whisper_context_default_params();
-
-    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;
-
-        if (params.dtw == "tiny")      cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
-        if (params.dtw == "tiny.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
-        if (params.dtw == "base")      cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
-        if (params.dtw == "base.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
-        if (params.dtw == "small")     cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
-        if (params.dtw == "small.en")  cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
-        if (params.dtw == "medium")    cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
-        if (params.dtw == "medium.en") cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
-        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 (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
-            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
-            return 3;
-        }
-    }
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
@@ -1013,29 +893,6 @@ int main(int argc, char ** argv) {
     // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
     whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
 
-    if (!params.grammar.empty()) {
-        auto & grammar = params.grammar_parsed;
-        if (is_file_exist(params.grammar.c_str())) {
-            // read grammar from file
-            std::ifstream ifs(params.grammar.c_str());
-            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
-            grammar = grammar_parser::parse(txt.c_str());
-        } else {
-            // read grammar from string
-            grammar = grammar_parser::parse(params.grammar.c_str());
-        }
-
-        // will be empty (default) if there are parse errors
-        if (grammar.rules.empty()) {
-            fprintf(stderr, "error: failed to parse grammar \"%s\"\n", params.grammar.c_str());
-            return 4;
-        } else {
-            fprintf(stderr, "%s: grammar:\n", __func__);
-            grammar_parser::print_grammar(stderr, grammar);
-            fprintf(stderr, "\n");
-        }
-    }
-
     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];
@@ -1048,28 +905,29 @@ int main(int argc, char ** argv) {
             continue;
         }
 
-        if (!whisper_is_multilingual(ctx)) {
-            if (params.language != "en" || params.translate) {
-                params.language = "en";
-                params.translate = false;
-                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-            }
-        }
-        if (params.detect_language) {
-            params.language = "auto";
-        }
-
-        if (!params.no_prints) {
-            // print system information
+        // print system information
+        {
             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
+        // print some info about the processing
+        {
             fprintf(stderr, "\n");
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            if (params.detect_language) {
+                params.language = "auto";
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
-                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
+                    params.n_threads, params.n_processors,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
                     params.tinydiarize ? "tdrz = 1, " : "",
@@ -1082,8 +940,7 @@ int main(int argc, char ** argv) {
         {
             whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
-            const bool use_grammar = (!params.grammar_parsed.rules.empty() && !params.grammar_rule.empty());
-            wparams.strategy = (params.beam_size > 1 || use_grammar) ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
 
             wparams.print_realtime   = false;
             wparams.print_progress   = params.print_progress;
@@ -1101,43 +958,23 @@ int main(int argc, char ** argv) {
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
             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.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 : params.temperature_inc;
-            wparams.temperature      = params.temperature;
-
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
-            wparams.no_timestamps    = params.no_timestamps;
-
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
-            const auto & grammar_parsed = params.grammar_parsed;
-            auto grammar_rules = grammar_parsed.c_rules();
-
-            if (use_grammar) {
-                if (grammar_parsed.symbol_ids.find(params.grammar_rule) == grammar_parsed.symbol_ids.end()) {
-                    fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, params.grammar_rule.c_str());
-                } else {
-                    wparams.grammar_rules = grammar_rules.data();
-                    wparams.n_grammar_rules = grammar_rules.size();
-                    wparams.i_start_rule = grammar_parsed.symbol_ids.at(params.grammar_rule);
-                    wparams.grammar_penalty = params.grammar_penalty;
-                }
-            }
-
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
                 wparams.new_segment_callback           = whisper_print_segment_callback;
@@ -1234,9 +1071,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if (!params.no_prints) {
-        whisper_print_timings(ctx);
-    }
+    whisper_print_timings(ctx);
     whisper_free(ctx);
 
     return 0;

examples/python/test_whisper_processor.py

@@ -1,7 +0,0 @@
-import whisper_processor
-
-try:
-    result = whisper_processor.process_audio("./audio/wake_word_detected16k.wav", "base.en")
-    print(result)
-except Exception as e:
-    print(f"Error: {e}")

examples/python/whisper_processor.py

@@ -1,54 +0,0 @@
-import subprocess
-import sys
-import os
-
-def process_audio(wav_file, model_name="base.en"):
-    """
-    Processes an audio file using a specified model and returns the processed string.
-
-    :param wav_file: Path to the WAV file
-    :param model_name: Name of the model to use
-    :return: Processed string output from the audio processing
-    :raises: Exception if an error occurs during processing
-    """
-
-    model = f"./models/ggml-{model_name}.bin"
-
-    # Check if the file exists
-    if not os.path.exists(model):
-        raise FileNotFoundError(f"Model file not found: {model} \n\nDownload a model with this command:\n\n> bash ./models/download-ggml-model.sh {model_name}\n\n")
-
-    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"
-
-    # Execute the command
-    process = subprocess.Popen(full_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-
-    # Get the output and error (if any)
-    output, error = process.communicate()
-
-    if error:
-        raise Exception(f"Error processing audio: {error.decode('utf-8')}")
-
-    # Process and return the output string
-    decoded_str = output.decode('utf-8').strip()
-    processed_str = decoded_str.replace('[BLANK_AUDIO]', '').strip()
-
-    return processed_str
-
-def main():
-    if len(sys.argv) >= 2:
-        wav_file = sys.argv[1]
-        model_name = sys.argv[2] if len(sys.argv) == 3 else "base.en"
-        try:
-            result = process_audio(wav_file, model_name)
-            print(result)
-        except Exception as e:
-            print(f"Error: {e}")
-    else:
-        print("Usage: python whisper_processor.py <wav_file> [<model_name>]")
-
-if __name__ == "__main__":
-    main()

examples/server/CMakeLists.txt

@@ -1,10 +0,0 @@
-set(TARGET server)
-add_executable(${TARGET} server.cpp httplib.h)
-
-include(DefaultTargetOptions)
-
-target_link_libraries(${TARGET} PRIVATE common json_cpp whisper ${CMAKE_THREAD_LIBS_INIT})
-
-if (WIN32)
-    target_link_libraries(${TARGET} PRIVATE ws2_32)
-endif()

examples/server/README.md

@@ -1,69 +0,0 @@
-# whisper.cpp http server
-
-Simple http server. WAV Files are passed to the inference model via http requests.
-
-https://github.com/ggerganov/whisper.cpp/assets/1991296/e983ee53-8741-4eb5-9048-afe5e4594b8f
-
-## Usage
-
-```
-./server -h
-
-usage: ./bin/server [options]
-
-options:
-  -h,        --help              [default] show this help message and exit
-  -t N,      --threads N         [4      ] number of threads to use during computation
-  -p N,      --processors N      [1      ] number of processors to use during computation
-  -ot N,     --offset-t N        [0      ] time offset in milliseconds
-  -on N,     --offset-n N        [0      ] segment index offset
-  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
-  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
-  -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [2      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
-  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
-  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
-  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
-  -tr,       --translate         [false  ] translate from source language to english
-  -di,       --diarize           [false  ] stereo audio diarization
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
-  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
-  -ps,       --print-special     [false  ] print special tokens
-  -pc,       --print-colors      [false  ] print colors
-  -pr,       --print-realtime    [false  ] print output in realtime
-  -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
-  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
-             --prompt PROMPT     [       ] initial prompt
-  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  --host HOST,                   [127.0.0.1] Hostname/ip-adress for the server
-  --port PORT,                   [8080   ] Port number for the server
-  --convert,                     [false  ] Convert audio to WAV, requires ffmpeg on the server
-```
-
-> [!WARNING]
-> **Do not run the server example with administrative privileges and ensure it's operated in a sandbox environment, especially since it involves risky operations like accepting user file uploads and using ffmpeg for format conversions. Always validate and sanitize inputs to guard against potential security threats.**
-
-## request examples
-
-**/inference**
-```
-curl 127.0.0.1:8080/inference \
--H "Content-Type: multipart/form-data" \
--F file="@<file-path>" \
--F temperature="0.0" \
--F temperature_inc="0.2" \
--F response_format="json"
-```
-
-**/load**
-```
-curl 127.0.0.1:8080/load \
--H "Content-Type: multipart/form-data" \
--F model="<path-to-model-file>"
-```

examples/stream.wasm/emscripten.cpp

@@ -103,11 +103,11 @@ void stream_main(size_t index) {
 
             {
                 const int n_segments = whisper_full_n_segments(ctx);
-                if (n_segments > 0) {
-                    const char * text = whisper_full_get_segment_text(ctx, n_segments - 1);
+                for (int i = n_segments - 1; i < n_segments; ++i) {
+                    const char * text = whisper_full_get_segment_text(ctx, i);
 
-                    const int64_t t0 = whisper_full_get_segment_t0(ctx, n_segments - 1);
-                    const int64_t t1 = whisper_full_get_segment_t1(ctx, n_segments - 1);
+                    const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                    const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
                     printf("transcribed: %s\n", text);
 

examples/stream/README.md

@@ -4,7 +4,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The `stream` tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
-```bash
+```java
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
@@ -14,7 +14,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 
 Setting the `--step` argument to `0` enables the sliding window mode:
 
-```bash
+```java
  ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
@@ -30,21 +30,17 @@ a transcription block that is suitable for parsing.
 The `stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev
 
-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2
 
 make stream
 ```
 
-Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
+Ensure you are at the root of the repo when running `make stream`.  Not within the `examples/stream` dir
+as the libraries needed like `common-sdl.h` are located within `examples`.  Attempting to compile within
 `examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
 
 ```bash

examples/stream/stream.cpp

@@ -14,6 +14,20 @@
 #include <fstream>
 
 
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t) {
+    int64_t sec = t/100;
+    int64_t msec = t - sec*100;
+    int64_t min = sec/60;
+    sec = sec - min*60;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
+
+    return std::string(buf);
+}
+
 // command-line parameters
 struct whisper_params {
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -27,6 +41,7 @@ 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;
@@ -35,7 +50,6 @@ 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";
@@ -61,6 +75,7 @@ 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; }
@@ -71,7 +86,6 @@ 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());
@@ -98,6 +112,7 @@ 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");
@@ -108,7 +123,6 @@ 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");
 }
 
@@ -152,10 +166,8 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
-    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
@@ -311,6 +323,7 @@ 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]
 
@@ -359,7 +372,7 @@ int main(int argc, char ** argv) {
                         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-                        std::string output = "[" + to_timestamp(t0, false) + " --> " + to_timestamp(t1, false) + "]  " + text;
+                        std::string output = "[" + to_timestamp(t0) + " --> " + to_timestamp(t1) + "]  " + text;
 
                         if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
                             output += " [SPEAKER_TURN]";

examples/sycl/CMakeLists.txt

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

examples/sycl/README.md

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

examples/sycl/build.sh

@@ -1,19 +0,0 @@
-#  MIT license
-#  Copyright (C) 2024 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DWHISPER_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
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v

examples/sycl/ls-sycl-device.cpp

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

examples/sycl/run-whisper.sh

@@ -1,17 +0,0 @@
-#!/bin/bash
-
-#  MIT license
-#  Copyright (C) 2024 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
-source /opt/intel/oneapi/setvars.sh
-
-if [ $# -gt 0 ]; then
-    export GGML_SYCL_DEVICE=$1
-else
-    export GGML_SYCL_DEVICE=0
-fi
-echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
-#export GGML_SYCL_DEBUG=1
-./build/bin/main -m models/ggml-base.en.bin -f samples/jfk.wav

examples/talk-llama/.gitignore

@@ -1,2 +1 @@
 audio.mp3
-to_speak.txt

examples/talk-llama/CMakeLists.txt

@@ -1,18 +1,25 @@
 if (WHISPER_SDL2)
     # talk-llama
     set(TARGET talk-llama)
-    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp unicode-data.cpp)
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 
-    if (WHISPER_CLBLAST)
-        set(CLBLAST_LIBNAME clblast)
-    endif ()
-    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CLBLAST_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET}
+        talk-llama.cpp
+        llama.cpp
+        ../common.cpp
+        ../common-sdl.cpp
+        ../../ggml.c
+        ../../ggml-alloc.c
+        ../../ggml-backend.c
+        ../../ggml-quants.c
+        ../../whisper.cpp)
 
-    if(WIN32)
-        # It requires Windows 8.1 or later for PrefetchVirtualMemory
-        target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
-    endif()
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 
     include(DefaultTargetOptions)
 endif ()

examples/talk-llama/README.md

@@ -15,13 +15,9 @@ https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-
 The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
-# Install SDL2
-# On Debian based linux distributions:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev
 
-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2
 

examples/talk-llama/eleven-labs.py

@@ -1,80 +1,20 @@
 import sys
-import argparse
-import textwrap
+import importlib.util
 
-parser = argparse.ArgumentParser(add_help=False,
-    formatter_class=argparse.RawTextHelpFormatter)
-parser.add_argument("-q", "--quick", action="store_true",
-    help="skip checking the required library")
-
-modes = parser.add_argument_group("action")
-modes.add_argument("inputfile", metavar="TEXTFILE",
-    nargs='?', type=argparse.FileType(), default=sys.stdin,
-    help="read the text file (default: stdin)")
-modes.add_argument("-l", "--list", action="store_true",
-    help="show the list of voices and exit")
-modes.add_argument("-h", "--help", action="help",
-    help="show this help and exit")
-
-selopts = parser.add_argument_group("voice selection")
-selmodes = selopts.add_mutually_exclusive_group()
-selmodes.add_argument("-n", "--name",
-    default="Arnold",
-    help="get a voice object by name (default: Arnold)")
-selmodes.add_argument("-v", "--voice", type=int, metavar="NUMBER",
-    help="get a voice object by number (see --list)")
-selopts.add_argument("-f", "--filter", action="append", metavar="KEY=VAL",
-    default=["use case=narration"],
-    help=textwrap.dedent('''\
-        filter voices by labels (default: "use case=narration")
-        this option can be used multiple times
-        filtering will be disabled if the first -f has no "=" (e.g. -f "any")
-        '''))
-
-outmodes = parser.add_argument_group("output")
-outgroup = outmodes.add_mutually_exclusive_group()
-outgroup.add_argument("-s", "--save", metavar="FILE",
-    default="audio.mp3",
-    help="save the TTS to a file (default: audio.mp3)")
-outgroup.add_argument("-p", "--play", action="store_true",
-    help="play the TTS with ffplay")
-
-args = parser.parse_args()
-
-if not args.quick:
-    import importlib.util
-    if importlib.util.find_spec("elevenlabs") is None:
-        print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
-        sys.exit()
-
-from elevenlabs import voices, generate, play, save
-
-if args.filter and "=" in args.filter[0]:
-    voicelist = voices()
-    for f in args.filter:
-        label, value = f.split("=")
-        voicelist = filter(lambda x: x.labels.get(label) == value, voicelist)
-    voicelist = list(voicelist)
-else:
-    voicelist = list(voices())
-
-if args.list:
-    for i, v in enumerate(voicelist):
-        print(str(i) + ": " + v.name + " " + str(v.labels))
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
     sys.exit()
 
-if args.voice:
-    voice = voicelist[args.voice % len(voicelist)]
-else:
-    voice = args.name
-    # if -n should consult -f, use the following
-    #voice = next(x for x in voicelist if x.name == args.name)
+from elevenlabs import generate, play, save
 
+# Get a Voice object, by name or UUID
+voice = "Arnold" #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
 audio = generate(
-    text=str(args.inputfile.read()),
-    voice=voice
+  text=str(sys.argv[2:]),
+  voice=voice
 )
-if args.play:
-    play(audio)
-else:
-    save(audio, args.save) 
+
+# Save the TTS to a file
+save(audio, "audio.mp3") 

examples/talk-llama/llama.h

@@ -2,8 +2,12 @@
 #define LLAMA_H
 
 #include "ggml.h"
-#include "ggml-backend.h"
-
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#else
+#define LLAMA_MAX_DEVICES 1
+#endif // GGML_USE_CUBLAS
 #include <stddef.h>
 #include <stdint.h>
 #include <stdio.h>
@@ -35,15 +39,15 @@
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
-#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
-#define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 6
+#define LLAMA_SESSION_VERSION 2
 
-#define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+#define LLAMA_SUPPORTS_GPU_OFFLOAD
+#endif
 
 #ifdef __cplusplus
 extern "C" {
@@ -63,36 +67,8 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-    };
-
-    // pre-tokenization types
-    enum llama_vocab_pre_type {
-        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
-        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
-        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
-        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
-        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
-        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
-        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
-    };
-
-    // note: these values should be synchronized with ggml_rope
-    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
-    enum llama_rope_type {
-        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM =  0,
-        LLAMA_ROPE_TYPE_NEOX =  2,
-        LLAMA_ROPE_TYPE_GLM  =  4,
+        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
     };
 
     enum llama_token_type {
@@ -126,43 +102,16 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_XS        = 22, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ1_S         = 24, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ4_NL        = 25, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_S         = 26, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ3_M         = 27, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ2_S         = 28, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
     enum llama_rope_scaling_type {
-        LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
-        LLAMA_ROPE_SCALING_TYPE_NONE        = 0,
-        LLAMA_ROPE_SCALING_TYPE_LINEAR      = 1,
-        LLAMA_ROPE_SCALING_TYPE_YARN        = 2,
-        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_YARN,
-    };
-
-    enum llama_pooling_type {
-        LLAMA_POOLING_TYPE_UNSPECIFIED = -1,
-        LLAMA_POOLING_TYPE_NONE = 0,
-        LLAMA_POOLING_TYPE_MEAN = 1,
-        LLAMA_POOLING_TYPE_CLS  = 2,
-    };
-
-    enum llama_split_mode {
-        LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
-        LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
-        LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
     };
 
     typedef struct llama_token_data {
@@ -177,7 +126,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef bool (*llama_progress_callback)(float progress, void * user_data);
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -187,7 +136,7 @@ extern "C" {
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
     // - seq_id : the sequence to which the respective token belongs
-    // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
+    // - logits : if zero, the logits for the respective token will not be output
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -197,7 +146,7 @@ extern "C" {
         llama_pos    *  pos;
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
-        int8_t       *  logits; // TODO: rename this to "output"
+        int8_t       *  logits;
 
         // NOTE: helpers for smooth API transition - can be deprecated in the future
         //       for future-proof code, use the above fields instead and ignore everything below
@@ -209,112 +158,54 @@ extern "C" {
         llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
-    enum llama_model_kv_override_type {
-        LLAMA_KV_OVERRIDE_TYPE_INT,
-        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
-        LLAMA_KV_OVERRIDE_TYPE_BOOL,
-        LLAMA_KV_OVERRIDE_TYPE_STR,
-    };
-
-    struct llama_model_kv_override {
-        enum llama_model_kv_override_type tag;
-
-        char key[128];
-
-        union {
-            int64_t val_i64;
-            double  val_f64;
-            bool    val_bool;
-            char    val_str[128];
-        };
-    };
-
     struct llama_model_params {
         int32_t n_gpu_layers; // number of layers to store in VRAM
-        enum llama_split_mode split_mode; // how to split the model across multiple GPUs
+        int32_t main_gpu;     // the GPU that is used for scratch and small tensors
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
-        // main_gpu interpretation depends on split_mode:
-        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
-        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
-        // LLAMA_SPLIT_LAYER: ignored
-        int32_t main_gpu;
-
-        // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
-        const float * tensor_split;
-
-        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
-        // If the provided progress_callback returns true, model loading continues.
-        // If it returns false, model loading is immediately aborted.
+        // called with a progress value between 0 and 1, pass NULL to disable
         llama_progress_callback progress_callback;
-
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
-        // override key-value pairs of the model meta data
-        const struct llama_model_kv_override * kv_overrides;
-
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
     };
 
     struct llama_context_params {
         uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
-        uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
-        uint32_t n_ubatch;          // physical maximum batch size
-        uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
+        uint32_t n_batch;           // prompt processing maximum batch size
         uint32_t n_threads;         // number of threads to use for generation
         uint32_t n_threads_batch;   // number of threads to use for batch processing
-
-        enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
-        enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
-                                                        // (ignored if no pooling layer)
+        int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
         float    rope_freq_base;   // RoPE base frequency, 0 = from model
         float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
-        float    yarn_ext_factor;  // YaRN extrapolation mix factor, negative = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, NaN = from model
         float    yarn_attn_factor; // YaRN magnitude scaling factor
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
-
-        ggml_backend_sched_eval_callback cb_eval;
-        void * cb_eval_user_data;
-
-        enum ggml_type type_k; // data type for K cache
-        enum ggml_type type_v; // data type for V cache
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
-        bool embeddings;  // if true, extract embeddings (together with logits)
-        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention
-
-        // Abort callback
-        // if it returns true, execution of llama_decode() will be aborted
-        // currently works only with CPU execution
-        ggml_abort_callback abort_callback;
-        void *              abort_callback_data;
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
+        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool embedding;  // embedding mode only
     };
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // itoken embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;      // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
     } llama_model_quantize_params;
 
     // grammar types
@@ -365,12 +256,6 @@ extern "C" {
         int32_t n_eval;
     };
 
-    // used in chat template
-    typedef struct llama_chat_message {
-        const char * role;
-        const char * content;
-    } llama_chat_message;
-
     // Helpers for getting default parameters
     LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
@@ -379,10 +264,7 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_backend_init(void);
-
-    //optional:
-    LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
+    LLAMA_API void llama_backend_init(bool numa);
 
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
@@ -402,51 +284,25 @@ extern "C" {
 
     LLAMA_API int64_t llama_time_us(void);
 
-    LLAMA_API size_t llama_max_devices(void);
-
-    LLAMA_API bool llama_supports_mmap       (void);
-    LLAMA_API bool llama_supports_mlock      (void);
-    LLAMA_API bool llama_supports_gpu_offload(void);
+    LLAMA_API int  llama_max_devices    (void);
+    LLAMA_API bool llama_mmap_supported (void);
+    LLAMA_API bool llama_mlock_supported(void);
 
     LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
 
-    LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
-    LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
-    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
-    LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
 
-    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
 
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model   * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model   * model);
-
-    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
+    LLAMA_API int llama_n_vocab    (const struct llama_model * model);
+    LLAMA_API int llama_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int llama_n_embd     (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
-    // Functions to access the model's GGUF metadata scalar values
-    // - The functions return the length of the string on success, or -1 on failure
-    // - The output string is always null-terminated and cleared on failure
-    // - GGUF array values are not supported by these functions
-
-    // Get metadata value as a string by key name
-    LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size);
-
-    // Get the number of metadata key/value pairs
-    LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
-
-    // Get metadata key name by index
-    LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
-
-    // Get metadata value as a string by index
-    LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
-
     // Get a string describing the model type
-    LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
 
     // Returns the total size of all the tensors in the model in bytes
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
@@ -458,7 +314,7 @@ extern "C" {
     LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
 
     // Returns 0 on success
-    LLAMA_API uint32_t llama_model_quantize(
+    LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
             const llama_model_quantize_params * params);
@@ -469,96 +325,38 @@ extern "C" {
     // The model needs to be reloaded before applying a new adapter, otherwise the adapter
     // will be applied on top of the previous one
     // Returns 0 on success
-    LLAMA_API int32_t llama_model_apply_lora_from_file(
-            const struct llama_model * model,
-                          const char * path_lora,
-                               float   scale,
-                          const char * path_base_model,
-                             int32_t   n_threads);
+    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                             int   n_threads),
+            "use llama_model_apply_lora_from_file instead");
 
-    // Apply a loaded control vector to a llama_context, or if data is NULL, clear
-    // the currently loaded vector.
-    // n_embd should be the size of a single layer's control, and data should point
-    // to an n_embd x n_layers buffer starting from layer 1.
-    // il_start and il_end are the layer range the vector should apply to (both inclusive)
-    // See llama_control_vector_load in common to load a control vector.
-    LLAMA_API int32_t llama_control_vector_apply(
-            struct llama_context * lctx,
-                     const float * data,
-                          size_t   len,
-                         int32_t   n_embd,
-                         int32_t   il_start,
-                         int32_t   il_end);
+    LLAMA_API int llama_model_apply_lora_from_file(
+            const struct llama_model * model,
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                             int   n_threads);
 
     //
     // KV cache
     //
 
-    // Information associated with an individual cell in the KV cache view.
-    struct llama_kv_cache_view_cell {
-        // The position for this cell. Takes KV cache shifts into account.
-        // May be negative if the cell is not populated.
-        llama_pos pos;
-    };
+    // Returns the number of tokens in the KV cache
+    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
 
-    // An updateable view of the KV cache.
-    struct llama_kv_cache_view {
-        // Number of KV cache cells. This will be the same as the context size.
-        int32_t n_cells;
-
-        // Maximum number of sequences that can exist in a cell. It's not an error
-        // if there are more sequences in a cell than this value, however they will
-        // not be visible in the view cells_sequences.
-        int32_t n_seq_max;
-
-        // Number of tokens in the cache. For example, if there are two populated
-        // cells, the first with 1 sequence id in it and the second with 2 sequence
-        // ids then you'll have 3 tokens.
-        int32_t token_count;
-
-        // Number of populated cache cells.
-        int32_t used_cells;
-
-        // Maximum contiguous empty slots in the cache.
-        int32_t max_contiguous;
-
-        // Index to the start of the max_contiguous slot range. Can be negative
-        // when cache is full.
-        int32_t max_contiguous_idx;
-
-        // Information for an individual cell.
-        struct llama_kv_cache_view_cell * cells;
-
-        // The sequences for each cell. There will be n_seq_max items per cell.
-        llama_seq_id * cells_sequences;
-    };
-
-    // Create an empty KV cache view. (use only for debugging purposes)
-    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);
-
-    // Free a KV cache view. (use only for debugging purposes)
-    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
-
-    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
-    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
-
-    // Returns the number of tokens in the KV cache (slow, use only for debug)
-    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
-
-    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
-
-    // Clear the KV cache - both cell info is erased and KV data is zeroed
+    // Clear the KV cache
     LLAMA_API void llama_kv_cache_clear(
             struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_cache_seq_rm(
+    LLAMA_API void llama_kv_cache_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -581,142 +379,76 @@ extern "C" {
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    // If the KV cache is RoPEd, the KV data is updated accordingly
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_add(
+    LLAMA_API void llama_kv_cache_seq_shift(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
                        llama_pos   delta);
 
-    // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_div(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                             int   d);
-
-    // Returns the largest position present in the KV cache for the specified sequence
-    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // Defragment the KV cache
-    // This will be applied:
-    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
-
     //
     // State / sessions
     //
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
-        "use llama_state_get_size instead");
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
 
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
     // Returns the number of bytes copied
-    LLAMA_API size_t llama_state_get_data(
+    LLAMA_API size_t llama_copy_state_data(
             struct llama_context * ctx,
                          uint8_t * dst);
-    LLAMA_API DEPRECATED(size_t llama_copy_state_data(
-            struct llama_context * ctx,
-                         uint8_t * dst),
-        "use llama_state_get_data instead");
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_state_set_data(
+    LLAMA_API size_t llama_set_state_data(
             struct llama_context * ctx,
-                   const uint8_t * src);
-    LLAMA_API DEPRECATED(size_t llama_set_state_data(
-            struct llama_context * ctx,
-                   const uint8_t * src),
-        "use llama_state_set_data instead");
+                         uint8_t * src);
 
     // Save/load session file
-    LLAMA_API bool llama_state_load_file(
+    LLAMA_API bool llama_load_session_file(
             struct llama_context * ctx,
                       const char * path_session,
                      llama_token * tokens_out,
                           size_t   n_token_capacity,
                           size_t * n_token_count_out);
-    LLAMA_API DEPRECATED(bool llama_load_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out),
-        "use llama_state_load_file instead");
 
-    LLAMA_API bool llama_state_save_file(
+    LLAMA_API bool llama_save_session_file(
             struct llama_context * ctx,
                       const char * path_session,
                const llama_token * tokens,
                           size_t   n_token_count);
-    LLAMA_API DEPRECATED(bool llama_save_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-               const llama_token * tokens,
-                          size_t   n_token_count),
-        "use llama_state_save_file instead");
-
-    // Get the exact size needed to copy the KV cache of a single sequence
-    LLAMA_API size_t llama_state_seq_get_size(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // Copy the KV cache of a single sequence into the specified buffer
-    LLAMA_API size_t llama_state_seq_get_data(
-            struct llama_context * ctx,
-                         uint8_t * dst,
-                    llama_seq_id   seq_id);
-
-    // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
-    // Returns:
-    //  - Positive: Ok
-    //  - Zero: Failed to load
-    LLAMA_API size_t llama_state_seq_set_data(
-            struct llama_context * ctx,
-                   const uint8_t * src,
-                    llama_seq_id   dest_seq_id);
-
-    LLAMA_API size_t llama_state_seq_save_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   seq_id,
-               const llama_token * tokens,
-                          size_t   n_token_count);
-
-    LLAMA_API size_t llama_state_seq_load_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   dest_seq_id,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out);
 
     //
     // Decoding
     //
 
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
+            struct llama_context * ctx,
+                     llama_token * tokens,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
+
+    // Same as llama_eval, but use float matrix input directly.
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
+            struct llama_context * ctx,
+                           float * embd,
+                         int32_t   n_tokens,
+                             int   n_past),
+            "use llama_decode() instead");
+
     // Return batch for single sequence of tokens starting at pos_0
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
@@ -746,7 +478,7 @@ extern "C" {
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
     // < 0 - error
-    LLAMA_API int32_t llama_decode(
+    LLAMA_API int llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
@@ -755,51 +487,21 @@ extern "C" {
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
     LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
-    // Set whether to use causal attention or not
-    // If set to true, the model will only attend to the past tokens
-    LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
-
-    // Set abort callback
-    LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Wait until all computations are finished
-    // This is automatically done when using one of the functions below to obtain the computation results
-    // and is not necessary to call it explicitly in most cases
-    LLAMA_API void llama_synchronize(struct llama_context * ctx);
-
-    // Token logits obtained from the last call to llama_decode()
-    // The logits for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // Rows: number of tokens for which llama_batch.logits[i] != 0
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
-    // Logits for the ith token. For positive indices, Equivalent to:
-    // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
-    // Negative indicies can be used to access logits in reverse order, -1 is the last logit.
-    // returns NULL for invalid ids.
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
     LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
 
-    // Get all output token embeddings.
-    // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,
-    // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // shape: [n_outputs*n_embd]
-    // Otherwise, returns NULL.
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
-    // Get the embeddings for the ith token. For positive indices, Equivalent to:
-    // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
-    // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
-    // shape: [n_embd] (1-dimensional)
-    // returns NULL for invalid ids.
-    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
-
-    // Get the embeddings for a sequence id
-    // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // shape: [n_embd] (1-dimensional)
-    LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
-
     //
     // Vocab
     //
@@ -810,23 +512,12 @@ extern "C" {
 
     LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
-    // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
-    LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
-
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
-    LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
-    LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
 
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t         llama_add_bos_token(const struct llama_model * model);
-
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);
-
-    // Codellama infill tokens
+    // codellama infill tokens
     LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
     LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
     LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
@@ -838,49 +529,28 @@ extern "C" {
 
     /// @details Convert the provided text into tokens.
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
-    /// @return Returns the number of tokens on success, no more than n_tokens_max
+    /// @return Returns the number of tokens on success, no more than n_max_tokens
     /// @return Returns a negative number on failure - the number of tokens that would have been returned
-    /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
-    ///                      as plaintext. Does not insert a leading space.
-    LLAMA_API int32_t llama_tokenize(
+    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
+    ///                Does not insert a leading space.
+    LLAMA_API int llama_tokenize(
         const struct llama_model * model,
                       const char * text,
-                         int32_t   text_len,
+                             int   text_len,
                      llama_token * tokens,
-                         int32_t   n_tokens_max,
-                            bool   add_special,
-                            bool   parse_special);
+                             int   n_max_tokens,
+                            bool   add_bos,
+                            bool   special);
 
     // Token Id -> Piece.
     // Uses the vocabulary in the provided context.
     // Does not write null terminator to the buffer.
     // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
-    // @param special If true, special tokens are rendered in the output.
-    LLAMA_API int32_t llama_token_to_piece(
+    LLAMA_API int llama_token_to_piece(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
-                               int32_t   length,
-                                  bool   special);
-
-    /// Apply chat template. Inspired by hf apply_chat_template() on python.
-    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-    /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
-    /// @param chat Pointer to a list of multiple llama_chat_message
-    /// @param n_msg Number of llama_chat_message in this chat
-    /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
-    /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
-    /// @param length The size of the allocated buffer
-    /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
-    LLAMA_API int32_t llama_chat_apply_template(
-              const struct llama_model * model,
-                            const char * tmpl,
-       const struct llama_chat_message * chat,
-                                size_t   n_msg,
-                                  bool   add_ass,
-                                  char * buf,
-                               int32_t   length);
+                                  int    length);
 
     //
     // Grammar
@@ -914,13 +584,13 @@ extern "C" {
                            float   penalty_present);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
-    /// @param logits Logits extracted from the original generation context.
-    /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
-    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    LLAMA_API void llama_sample_apply_guidance(
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
+    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
-                             float * logits,
-                             float * logits_guidance,
+            llama_token_data_array * candidates,
+              struct llama_context * guidance_ctx,
                              float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
@@ -932,7 +602,7 @@ extern "C" {
     LLAMA_API void llama_sample_top_k(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
-                         int32_t   k,
+                             int   k,
                           size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
@@ -963,19 +633,17 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
-    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
-    LLAMA_API void llama_sample_entropy(
-            struct llama_context * ctx,
-          llama_token_data_array * candidates_p,
-                           float   min_temp,
-                           float   max_temp,
-                           float   exponent_val);
-
     LLAMA_API void llama_sample_temp(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
                            float   temp);
 
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");
+
     /// @details Apply constraints from grammar
     LLAMA_API void llama_sample_grammar(
             struct llama_context * ctx,
@@ -993,7 +661,7 @@ extern "C" {
           llama_token_data_array * candidates,
                            float   tau,
                            float   eta,
-                         int32_t   m,
+                             int   m,
                            float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
@@ -1014,7 +682,7 @@ extern "C" {
             struct llama_context * ctx,
           llama_token_data_array * candidates);
 
-    /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
+    /// @details Randomly selects a token from the candidates based on their probabilities.
     LLAMA_API llama_token llama_sample_token(
             struct llama_context * ctx,
           llama_token_data_array * candidates);
@@ -1066,18 +734,8 @@ extern "C" {
         llama_beam_search_callback_fn_t   callback,
                                    void * callback_data,
                                  size_t   n_beams,
-                                int32_t   n_past,
-                                int32_t   n_predict);
-
-    /// @details Build a split GGUF final path for this chunk.
-    ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
-    //  Returns the split_path length.
-    LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
-
-    /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
-    ///          llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
-    //  Returns the split_prefix length.
-    LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
+                                    int   n_past,
+                                    int   n_predict);
 
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
@@ -1101,49 +759,15 @@ extern "C" {
 // Internal API to be implemented by llama.cpp and used by tests/benchmarks only
 #ifdef LLAMA_API_INTERNAL
 
-#include <random>
-#include <string>
 #include <vector>
+#include <string>
 
 struct ggml_tensor;
 
-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 {
-    const std::vector<std::vector<llama_grammar_element>>   rules;
-    std::vector<std::vector<const llama_grammar_element *>> stacks;
-
-    // buffer for partially generated UTF-8 sequence from accepted tokens
-    llama_partial_utf8                                      partial_utf8;
-};
-
-struct llama_grammar_candidate {
-    size_t               index;
-    const uint32_t     * code_points;
-    llama_partial_utf8   partial_utf8;
-};
-
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
     struct llama_context * ctx
 );
 
-void llama_grammar_accept(
-        const std::vector<std::vector<llama_grammar_element>>         & rules,
-        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
-        const uint32_t                                                  chr,
-        std::vector<std::vector<const llama_grammar_element *>>       & new_stacks);
-
-std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        const std::string & src,
-        llama_partial_utf8   partial_start);
-
-// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
-// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
-llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
-
 #endif // LLAMA_API_INTERNAL
 
 #endif // LLAMA_H

examples/talk-llama/speak

@@ -1,40 +1,24 @@
 #!/bin/bash
 
 # Usage:
-#  speak <voice_id> <textfile>
+#  speak.sh <voice_id> <text-to-speak>
 
-function installed() { command -v $1 >/dev/null 2>&1; }
-
-if installed espeak; then
-  espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 -f $2
-
-elif installed piper && installed aplay; then
-  cat $2 | piper --model ~/en_US-lessac-medium.onnx --output-raw | aplay -q -r 22050 -f S16_LE -t raw -
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
 
 # for Mac
-elif installed say; then
-  say -f $2
+say "$2"
 
 # Eleven Labs
-elif installed python3 && \
-  python3 -c 'import importlib.util; exit(not importlib.util.find_spec("elevenlabs"))' && \
-  installed ffplay; then
-    # It's possible to use the API for free with limited number of characters.
-    # To increase this limit register to https://beta.elevenlabs.io to get an api key
-    # and paste it after 'ELEVEN_API_KEY='
-    # Keep the line commented to use the free version without api key
-    #export ELEVEN_API_KEY=your_api_key
-    wd=$(dirname $0)
-    script=$wd/eleven-labs.py
-    python3 $script -q -p -v $1 $2 >/dev/null 2>&1
-
-    # Uncomment to keep the audio file
-    #python3 $script -q -s ./audio.mp3 -v $1 $2 >/dev/null 2>&1
-    #ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
-
-else
-  echo 'Install espeak ("brew install espeak" or "apt-get install espeak"),'
-  echo 'piper ("pip install piper-tts" or https://github.com/rhasspy/piper) with aplay,'
-  echo 'or elevenlabs ("pip install elevenlabs") with ffplay.'
-  echo '(export ELEVEN_API_KEY if you have an api key from https://beta.elevenlabs.io)'
-fi
+# To use it, install the elevenlabs module from pip (pip install elevenlabs)
+# It's possible to use the API for free with limited number of characters. To increase this limit register to https://beta.elevenlabs.io to get an api key and paste it after 'ELEVEN_API_KEY='
+#Keep the line commented to use the free version whitout api key
+#
+#export ELEVEN_API_KEY=your_api_key
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1

examples/talk-llama/speak.bat

@@ -1 +1 @@
-@powershell -ExecutionPolicy Bypass -F examples\talk-llama\speak.ps1 %1 %2
+@powershell -ExecutionPolicy Bypass -F examples\talk\speak.ps1 %1 %2