release : v1.7.5

ggml-ci

.devops/cublas.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libsdl2-dev wget
+    apt-get install -y build-essential git cmake libsdl2-dev wget git
 
 WORKDIR /app
 

.devops/main-cuda.Dockerfile

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

.devops/main.Dockerfile

@@ -2,7 +2,7 @@ FROM ubuntu:22.04 AS build
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y build-essential wget cmake \
+  apt-get install -y build-essential wget cmake git \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY .. .
@@ -12,7 +12,7 @@ FROM ubuntu:22.04 AS runtime
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg libsdl2-dev wget cmake \
+  apt-get install -y curl ffmpeg libsdl2-dev wget cmake git \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app

.github/workflows/bindings-go.yml

@@ -10,8 +10,8 @@ on:
       - whisper.h
 
 jobs:
-  ubuntu-latest:
-    runs-on: ubuntu-latest
+  ubuntu-22:
+    runs-on: ubuntu-22.04
     steps:
       - uses: actions/setup-go@v5
         with:

.github/workflows/bindings-ruby.yml

@@ -19,7 +19,12 @@ on:
       - ggml/**/*.m
       - ggml/**/*.metal
       - scripts/get-flags.mk
-      - examples/dr_wav.h
+      - examples/common.h
+      - examples/common.cpp
+      - examples/common-whisper.h
+      - examples/common-whisper.cpp
+      - examples/stb_vorbis.c
+      - examples/miniaudio.h
   pull_request:
     paths:
       - bindings/ruby/**
@@ -39,11 +44,16 @@ on:
       - ggml/**/*.m
       - ggml/**/*.metal
       - scripts/get-flags.mk
-      - examples/dr_wav.h
+      - examples/common.h
+      - examples/common.cpp
+      - examples/common-whisper.h
+      - examples/common-whisper.cpp
+      - examples/stb_vorbis.c
+      - examples/miniaudio.h
 
 jobs:
-  ubuntu-latest:
-    runs-on: ubuntu-latest
+  ubuntu-22:
+    runs-on: ubuntu-22.04
     defaults:
       run:
         working-directory: bindings/ruby

.github/workflows/build.yml

@@ -6,18 +6,82 @@ on:
       - master
   pull_request:
     types: [opened, synchronize, reopened]
+  workflow_dispatch:
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+      pre_release_tag:
+        description: 'Pre-release tag name'
+        required: false
+        type: string
+      run_type:
+        description: 'Workflow type to run'
+        required: true
+        type: choice
+        options:
+          - full-ci
+          - release-only
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+permissions:
+  contents: write  # for creating release
+
 env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   ubuntu_image: "ubuntu:22.04"
   VCPKG_BINARY_SOURCES: "clear;x-gha,readwrite"
 
 jobs:
-  ubuntu-latest:
+  determine-tag:
     runs-on: ubuntu-latest
+    outputs:
+      tag_name: ${{ steps.tag.outputs.name }}
+
+    steps:
+      - name: Checkout with full history
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER=$(git rev-list --count HEAD)
+          SHORT_HASH=$(git rev-parse --short=7 HEAD)
+          CUSTOM_TAG="${{ github.event.inputs.pre_release_tag }}"
+
+          echo "Raw values:"
+          echo "BUILD_NUMBER: $BUILD_NUMBER"
+          echo "SHORT_HASH: $SHORT_HASH"
+          echo "BRANCH_NAME: ${{ env.BRANCH_NAME }}"
+          echo "CUSTOM_TAG: $CUSTOM_TAG"
+
+          # Use custom tag if provided
+          if [[ -n "$CUSTOM_TAG" ]]; then
+            echo "Using custom tag"
+            TAG_NAME="${CUSTOM_TAG}"
+          elif [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "Using master branch format"
+            TAG_NAME="b${BUILD_NUMBER}"
+          else
+            echo "Using non-master branch format"
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            TAG_NAME="${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}"
+          fi
+
+          echo "Final tag name: $TAG_NAME"
+          echo "name=$TAG_NAME" >> $GITHUB_OUTPUT
+
+  ubuntu-22:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -38,12 +102,14 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential libsdl2-dev cmake
+            apt install -y build-essential libsdl2-dev cmake git
             cmake -B build
             cmake --build build --config Release -j $(nproc)'
 
-  ubuntu-latest-arm64:
-    runs-on: ubuntu-latest
+  ubuntu-22-arm64:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -64,12 +130,14 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential libsdl2-dev cmake
+            apt install -y build-essential libsdl2-dev cmake git
             cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
             cmake --build build --config Release -j $(nproc)'
 
-  ubuntu-latest-arm-v7:
-    runs-on: ubuntu-latest
+  ubuntu-22-arm-v7:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -90,17 +158,30 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential libsdl2-dev cmake
+            apt install -y build-essential libsdl2-dev cmake git
             cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
             cmake --build build --config Release -j $(nproc)'
 
   macOS-latest:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: macOS-latest
 
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
     steps:
       - name: Clone
+        id: checkout
         uses: actions/checkout@v4
 
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: macOS-latest-swift
+          evict-old-files: 1d
+
       - name: Dependencies
         run: |
           brew update
@@ -108,29 +189,39 @@ jobs:
 
       - name: Build
         run: |
-          cmake -B build
-          cmake --build build --config Release
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DWHISPER_BUILD_EXAMPLES=OFF \
+            -DWHISPER_BUILD_TESTS=OFF \
+            -DWHISPER_BUILD_SERVER=OFF \
+            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-#  freeBSD-latest:
-#    runs-on: macos-12
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Build
-#        uses: cross-platform-actions/action@v0.24.0
-#        with:
-#          operating_system: freebsd
-#          version: '13.3'
-#          run: |
-#            sudo pkg update
-#            sudo pkg install -y gmake sdl2 cmake
-#            cmake -B build
-#            cmake --build build --config Release
 
-  ubuntu-latest-gcc:
-    runs-on: ubuntu-latest
+  freeBSD-latest:
+    runs-on: macos-13
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Build
+        uses: cross-platform-actions/action@v0.27.0
+        with:
+          operating_system: freebsd
+          version: '14.2'
+          run: |
+            sudo pkg update
+            sudo pkg install -y gmake sdl2 cmake git
+            cmake -B build
+            cmake --build build --config Release
+
+  ubuntu-22-gcc:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -152,13 +243,15 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential cmake libsdl2-dev
+            apt install -y build-essential cmake libsdl2-dev git
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
             make
             ctest -L gh --output-on-failure'
 
-  ubuntu-latest-gcc-arm64:
-    runs-on: ubuntu-latest
+  ubuntu-22-gcc-arm64:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -180,13 +273,15 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential cmake libsdl2-dev
+            apt install -y build-essential cmake libsdl2-dev git
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
             make
             ctest -L gh --output-on-failure'
 
-  ubuntu-latest-gcc-arm-v7:
-    runs-on: ubuntu-latest
+  ubuntu-22-gcc-arm-v7:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -208,13 +303,15 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential cmake libsdl2-dev
+            apt install -y build-essential cmake libsdl2-dev git
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
             make
             ctest -L gh --output-on-failure'
 
-  ubuntu-latest-clang:
-    runs-on: ubuntu-latest
+  ubuntu-22-clang:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -239,13 +336,15 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y clang build-essential cmake libsdl2-dev
+            apt install -y clang build-essential cmake libsdl2-dev git
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
             make
             ctest -L gh --output-on-failure'
 
-  ubuntu-latest-gcc-sanitized:
-    runs-on: ubuntu-latest
+  ubuntu-22-gcc-sanitized:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       fail-fast: false
@@ -267,12 +366,16 @@ jobs:
             -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
+            apt install -y build-essential cmake git
+            cmake . -DCMAKE_BUILD_TYPE=Debug \
+              -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON \
+              -DGGML_OPENMP=OFF
             make
             ctest -L gh --output-on-failure'
 
   ubuntu-22-cmake-sycl:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: ubuntu-22.04
 
     strategy:
@@ -302,12 +405,12 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp git
 
       - name: install oneAPI MKL library
         shell: bash
         run: |
-          sudo apt install intel-oneapi-mkl-devel
+          sudo apt install intel-oneapi-mkl-devel git
 
       - name: Clone
         id: checkout
@@ -323,6 +426,8 @@ jobs:
           cmake --build . --config Release -j $(nproc)
 
   ubuntu-22-cmake-sycl-fp16:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: ubuntu-22.04
 
     strategy:
@@ -352,7 +457,7 @@ jobs:
         shell: bash
         run: |
           sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp git
 
       - name: install oneAPI MKL library
         shell: bash
@@ -373,6 +478,8 @@ jobs:
           cmake --build . --config Release -j $(nproc)
 
   windows-msys2:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-latest
 
     strategy:
@@ -393,6 +500,7 @@ jobs:
           msystem: ${{matrix.sys}}
           install: >-
             base-devel
+            git
             mingw-w64-${{matrix.env}}-toolchain
             mingw-w64-${{matrix.env}}-cmake
             mingw-w64-${{matrix.env}}-SDL2
@@ -416,6 +524,8 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-latest
 
     strategy:
@@ -476,6 +586,8 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   windows-blas:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-latest
 
     strategy:
@@ -549,6 +661,8 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   windows-cublas:
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
     runs-on: windows-2019
     strategy:
       matrix:
@@ -565,15 +679,134 @@ jobs:
       - name: Clone repository
         uses: actions/checkout@v4
 
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ${{ github.job }}-${{ matrix.cuda-toolkit }}-${{ matrix.build }}
+          variant: sccache
+          evict-old-files: 5d
+
+      - name: Install Cuda Toolkit 11.8.0
+        if: ${{ matrix.cuda-toolkit == '11.8.0' }}
+        run: |
+          $CUDA_VERSION = ${{ matrix.cuda-toolkit }}
+          $CUDA_TOOLKIT_DIR = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v$CUDA_VERSION"
+          $CUDA_DOWNLOAD = "https://developer.download.nvidia.com/compute/cuda/redist"
+
+          # Components versions
+          $CUDART_VER = "11.8.89"
+          $NVCC_VER   = "11.8.89"
+          $NVRTC_VER  = "11.8.89"
+          $CUBLAS_VER = "11.8.1.74"
+          $NVTX_VER   = "11.8.86"
+          $VS_VER     = "11.8.86"
+          $NVPROF_VER = "11.8.87"
+          $CCCL_VER   = "11.8.89"
+
+          # Create the directory where the CUDA Toolkit will be installed
+          mkdir -p $CUDA_TOOLKIT_DIR
+
+          # Install unzip to extract the downloaded files
+          choco install unzip -y
+
+          # Download all the required components
+          curl -O "$CUDA_DOWNLOAD/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-${CUDART_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-${NVCC_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-${NVRTC_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/libcublas/windows-x86_64/libcublas-windows-x86_64-${CUBLAS_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-${NVTX_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-${VS_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-${NVPROF_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-${CCCL_VER}-archive.zip"
+
+          # Extract all the downloaded files to the CUDA Toolkit directory
+          unzip '*.zip' -d $CUDA_TOOLKIT_DIR
+
+          # Copy all the extracted files to the main CUDA Toolkit directory
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_cudart-windows-x86_64-${CUDART_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvcc-windows-x86_64-${NVCC_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvrtc-windows-x86_64-${NVRTC_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\libcublas-windows-x86_64-${CUBLAS_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvtx-windows-x86_64-${NVTX_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvprof-windows-x86_64-${NVPROF_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_cccl-windows-x86_64-${CCCL_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\visual_studio_integration-windows-x86_64-${VS_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+
+          # Visual Studio integration
+          xcopy "$CUDA_TOOLKIT_DIR\visual_studio_integration-windows-x86_64-${VS_VER}-archive\visual_studio_integration\MSBuildExtensions\*" "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\MSBuild\Microsoft\VC\v160\BuildCustomizations" /E /I /H /Y
+
+          # Set environment variables
+          echo "$CUDA_TOOLKIT_DIR\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "$CUDA_TOOLKIT_DIR\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=$CUDA_TOOLKIT_DIR" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V11_8=$CUDA_TOOLKIT_DIR" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+        
+      - name: Install Cuda Toolkit 12.2.0
+        if: ${{ matrix.cuda-toolkit == '12.2.0' }}
+        run: |
+          $CUDA_VERSION = ${{ matrix.cuda-toolkit }}
+          $CUDA_TOOLKIT_DIR = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v$CUDA_VERSION"
+          $CUDA_DOWNLOAD = "https://developer.download.nvidia.com/compute/cuda/redist"
+
+          # Components versions
+          $CUDART_VER   = "12.2.140"
+          $NVCC_VER     = "12.2.140"
+          $NVRTC_VER    = "12.2.140"
+          $CUBLAS_VER   = "12.2.5.6"
+          $NVTX_VER     = "12.2.140"
+          $PROFILER_VER = "12.2.140"
+          $VS_VER       = "12.2.140"
+          $NVPROF_VER   = "12.2.142"
+          $CCCL_VER     = "12.2.140"
+
+          # Create the directory where the CUDA Toolkit will be installed
+          mkdir -p $CUDA_TOOLKIT_DIR
+
+          # Install unzip to extract the downloaded files
+          choco install unzip -y
+
+          # Download all the required components
+          curl -O "$CUDA_DOWNLOAD/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-${CUDART_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-${NVCC_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-${NVRTC_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/libcublas/windows-x86_64/libcublas-windows-x86_64-${CUBLAS_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-${NVTX_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-${PROFILER_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-${VS_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-${NVPROF_VER}-archive.zip"
+          curl -O "$CUDA_DOWNLOAD/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-${CCCL_VER}-archive.zip"
+
+          # Extract all the downloaded files to the CUDA Toolkit directory
+          unzip -q '*.zip' -d $CUDA_TOOLKIT_DIR
+
+          # Copy all the extracted files to the main CUDA Toolkit directory
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_cudart-windows-x86_64-${CUDART_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvcc-windows-x86_64-${NVCC_VER}-archive\*"     "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvrtc-windows-x86_64-${NVRTC_VER}-archive\*"   "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\libcublas-windows-x86_64-${CUBLAS_VER}-archive\*"   "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvtx-windows-x86_64-${NVTX_VER}-archive\*"     "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_nvprof-windows-x86_64-${NVPROF_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_cccl-windows-x86_64-${CCCL_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\cuda_profiler_api-windows-x86_64-${PROFILER_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+          xcopy "$CUDA_TOOLKIT_DIR\visual_studio_integration-windows-x86_64-${VS_VER}-archive\*" "$CUDA_TOOLKIT_DIR" /E /I /H /Y
+
+          # Visual Studio integration
+          xcopy "$CUDA_TOOLKIT_DIR\visual_studio_integration-windows-x86_64-${VS_VER}-archive\visual_studio_integration\MSBuildExtensions\*" "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\MSBuild\Microsoft\VC\v160\BuildCustomizations" /E /I /H /Y
+
+          # Set environment variables
+          echo "$CUDA_TOOLKIT_DIR\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "$CUDA_TOOLKIT_DIR\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=$CUDA_TOOLKIT_DIR" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V12_2=$CUDA_TOOLKIT_DIR" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v2
 
-      - name: Install CUDA Toolkit
-        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
-        with:
-          cuda: '${{ matrix.cuda-toolkit }}'
-
       - name: Install 7-Zip
         run: choco install 7zip -y
 
@@ -584,26 +817,31 @@ jobs:
           7z x sdl2.zip
           echo "SDL2_DIR=${{ github.workspace }}\SDL2-${{ matrix.sdl2_ver }}\cmake" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "${{ github.workspace }}\SDL2-${{ matrix.sdl2_ver }}\cmake" > SDL2_PATH.txt
- 
-      - name: Configure CMake
-        shell: cmd
-        run: |
-          cmake -S . -B ./build -A ${{ matrix.arch }} ^
-            -DCMAKE_BUILD_TYPE=${{ matrix.build }} ^
-            -DGGML_CUDA=${{ matrix.cublas }} ^
-            -DCMAKE_CUDA_ARCHITECTURES=all  ^
-            -DWHISPER_SDL2=${{ matrix.sdl2 }} ^
-            -DSDL2_DIR="%SDL2_DIR%"
-  
+
+      - name: Install cmake
+        run: choco install cmake
+
       - name: Build Project
         shell: cmd
         run: |
-          cd ./build
-          cmake --build . --config ${{ matrix.build }}          
+          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          cmake --version
+          where cmake
+          cmake -S . -B build -G "Ninja Multi-Config" ^
+            -DCMAKE_BUILD_TYPE=${{ matrix.build }} ^
+            -DGGML_CUDA=${{ matrix.cublas }} ^
+            -DWHISPER_SDL2=${{ matrix.sdl2 }} ^
+            -DSDL2_DIR="%SDL2_DIR%"
+          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
+          cmake --build build --config ${{ matrix.build }} -j %NUMBER_OF_PROCESSORS%
+
+      - name: Check sccache status after build
+        run: |
+          sccache --show-stats
 
       - name: Copy CUDA DLLs
         run: |
-          Get-ChildItem "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/" -Filter "*.dll" | 
+          Get-ChildItem "$env:CUDA_PATH\bin\" -Filter "*.dll" |
           Copy-Item -Destination "build/bin/${{ matrix.build }}"
 
       - name: Copy SDL2.dll
@@ -617,7 +855,9 @@ jobs:
           path: build/bin/${{ matrix.build }}
 
   emscripten:
-    runs-on: ubuntu-latest
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     strategy:
       matrix:
@@ -640,6 +880,7 @@ jobs:
 
   ios-xcode-build:
     runs-on: macos-latest
+    needs: determine-tag
 
     strategy:
       matrix:
@@ -670,21 +911,39 @@ jobs:
             -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-          sudo cmake --install . --config Release
 
       - name: xcodebuild for swift package
         id: xcodebuild
         run: |
-          xcodebuild -scheme whisper-Package -destination 'generic/platform=iOS'
+          ./build-xcframework.sh
 
-#- name: Build objc example
-#  run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphoneos build
+      - name: Build objc example
+        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO FRAMEWORK_FOLDER_PATH=./build-ios build
 
       - name: Build swiftui example
-        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
+        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
+                github.event.inputs.create_release == 'true' ||
+                github.event.inputs.pre_release_tag != '' }}
+        run: |
+          zip --symlinks -r whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip build-apple/whisper.xcframework
+
+      - name: Upload artifacts
+        if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
+                github.event.inputs.create_release == 'true' ||
+                github.event.inputs.pre_release_tag != '' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework.zip
+          name: whisper-${{ needs.determine-tag.outputs.tag_name }}-xcframework
 
   android:
-    runs-on: ubuntu-latest
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     steps:
       - name: Clone
@@ -712,31 +971,30 @@ jobs:
           cd whisper/examples/whisper.android
           ./gradlew assembleRelease --no-daemon
 
-# TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602
-#  android_java:
-#    runs-on: ubuntu-latest
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: set up JDK 11
-#        uses: actions/setup-java@v4
-#        with:
-#          java-version: '11'
-#          distribution: 'temurin'
-#          cache: gradle
-#
-#      - name: Setup Android SDK
-#        uses: android-actions/setup-android@v3
-#        with:
-#          cmdline-tools-version: 9.0
-#
-#      - name: Build
-#        run: |
-#          cd examples/whisper.android.java
-#          chmod +x ./gradlew
-#          ./gradlew assembleRelease
+  android_java:
+    runs-on: ubuntu-22.04
+
+    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
 
 # TODO: disabled because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/9686220096/job/26735899598
 #  java:
@@ -783,7 +1041,9 @@ jobs:
 #          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
 
   quantize:
-    runs-on: ubuntu-latest
+    if: ${{ github.event_name == 'push' || github.event_name == 'pull_request' ||
+            github.event.inputs.run_type == 'full-ci' }}
+    runs-on: ubuntu-22.04
 
     steps:
       - name: Clone
@@ -795,3 +1055,95 @@ jobs:
           cmake -B build
           cmake --build build --config Release
           ./build/bin/quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0
+
+  release:
+    if: ${{ github.event.inputs.create_release == 'true' || github.event.inputs.pre_release_tag != '' }}
+
+    runs-on: ubuntu-latest
+
+    needs:
+      - determine-tag
+      - ios-xcode-build
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: release
+          evict-old-files: 1d
+
+      # Downloads all the artifacts from the previous jobs
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+
+      - name: Create release
+        id: create_release
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ needs.determine-tag.outputs.tag_name }}
+          prerelease: ${{ github.event.inputs.pre_release_tag != '' }}
+
+      - name: Upload release
+        id: upload_release
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact/release')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                });
+              }
+            }
+
+  coreml-base-en:
+    if: ${{ (github.event_name == 'push' && github.ref == 'refs/heads/master') ||
+            github.event.inputs.create_release == 'true' ||
+            github.event.inputs.pre_release_tag != '' }}
+    runs-on: macos-latest
+    needs: determine-tag
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set environment variables
+        id: set_vars
+        run: |
+          echo "MODEL_NAME=base.en" >> $GITHUB_ENV
+          echo "GEN_MODEL_NAME=whisper-${{ needs.determine-tag.outputs.tag_name }}-ggml-base.en-encoder.mlmodelc" >> $GITHUB_ENV
+
+      - name: Download model
+        run: |
+          ./models/download-ggml-model.sh ${{ env.MODEL_NAME }}
+
+      - name: Generate CoreML model
+        run: |
+          python3.11 -m venv venv
+          source venv/bin/activate
+          pip install ane_transformers openai-whisper coremltools
+          ./models/generate-coreml-model.sh ${{ env.MODEL_NAME }}

.github/workflows/docker.yml

@@ -11,7 +11,7 @@ jobs:
     name: Push Docker image to Docker Hub
     if: github.event.pull_request.draft == false
 
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-22.04
     env:
       COMMIT_SHA: ${{ github.sha }}
     strategy:
@@ -28,6 +28,8 @@ jobs:
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v3
+        with:
+          image: tonistiigi/binfmt:qemu-v7.0.0-28
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3

.github/workflows/examples-wasm.yml

@@ -0,0 +1,91 @@
+name: Examples WASM
+on:
+  push:
+    branches: ["master"]
+
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+concurrency:
+  group: "pages"
+  cancel-in-progress: false
+
+jobs:
+  deploy-wasm-github-pages:
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Pages
+        uses: actions/configure-pages@v4
+
+      - name: Setup emsdk
+        uses: mymindstorm/setup-emsdk@v14
+
+      - name: Build WASM Examples
+        # Enable for real build later in whisper.cpp
+        run: |
+          mkdir -p build-em && cd build-em
+          emcmake cmake .. -DCMAKE_BUILD_TYPE=Release
+          make -j
+
+      - name: Create staging directory
+        run: mkdir -p staging
+
+      - name: Create .nojekyll file in staging directory
+        run: touch staging/.nojekyll
+
+      - name: Copy application files
+        run: |
+          build_dir=build-em/bin
+
+          ls ${build_dir}
+
+          # command.wasm
+          target_dir=staging/command.wasm
+          mkdir -p ${target_dir}
+          cp ${build_dir}/command.wasm/{index.html,command.js,helpers.js} ${target_dir}
+          cp ${build_dir}/libcommand.js ${target_dir}
+
+          # bench.wasm
+          target_dir=staging/bench.wasm
+          mkdir -p ${target_dir}
+          cp ${build_dir}/bench.wasm/{index.html,bench.js,helpers.js} ${target_dir}
+          cp ${build_dir}/libbench.js ${target_dir}
+
+          # stream.wasm
+          target_dir=staging/stream.wasm
+          mkdir -p ${target_dir}
+          cp ${build_dir}/stream.wasm/{index.html,stream.js,helpers.js} ${target_dir}
+          cp ${build_dir}/libstream.js ${target_dir}
+
+          # whisper.wasm (this will be the main example page)
+          target_dir=staging
+          mkdir -p ${target_dir}
+          cp ${build_dir}/whisper.wasm/{index.html,main.js,helpers.js} ${target_dir}
+          cp ${build_dir}/libmain.js ${target_dir}
+
+          # Copy Cross-Origin Isolation service worker
+          cp -v examples/coi-serviceworker.js staging/
+
+      - name: List files in staging directory (for debugging)
+        run: |
+          echo "Files in staging directory:"
+          find staging -type f | sort
+
+      - name: Upload artifact
+        uses: actions/upload-pages-artifact@v3
+        with:
+          path: ./staging
+
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4

.github/workflows/examples.yml

@@ -10,8 +10,8 @@ on:
       - whisper.h
 
 jobs:
-  addon_node-ubuntu-latest:
-    runs-on: ubuntu-latest
+  addon_node-ubuntu-22:
+    runs-on: ubuntu-22.04
     strategy:
       matrix:
         node-version: [ 16.x, 18.x ]
@@ -22,7 +22,7 @@ jobs:
       - name: Dependencies
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential git
           sudo apt-get install cmake
           sudo apt-get install libsdl2-dev
 

.gitignore

@@ -58,3 +58,5 @@ cmake-build-debug/
 .cxx/
 .gradle/
 local.properties
+.log
+.exe

AUTHORS

@@ -1,34 +1,51 @@
-# date: Tue Apr  9 20:27:03 EEST 2024
+# date: Tue Feb  4 13:03:35 EET 2025
 # 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>
+65a <10104049+65a@users.noreply.github.com>
+AIWintermuteAI <32562299+AIWintermuteAI@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>
+Adam Jones <domdomegg+git@gmail.com>
+Adrien Gallouët <adrien@gallouet.fr>
+Adrien Gallouët <angt@huggingface.co>
 AfryMask <AfryMask@163.com>
 Ahmad Bilal <ahmad.bilal@empglabs.com>
+Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com>
 AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
+AidanBeltonS <aidan.belton@codeplay.com>
+Akarshan Biswas <akarshan.biswas@gmail.com>
+Akarshan Biswas <akarshanbiswas@fedoraproject.org>
 Akash Mahajan <akash7190@gmail.com>
 Akash Mahajan <akashmjn@stanford.edu>
 Al Hoang <3811822-hoanga@users.noreply.gitlab.com>
 Alan <unknown>
+Albert Jin <albert.jin@gmail.com>
+Alberto Cabrera Pérez <alberto.cabrera@codeplay.com>
+Alberto Cabrera Pérez <alberto.cabrera@intel.com>
 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>
+Alex O'Connell <35843486+acon96@users.noreply.github.com>
 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>
+Alter <0x7c48@gmail.com>
 Ananta Bastola <anantarajbastola@gmail.com>
+Andreas Kieslinger <47689530+aendk@users.noreply.github.com>
+Andreas Lubbe <git@lubbe.org>
 Andreu Huguet <andreuhuguet@gmail.com>
 Andrew Huynh <a5thuynh@gmail.com>
+Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com>
 Andrew S <andrews54757@gmail.com>
 Andy Maloney <asmaloney@gmail.com>
 Anton Kostin <masguit42@users.noreply.github.com>
@@ -40,8 +57,11 @@ AustinMroz <austinmroz@utexas.edu>
 Avik Sengupta <avik@sengupta.net>
 Bader-eddine Ouaich <49657842+baderouaich@users.noreply.github.com>
 Baffin Lee <baffinlee@gmail.com>
+Ben Ashbaugh <ben.ashbaugh@intel.com>
 Ben Nortier <bjnortier@gmail.com>
 Benjamin Heiniger <benjamin.heiniger@bluewin.ch>
+Bernhard M. Wiedemann <githubbmwprimary@lsmod.de>
+Binozo <70137898+Binozo@users.noreply.github.com>
 Bo-Yi Wu <appleboy.tw@gmail.com>
 Boris Bliznioukov <blib@mail.com>
 Borislav Stanimirov <b.stanimirov@abv.bg>
@@ -49,47 +69,86 @@ 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>
+Carlos Zoido <mrgalleta@gmail.com>
 Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
+CarterLi999 <664681047@qq.com>
 ChangSeok Oh <shivamidow@users.noreply.github.com>
+Changyeon Kim <cyzero.kim@samsung.com>
 Chaoqun <27287694+OpenWaygate@users.noreply.github.com>
+Charles Xu <63788048+chaxu01@users.noreply.github.com>
+Charles Xu <charles.xu@arm.com>
+Chen Xi <xi2.chen@intel.com>
+Chen Xi <xixichen08@foxmail.com>
+Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
 Chia-Hsiang Cheng <88014292+garychia@users.noreply.github.com>
 Chidi Williams <williamschidi1@gmail.com>
+Chris Elrod <elrodc@gmail.com>
 Christian <12550267+iceychris@users.noreply.github.com>
+Christian Kastner <ckk@kvr.at>
 Clifford Heath <clifford.heath@gmail.com>
+Clint Herron <hanclinto@gmail.com>
 Colin <github@whoisc.cc>
+Conrad Kramer <conrad@conradkramer.com>
+Corey Earwood <iamcgn+github@gmail.com>
+CrispStrobe <154636388+CrispStrobe@users.noreply.github.com>
+DAN™ <dranger003@gmail.com>
 DGdev91 <DGdev91@users.noreply.github.com>
 Damian Czaja <trojan295@protonmail.com>
+Dan Johansson <164997844+eddnjjn@users.noreply.github.com>
+Dan Johansson <dan.johansson@arm.com>
 Daniel Bevenius <daniel.bevenius@gmail.com>
+Daniel Valdivia <18384552+dvaldivia@users.noreply.github.com>
+Daniel Ziegenberg <daniel@ziegenberg.at>
+Daniele <57776841+daniandtheweb@users.noreply.github.com>
+Dave <dave-fl@users.noreply.github.com>
+Dave Airlie <airlied@gmail.com>
+Dave Airlie <airlied@redhat.com>
+Daven Sanassy <daven@vochlea.co.uk>
 David <dnhkng@gmail.com>
 David Thorpe <djt@mutablelogic.com>
+DavidKorczynski <david@adalogics.com>
 Davidson Francis <davidsondfgl@gmail.com>
 Dener Stassun <denerstassun@gmail.com>
+Dibakar Gope <dibakar.gope@arm.com>
 Didzis Gosko <didzis@users.noreply.github.com>
+Diego Devesa <slarengh@gmail.com>
 Digipom <admin@digipom.com>
 Dimo <dimo@ieee.org>
+Djip007 <3705339+Djip007@users.noreply.github.com>
+Djip007 <djip.perois@free.fr>
 Dody Suria Wijaya <dodysw@gmail.com>
+Dou Xinpeng <15529241576@163.com>
+Dou Xinpeng <81913537+Dou-Git@users.noreply.github.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 Curtin <ericcurtin17@gmail.com>
 Eric Swanson <eswanson@alloscomp.com>
 Eric Tendian <erictendian@gmail.com>
+Eric Zhang <34133756+EZForever@users.noreply.github.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>
+Faisal Zaghloul <quic_fzaghlou@quicinc.com>
 Fangjun Kuang <csukuangfj@gmail.com>
 Felix <stenbackfelix@gmail.com>
 Finn Voorhees <finnvoorhees@gmail.com>
+FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
 FlippFuzz <41221030+FlippFuzz@users.noreply.github.com>
+Frankie Robertson <frankier@users.noreply.github.com>
 Gang Chen <goncha@gmail.com>
 Gavin Cai <gavin1818@hotmail.com>
 George Hindle <george@georgehindle.com>
 Georgi Gerganov <ggerganov@gmail.com>
+Gilad S <7817232+giladgd@users.noreply.github.com>
+Gilad S <giladgd@users.noreply.github.com>
+Gilad S. <7817232+giladgd@users.noreply.github.com>
 GitAritron <103900385+GitAritron@users.noreply.github.com>
 GiviMAD <GiviMAD@users.noreply.github.com>
 Gleicon Moraes <gleicon@gmail.com>
@@ -98,41 +157,66 @@ 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>
+Haus1 <haus.xda@gmail.com>
 Herman Semenov <GermanAizek@yandex.ru>
+HimariO <dsfhe49854@gmail.com>
+Hong Bo PENG <penghb@cn.ibm.com>
 Hrishikesh Barman <geekodour@users.noreply.github.com>
+Hugo <hugo@whynothugo.nl>
 Ian Bicking <ian@ianbicking.org>
 Ian Bull <irbull@eclipsesource.com>
+Ihar Hrachyshka <ihrachys@redhat.com>
 Ikko Ashimine <eltociear@gmail.com>
+Ikko Eltociear Ashimine <eltociear@gmail.com>
 InconsolableCellist <23345188+InconsolableCellist@users.noreply.github.com>
 Ismatulla Mansurov <47342870+sapoepsilon@users.noreply.github.com>
+Ivan <nekotekina@gmail.com>
+Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
 Ivan Gorin <ivangorin21@gmail.com>
+Ivo von Putzer Reibegg <ivo.putzer@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>
+Jayant <jayantyadav202@gmail.com>
+Jeff Bolz <jbolz@nvidia.com>
+Jeroen Mostert <jeroen.mostert@cm.com>
 Jhen-Jie Hong <developer@jhen.me>
 Jhen-Jie Hong <iainst0409@gmail.com>
 JidongZhang-THU <1119708529@qq.com>
 Jo Liss <joliss42@gmail.com>
+Joe Todd <joe.todd@codeplay.com>
 Johan <jr.raffin@gmail.com>
 Johannes Gäßler <johannesg@5d6.de>
 John Balis <phobossystems@gmail.com>
+JohnnyB <jboero@users.noreply.github.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>
+Josscii <jossciiweiyi@gmail.com>
 Judd <foldl@users.noreply.github.com>
 Jumper775 <78500318+jumpers775@users.noreply.github.com>
+Jun Hee Yoo <contact.jhyoo@gmail.com>
+Junil Kim <logyourself@gmail.com>
+Justina Cho <justcho5@gmail.com>
 Justine Tunney <jtunney@gmail.com>
+Justine Tunney <jtunney@mozilla.com>
+KITAITI Makoto <KitaitiMakoto@gmail.com>
 KP Kaiser <kirk@zothcorp.com>
 Kamilake <exjang0@gmail.com>
+Karol Kontny <82021046+kkontny@users.noreply.github.com>
+Karthick <j.karthic2004@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>
+Kendrick Taylor <kendrick@circuitsix.com>
 Kevin Brothaler <admin@digipom.com>
+Kevin Gibbons <bakkot@gmail.com>
+Konosuke Sakai <konosuke@konosuke.work>
 Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
 Kreijstal <rainb@tfwno.gf>
 Kylin <56434533+KyL0N@users.noreply.github.com>
@@ -147,56 +231,110 @@ 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>
+Ma Mingfei <mingfei.ma@intel.com>
 Maciek <maciek.mab122@gmail.com>
+Mahesh Madhav <67384846+heshpdx@users.noreply.github.com>
 Marcin Mielniczuk <marmistrz.dev@zoho.eu>
+Mark Karpelès <MagicalTux@users.noreply.github.com>
+Mark Zhuang <zhuangqiubin@gmail.com>
+Markus Tavenrath <mtavenrath@users.noreply.github.com>
+Martin Delille <martin@delille.org>
 Martin Warnaar <martinwarnaar@gmail.com>
+Masaya, Kato <62578291+msy-kato@users.noreply.github.com>
 Matheus de Sousa <23645013+keyehzy@users.noreply.github.com>
+Mathieu Baudier <mbaudier@argeo.org>
 Mathijs de Bruin <mathijs@mathijsfietst.nl>
 Matija Pevec <mightymatth@users.noreply.github.com>
+Matt Stephenson <mstephenson6@users.noreply.github.com>
+Max Krasnyansky <max.krasnyansky@gmail.com>
+Max Krasnyansky <quic_maxk@quicinc.com>
 Maximiliano Levi <8160966+maxilevi@users.noreply.github.com>
 Meng, Hengyu <hengyu.meng@intel.com>
+Mengqing Cao <cmq0113@163.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>
+Molly Sophia <mollysophia379@gmail.com>
 Murilo Santana <mvrilo@gmail.com>
+NETZkultur GmbH <mulholland@netzkultur.de>
+Natsu <chino@hotococoa.moe>
 Neil Chudleigh <nchudleigh@users.noreply.github.com>
+Neo Zhang <14088817+arthw@users.noreply.github.com>
 Neo Zhang Jianyu <jianyu.zhang@intel.com>
 Neuman Vong <neuman.vong@gmail.com>
+Nicholai Tukanov <nicholaitukanov@gmail.com>
 Nicholas Albion <nalbion@yahoo.com>
+Nico Bosshard <nico@bosshome.ch>
+Nicolò Scipione <nicolo.scipione@codeplay.com>
 Niels Mayer <Niels.Mayer@gmail.com>
+Nikita Sarychev <42014488+sARY77@users.noreply.github.com>
+Nikolaj Olsson <nikse.dk@gmail.com>
 Okabintaro <103938900+Okabintaro@users.noreply.github.com>
 Oleg Sidorov <me@whitebox.io>
 Oleg Sidorov <oleg@sidorov.nl>
+Olivier Chafik <ochafik@users.noreply.github.com>
 Ondrej Kokes <ondrej.kokes@gmail.com>
 Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
+PAB <pierreantoine.bannier@gmail.com>
 Paul Tsochantaris <ptsochantaris@icloud.com>
+Pedro Probst <pprobst@insiberia.net>
+Peng <hzp1024@qq.com>
+Peter <peter277@users.noreply.github.com>
 Philipp Zabel <philipp.zabel@gmail.com>
 Philippe Normand <phil@base-art.net>
+Philippe Normand <philn@igalia.com>
+Plamen Minev <pacominev@gmail.com>
+Prashant Vithule <119530321+Vithulep@users.noreply.github.com>
 Przemysław Pawełczyk <przemoc@gmail.com>
 Qianhe Chen <54462604+chenqianhe@users.noreply.github.com>
+R0CKSTAR <xiaodong.ye@mthreads.com>
+R0CKSTAR <yeahdongcn@gmail.com>
+Radoslav Gerganov <rgerganov@gmail.com>
 Radosław Gryta <radek.gryta@gmail.com>
+Rahul Vadhyar <107788610+RahulVadhyar@users.noreply.github.com>
+Raiya Araki <83504221+rai62@users.noreply.github.com>
 Reinforce-II <fate@eastal.com>
 Reinis Muiznieks <muiznieks.reinis@gmail.com>
 RelatedTitle <r3latedtitle@gmail.com>
+Rémy Oudompheng <oudomphe@phare.normalesup.org>
 RhinoDevel <RhinoDevel@users.noreply.github.com>
 Rich Jones <miserlou@gmail.com>
+Robert Ormandi <52251610+ormandi@users.noreply.github.com>
 Robin <robin.xw@hotmail.com>
 Roddur Dasgupta <roddurd@gmail.com>
 Roland Rabien <figbug@gmail.com>
+Romain Biessy <romain.biessy@codeplay.com>
+Ronsor <ronsor@ronsor.pw>
 Rotem Dan <rotemdan@gmail.com>
 Ryan Hitchman <hitchmanr@gmail.com>
 Ryan Metcalfe <107415876+RyanMetcalfeInt8@users.noreply.github.com>
 RyanChang <ftes90015@gmail.com>
+SRHMorris <69468379+SRHMorris@users.noreply.github.com>
+SXX <sxx1136965276@gmail.com>
+Sacha Arbonel <sacha.arbonel@hotmail.fr>
+Salman Faroz <stsfaroz@gmail.com>
+Salvatore Mesoraca <s.mesoraca16@gmail.com>
 Sam <49637763+Onlyartist9@users.noreply.github.com>
 Sam Pullara <spullara@gmail.com>
+Samuel Durante <44513615+samueldurantes@users.noreply.github.com>
 Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
+Sandro Hanea <40202887+sandrohanea@users.noreply.github.com>
+Sergio López <slp@redhat.com>
 Sergio López <slp@sinrega.org>
+Shanshan Shen <467638484@qq.com>
+Shijie <821898965@qq.com>
+Shupei Fan <dymarkfan@outlook.com>
 Siddharth Ramakrishnan <srr2141@columbia.edu>
+Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
 Simon Moisselin <simon.moisstoll@gmail.com>
 Sindre Sorhus <sindresorhus@gmail.com>
 Slava Primenko <primenko.s@gmail.com>
+Srihari-mcw <96763064+Srihari-mcw@users.noreply.github.com>
+Stavros Panakakis <53979866+Stavrospanakakis@users.noreply.github.com>
+Stefan Sydow <s.sydow@heinlein-video.de>
+Stefan Sydow <stefan@sydow.email>
 Syahmi Azhar <prsyahmi@gmail.com>
 Syed Jafri <syedjafri97@gmail.com>
 Sơn Phan Trung <phantrungson17@gmail.com>
@@ -205,37 +343,63 @@ 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>
+Thamster <Thamster@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>
+Tim Miller <drasticactions@users.noreply.github.com>
 Timothy Cronin <40186632+4imothy@users.noreply.github.com>
 Tobrun <tobrun.van.nuland@gmail.com>
 Todd <taf2@users.noreply.github.com>
+Toliver <teejae@gmail.com>
 Tong Li <31761981+litongjava@users.noreply.github.com>
+Tony Wasserka <4840017+neobrain@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>
+UsernamesLame <156965854+UsernamesLame@users.noreply.github.com>
 Vadim Peretokin <vperetokin@hey.com>
 Valentin Gosu <1454649+valenting@users.noreply.github.com>
+Vin Misra <vinith@alum.mit.edu>
 Vulcan <93451215+trholding@users.noreply.github.com>
 WhiteOlivierus <36532695+WhiteOlivierus@users.noreply.github.com>
+William Tambellini <william.tambellini@gmail.com>
+William Tambellini <wtambellini@sdl.com>
+Wilson Silva <wilson.dsigns@gmail.com>
 Xiang (Kevin) Li <kevinli020508@gmail.com>
 Xiao-Yong Jin <jinxiaoyong@gmail.com>
 XiaotaoChen <chenxiaotao1234@gmail.com>
+Xingchen Song(宋星辰) <xingchensong1996@163.com>
+Xinpeng Dou <81913537+Dou-Git@users.noreply.github.com>
+Xuan Son Nguyen <thichthat@gmail.com>
 Yajing Tang <phillis@google.com>
 Yang Shen <aplshenyang@gmail.com>
 Yunès <jean.baptiste.yunes@free.fr>
+Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
+Yusuf Redžić <48274562+redzic@users.noreply.github.com>
 ZaBlazzingZephyrus <119159668+blazingzephyr@users.noreply.github.com>
+Zhenwei Jin <109658203+kylo5aby@users.noreply.github.com>
+Zhiyuan Li <lizhiyuan@uniartisan.com>
+Zhiyuan Li <uniartisan2017@gmail.com>
 Zigfrid Zvezdin <ziggerZZ@gmail.com>
 Zollner <24618122+Zolliner@users.noreply.github.com>
+a3sh <38979186+A3shTnT@users.noreply.github.com>
+ag2s20150909 <19373730+ag2s20150909@users.noreply.github.com>
+agray3 <agray3@users.noreply.github.com>
 ai-at-home <149282006+ai-at-home@users.noreply.github.com>
+aldorof <aldorof@users.noreply.github.com>
 alonfaraj <alonfaraj@gmail.com>
+amd-dwang <dong.wang@amd.com>
+amritahs-ibm <amritahs@linux.vnet.ibm.com>
 andypayne <apayne@gmail.com>
 ardfork <134447697+ardfork@users.noreply.github.com>
+arizhih <40765267+arizhih@users.noreply.github.com>
 automaticcat <daogiatuank54@gmail.com>
+bandoti <141645996+bandoti@users.noreply.github.com>
 be-next <jerome.ramette@gmail.com>
 bert hubert <bert@hubertnet.nl>
+billyct <billy_allen@126.com>
 bmwl <brian.marshall@tolko.com>
 bobqianic <129547291+bobqianic@users.noreply.github.com>
 bocytko <bocytko+github@gmail.com>
@@ -248,7 +412,9 @@ 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>
+compilade <git@compilade.net>
 conradg <conradjgodfrey@gmail.com>
+crummyh <elijah@crums.us>
 ddpasa <112642920+ddpasa@users.noreply.github.com>
 denersc <denerstassun@gmail.com>
 dscripka <dscripka@users.noreply.github.com>
@@ -256,28 +422,55 @@ duthils <duthils@duthils.net>
 ecneladis <ecneladis@users.noreply.github.com>
 faker <nspyia2002@gmail.com>
 fitzsim <fitzsim@fitzsim.org>
+fj-y-saito <85871716+fj-y-saito@users.noreply.github.com>
 fraxy-v <65565042+fraxy-v@users.noreply.github.com>
 genevera (she/her) <genevera@users.noreply.github.com>
 geniusnut <geniusnut@gmail.com>
+gilbertgong <gilbert.gong@gmail.com>
+gn64 <yukikaze.jp@gmail.com>
+goldwaving <77494627+goldwaving@users.noreply.github.com>
 greeshmay <greeshmay@gmail.com>
+haopeng <657407891@qq.com>
+hipudding <huafengchun@gmail.com>
+hsinhoyeh <yhh92u@gmail.com>
 hydai <z54981220@gmail.com>
 iamthad <thadeus.j.fleming@gmail.com>
+issixx <46835150+issixx@users.noreply.github.com>
 james wolf <contractorwolf@hotmail.com>
+jdomke <28772296+jdomke@users.noreply.github.com>
+jettoblack <jettoblack@gmail.com>
+jiez <373447296@qq.com>
 joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
 jorismertz <35079666+jorismertz@users.noreply.github.com>
+junchao-loongson <68935141+junchao-loongson@users.noreply.github.com>
 junkfood <69683722+JunkFood02@users.noreply.github.com>
 jwijffels <jwijffels@bnosac.be>
+k.h.lai <adrian.k.h.lai@outlook.com>
 kamranjon <kamranjon@gmail.com>
 katsu560 <katsu560oo-@docomo.ne.jp>
 kennethge <57784063+kenneth-ge@users.noreply.github.com>
 keyehzy <msamuel@aluno.puc-rio.br>
+kunnis <kunnis@users.noreply.github.com>
+l3utterfly <gc.pthzfoldr@gmail.com>
 leejet <leejet714@gmail.com>
+leo-pony <nengjunma@outlook.com>
+lhez <quic_lih@quicinc.com>
 litong <31761981+litongjava@users.noreply.github.com>
+liuwei-git <14815172+liuwei-git@users.noreply.github.com>
 lnyan <lkwq007@gmail.com>
+luoyu-intel <yu.luo@intel.com>
 m.bell <m.bell@techsmith.com>
+mahorozte <41834471+mahorozte@users.noreply.github.com>
+mashizora <30516315+mashizora@users.noreply.github.com>
+matt23654 <matthew.webber@protonmail.com>
+matteo <matteogeniaccio@yahoo.it>
+mgrachten <maarten@grachten.eu>
 mkiol <mkiol@users.noreply.github.com>
+mky_coder <47767389+mkycoder@users.noreply.github.com>
 novag <7754358+novag@users.noreply.github.com>
 pajowu <pajowu@pajowu.de>
+pengxin99 <pengxin.yuan@intel.com>
+petterreinholdtsen <pere-github@hungry.com>
 polarmoon <90010972+polarmoon@users.noreply.github.com>
 rlapray <lapray.romain@gmail.com>
 sandrohanea <40202887+sandrohanea@users.noreply.github.com>
@@ -287,15 +480,31 @@ shikokuchuo <53399081+shikokuchuo@users.noreply.github.com>
 slaren <slarengh@gmail.com>
 slashlib <slashlib@users.noreply.github.com>
 snadampal <87143774+snadampal@users.noreply.github.com>
+someone13574 <81528246+someone13574@users.noreply.github.com>
 st-gr <38470677+st-gr@users.noreply.github.com>
+stduhpf <stephduh@live.fr>
+stormofice <58337328+stormofice@users.noreply.github.com>
 texmex76 <40733439+texmex76@users.noreply.github.com>
 thefinaldegree <thefinaldegree@gmail.com>
+thewh1teagle <61390950+thewh1teagle@users.noreply.github.com>
+toboil-features <160222185+toboil-features@users.noreply.github.com>
 trixirt <trix@redhat.com>
 ulatekh <ulatekh@yahoo.com>
 undef <undefdev@gmail.com>
+uvos <devnull@uvos.xyz>
+uvos <philipp@uvos.xyz>
+valVk <valVk@users.noreply.github.com>
 venkr <venkateshrameshkumar+1@gmail.com>
 vicalloy <zbirder@gmail.com>
+wangshuai09 <391746016@qq.com>
+woachk <24752637+woachk@users.noreply.github.com>
+xctan <axunlei@gmail.com>
 xdrudis <xavierdrudis@yahoo.es>
+yuri@FreeBSD <yuri@FreeBSD>
+zhangjixiong <code.zjx@gmail.com>
+zhentaoyu <zhentao.yu@intel.com>
 zhouwg <6889919+zhouwg@users.noreply.github.com>
+zhouwg <zhouwg2000@gmail.com>
+谢乃闻 <sienaiwun@users.noreply.github.com>
 布客飞龙 <562826179@qq.com>
 Артём Земляк <azemlyak@smart-consulting.ru>

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
 project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.7.4)
+project("whisper.cpp" VERSION 1.7.5)
 include(CheckIncludeFileCXX)
 
 set(SOVERSION 1)
@@ -38,8 +38,13 @@ if (EMSCRIPTEN)
 
     # TODO: without these, we get the following error:
     #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
-    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
+
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -s TOTAL_STACK=5242880")
+    set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -s TOTAL_STACK=5242880")
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-deprecated")
 else()
     if (MINGW)
         set(BUILD_SHARED_LIBS_DEFAULT OFF)
@@ -62,7 +67,8 @@ option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"
 option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
 
 # build
-option(WHISPER_FATAL_WARNINGS "whisper: enable -Werror flag" OFF)
+option(WHISPER_FATAL_WARNINGS  "whisper: enable -Werror flag"               OFF)
+option(WHISPER_USE_SYSTEM_GGML "whisper: use system-installed GGML library" OFF)
 
 # sanitizers
 option(WHISPER_SANITIZE_THREAD    "whisper: enable thread sanitizer"    OFF)
@@ -121,7 +127,15 @@ whisper_option_depr(WARNING     WHISPER_SYCL_F16            GGML_SYCL_F16)
 #
 
 if (NOT TARGET ggml)
-    add_subdirectory(ggml)
+    if (WHISPER_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+        if (NOT ggml_FOUND)
+            message(FATAL_ERROR "System-installed GGML library not found.")
+        endif()
+        add_library(ggml ALIAS ggml::ggml)
+    else()
+        add_subdirectory(ggml)
+    endif()
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 add_subdirectory(src)
@@ -176,8 +190,8 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
 #
 
 if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    #include(CTest)
-    #add_subdirectory(tests)
+    include(CTest)
+    add_subdirectory(tests)
 endif ()
 
 if (WHISPER_BUILD_EXAMPLES)

Makefile

@@ -18,17 +18,6 @@ samples:
 	@wget --quiet --show-progress -O samples/mm1.wav https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav
 	@wget --quiet --show-progress -O samples/a13.mp3 https://upload.wikimedia.org/wikipedia/commons/transcoded/6/6f/Apollo13-wehaveaproblem.ogg/Apollo13-wehaveaproblem.ogg.mp3
 	@wget --quiet --show-progress -O samples/diffusion2023-07-03.flac https://archive.org/download/diffusion2023-07-03/diffusion2023-07-03.flac
-	@echo "Converting to 16-bit WAV ..."
-	@ffmpeg -loglevel -0 -y -i samples/gb0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb0.wav
-	@ffmpeg -loglevel -0 -y -i samples/gb1.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb1.wav
-	@ffmpeg -loglevel -0 -y -i samples/hp0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/hp0.wav
-	@rm samples/*.ogg
-	@ffmpeg -loglevel -0 -y -i samples/mm1.wav -ar 16000 -ac 1 -c:a pcm_s16le samples/mm0.wav
-	@rm samples/mm1.wav
-	@ffmpeg -loglevel -0 -y -i samples/a13.mp3 -ar 16000 -ac 1 -c:a pcm_s16le -ss 00:00:00 -to 00:00:30 samples/a13.wav
-	@rm samples/a13.mp3
-	@ffmpeg -loglevel -0 -y -i samples/diffusion2023-07-03.flac -ar 16000 -ac 1 -c:a pcm_s16le samples/diffusion2023-07-03.wav
-	@rm samples/diffusion2023-07-03.flac
 
 #
 # Models
@@ -59,7 +48,7 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 larg
 	@echo "Running $@ on all samples in ./samples ..."
 	@echo "==============================================="
 	@echo ""
-	@for f in samples/*.wav; do \
+	@for f in samples/*$(.flac .mp3 .ogg .wav); do \
 		echo "----------------------------------------------" ; \
 		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
 	    echo "----------------------------------------------" ; \

Package.swift

@@ -1,19 +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: [
-        .systemLibrary(name: "whisper", pkgConfig: "whisper"),
-    ]
-)

README.md

@@ -7,14 +7,17 @@
 [![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.7.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+> [!NOTE]
+> New maintenance roadmap: https://github.com/ggerganov/whisper.cpp/discussions/2788
+
+Stable: [v1.7.5](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.5) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
 - Plain C/C++ implementation without dependencies
 - Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](#core-ml-support)
 - AVX intrinsics support for x86 architectures
-- VSX intrinsics support for POWER architectures
+- [VSX intrinsics support for POWER architectures](#power-vsx-intrinsics)
 - Mixed F16 / F32 precision
 - [Integer quantization support](#quantization)
 - Zero memory allocations at runtime
@@ -136,6 +139,20 @@ make -j large-v3-turbo
 | medium | 1.5 GiB | ~2.1 GB |
 | large  | 2.9 GiB | ~3.9 GB |
 
+## POWER VSX Intrinsics
+
+`whisper.cpp` supports POWER architectures and includes code which
+significantly speeds operation on Linux running on POWER9/10, making it
+capable of faster-than-realtime transcription on underclocked Raptor
+Talos II. Ensure you have a BLAS package installed, and replace the
+standard cmake setup with:
+
+```bash
+# build with GGML_BLAS defined
+cmake -B build -DGGML_BLAS=1
+cmake --build build --config Release
+./build/bin/whisper-cli [ .. etc .. ]
+
 ## Quantization
 
 `whisper.cpp` supports integer quantization of the Whisper `ggml` models.
@@ -167,11 +184,11 @@ 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.
+  - Python 3.11 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:
-    - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
-    - To activate the environment, use: `conda activate py310-whisper`
+    - To create an environment, use: `conda create -n py311-whisper python=3.11 -y`
+    - To activate the environment, use: `conda activate py311-whisper`
 
 - Generate a Core ML model. For example, to generate a `base.en` model, use:
 
@@ -360,6 +377,38 @@ Run the inference examples as usual, for example:
 - If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
 - If you run successfully with your Ascend NPU device, please help update the table `Verified devices`.
 
+## Docker
+
+### Prerequisites
+
+- 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:
@@ -378,7 +427,8 @@ For detailed instructions on how to use Conan, please refer to the [Conan docume
 
 This is a naive example of performing real-time inference on audio from your microphone.
 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).
+More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10). 
+You will need to have [sdl2](https://wiki.libsdl.org/SDL2/Installation) installed for it to work properly. 
 
 ```bash
 cmake -B build -DWHISPER_SDL2=ON

Sources/whisper/module.modulemap

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

Sources/whisper/whisper.h

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

bindings/go/Makefile

@@ -11,11 +11,11 @@ UNAME_M := $(shell uname -m)
 endif
 
 GGML_METAL_PATH_RESOURCES := $(abspath ../..)
-BUILD_DIR := build
+BUILD_DIR := build_go
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
 INCLUDE_PATH := $(abspath ../../include):$(abspath ../../ggml/include)
-LIBRARY_PATH := $(abspath ../..)
+LIBRARY_PATH := $(abspath ../../${BUILD_DIR}/src:$(abspath ../../${BUILD_DIR}/ggml/src))
 
 ifeq ($(GGML_CUDA),1)
 	LIBRARY_PATH := $(LIBRARY_PATH):$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib/
@@ -29,8 +29,10 @@ endif
 all: clean whisper examples
 
 whisper: mkdir
-	@echo Build whisper
-	@${MAKE} -C ../.. libwhisper.a
+	cmake -S ../.. -B ../../${BUILD_DIR} \
+		-DCMAKE_BUILD_TYPE=Release \
+		-DBUILD_SHARED_LIBS=OFF
+	cmake --build ../../${BUILD_DIR} --target whisper
 
 test: model-small whisper modtidy
 ifeq ($(UNAME_S),Darwin)

bindings/go/README.md

@@ -31,7 +31,7 @@ func main() {
 	if err != nil {
 		panic(err)
 	}
-	if err := context.Process(samples, nil, nil); err != nil {
+	if err := context.Process(samples, nil, nil, nil); err != nil {
 		return err
 	}
 

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

@@ -9,22 +9,23 @@ import (
 // ContextForSignal returns a context object which is cancelled when a signal
 // is received. It returns nil if no signal parameter is provided
 func ContextForSignal(signals ...os.Signal) context.Context {
-	if len(signals) == 0 {
-		return nil
-	}
+    if len(signals) == 0 {
+        return nil
+    }
 
-	ch := make(chan os.Signal)
-	ctx, cancel := context.WithCancel(context.Background())
+    ch := make(chan os.Signal, 1) // Buffered channel with space for 1 signal
+    ctx, cancel := context.WithCancel(context.Background())
 
-	// Send message on channel when signal received
-	signal.Notify(ch, signals...)
+    // Send message on channel when signal received
+    signal.Notify(ch, signals...)
 
-	// When any signal received, call cancel
-	go func() {
-		<-ch
-		cancel()
-	}()
+    // When any signal is received, call cancel
+    go func() {
+        <-ch
+        cancel()
+    }()
 
-	// Return success
-	return ctx
+    // Return success
+    return ctx
 }
+

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

@@ -9,6 +9,7 @@ import (
 	"net/url"
 	"os"
 	"path/filepath"
+	"strings"
 	"syscall"
 	"time"
 )
@@ -17,14 +18,27 @@ import (
 // CONSTANTS
 
 const (
-	srcUrl  = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main" // The location of the models
-	srcExt  = ".bin"                                                      // Filename extension
-	bufSize = 1024 * 64                                                   // Size of the buffer used for downloading the model
+	srcUrl  = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/" // The location of the models
+	srcExt  = ".bin"                                                       // Filename extension
+	bufSize = 1024 * 64                                                    // Size of the buffer used for downloading the model
 )
 
 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", "large-v3-turbo"}
+	modelNames = []string{
+		"tiny", "tiny-q5_1", "tiny-q8_0",
+		"tiny.en", "tiny.en-q5_1", "tiny.en-q8_0",
+		"base", "base-q5_1", "base-q8_0",
+		"base.en", "base.en-q5_1", "base.en-q8_0",
+		"small", "small-q5_1", "small-q8_0",
+		"small.en", "small.en-q5_1", "small.en-q8_0",
+		"medium", "medium-q5_0", "medium-q8_0",
+		"medium.en", "medium.en-q5_0", "medium.en-q8_0",
+		"large-v1",
+		"large-v2", "large-v2-q5_0", "large-v2-q8_0",
+		"large-v3", "large-v3-q5_0",
+		"large-v3-turbo", "large-v3-turbo-q5_0", "large-v3-turbo-q8_0",
+	}
 )
 
 var (
@@ -44,7 +58,25 @@ var (
 func main() {
 	flag.Usage = func() {
 		name := filepath.Base(flag.CommandLine.Name())
-		fmt.Fprintf(flag.CommandLine.Output(), "Usage: %s [options] <model>\n\n", name)
+		fmt.Fprintf(flag.CommandLine.Output(), `
+			Usage: %s [options] [<model>...]
+
+			Options:
+  			-out string     Specify the output folder where models will be saved.
+                  			Default: Current working directory.
+  			-timeout duration Set the maximum duration for downloading a model.
+            			      Example: 10m, 1h (default: 30m0s).
+  			-quiet           Suppress all output except errors.
+
+			Examples:
+  			1. Download a specific model:
+     			%s -out ./models tiny-q8_0
+
+			  2. Download all models:
+     			%s -out ./models
+
+			`, name, name, name)
+
 		flag.PrintDefaults()
 	}
 	flag.Parse()
@@ -114,23 +146,87 @@ func GetOut() (string, error) {
 // GetModels returns the list of models to download
 func GetModels() []string {
 	if flag.NArg() == 0 {
-		return modelNames
-	} else {
-		return flag.Args()
+		fmt.Println("No model specified.")
+		fmt.Println("Preparing to download all models...")
+
+		// Calculate total download size
+		fmt.Println("Calculating total download size...")
+		totalSize, err := CalculateTotalDownloadSize(modelNames)
+		if err != nil {
+			fmt.Println("Error calculating download sizes:", err)
+			os.Exit(1)
+		}
+
+		fmt.Println("View available models: https://huggingface.co/ggerganov/whisper.cpp/tree/main")
+		fmt.Printf("Total download size: %.2f GB\n", float64(totalSize)/(1024*1024*1024))
+		fmt.Println("Would you like to download all models? (y/N)")
+
+		// Prompt for user input
+		var response string
+		fmt.Scanln(&response)
+		if response != "y" && response != "Y" {
+			fmt.Println("Aborting. Specify a model to download.")
+			os.Exit(0)
+		}
+
+		return modelNames // Return all models if confirmed
 	}
+	return flag.Args() // Return specific models if arguments are provided
+}
+
+func CalculateTotalDownloadSize(models []string) (int64, error) {
+	var totalSize int64
+	client := http.Client{}
+
+	for _, model := range models {
+		modelURL, err := URLForModel(model)
+		if err != nil {
+			return 0, err
+		}
+
+		// Issue a HEAD request to get the file size
+		req, err := http.NewRequest("HEAD", modelURL, nil)
+		if err != nil {
+			return 0, err
+		}
+
+		resp, err := client.Do(req)
+		if err != nil {
+			return 0, err
+		}
+		resp.Body.Close()
+
+		if resp.StatusCode != http.StatusOK {
+			fmt.Printf("Warning: Unable to fetch size for %s (HTTP %d)\n", model, resp.StatusCode)
+			continue
+		}
+
+		size := resp.ContentLength
+		totalSize += size
+	}
+	return totalSize, nil
 }
 
 // URLForModel returns the URL for the given model on huggingface.co
 func URLForModel(model string) (string, error) {
+	// Ensure "ggml-" prefix is added only once
+	if !strings.HasPrefix(model, "ggml-") {
+		model = "ggml-" + model
+	}
+
+	// Ensure ".bin" extension is added only once
 	if filepath.Ext(model) != srcExt {
 		model += srcExt
 	}
+
+	// Parse the base URL
 	url, err := url.Parse(srcUrl)
 	if err != nil {
 		return "", err
-	} else {
-		url.Path = filepath.Join(url.Path, model)
 	}
+
+	// Ensure no trailing slash in the base URL
+	url.Path = fmt.Sprintf("%s/%s", strings.TrimSuffix(url.Path, "/"), model)
 	return url.String(), nil
 }
 

bindings/go/examples/go-whisper/process.go

@@ -67,7 +67,7 @@ func Process(model whisper.Model, path string, flags *Flags) error {
 	// Process the data
 	fmt.Fprintf(flags.Output(), "  ...processing %q\n", path)
 	context.ResetTimings()
-	if err := context.Process(data, cb, nil); err != nil {
+	if err := context.Process(data, nil, cb, nil); err != nil {
 		return err
 	}
 

bindings/go/pkg/whisper/context.go

@@ -71,6 +71,10 @@ func (context *context) Language() string {
 	return whisper.Whisper_lang_str(context.params.Language())
 }
 
+func (context *context) DetectedLanguage() string {
+	return whisper.Whisper_lang_str(context.model.ctx.Whisper_full_lang_id())
+}
+
 // Set translate flag
 func (context *context) SetTranslate(v bool) {
 	context.params.SetTranslate(v)
@@ -189,6 +193,7 @@ func (context *context) WhisperLangAutoDetect(offset_ms int, n_threads int) ([]f
 // Process new sample data and return any errors
 func (context *context) Process(
 	data []float32,
+	callEncoderBegin EncoderBeginCallback,
 	callNewSegment SegmentCallback,
 	callProgress ProgressCallback,
 ) error {
@@ -203,7 +208,20 @@ func (context *context) Process(
 	// We don't do parallel processing at the moment
 	processors := 0
 	if processors > 1 {
-		if err := context.model.ctx.Whisper_full_parallel(context.params, data, processors, nil, func(new int) {
+		if err := context.model.ctx.Whisper_full_parallel(context.params, data, processors, callEncoderBegin,
+			func(new int) {
+				if callNewSegment != nil {
+					num_segments := context.model.ctx.Whisper_full_n_segments()
+					s0 := num_segments - new
+					for i := s0; i < num_segments; i++ {
+						callNewSegment(toSegment(context.model.ctx, i))
+					}
+				}
+			}); err != nil {
+			return err
+		}
+	} else if err := context.model.ctx.Whisper_full(context.params, data, callEncoderBegin,
+		func(new int) {
 			if callNewSegment != nil {
 				num_segments := context.model.ctx.Whisper_full_n_segments()
 				s0 := num_segments - new
@@ -211,22 +229,11 @@ func (context *context) Process(
 					callNewSegment(toSegment(context.model.ctx, i))
 				}
 			}
-		}); err != nil {
-			return err
-		}
-	} else if err := context.model.ctx.Whisper_full(context.params, data, nil, func(new int) {
-		if callNewSegment != nil {
-			num_segments := context.model.ctx.Whisper_full_n_segments()
-			s0 := num_segments - new
-			for i := s0; i < num_segments; i++ {
-				callNewSegment(toSegment(context.model.ctx, i))
+		}, func(progress int) {
+			if callProgress != nil {
+				callProgress(progress)
 			}
-		}
-	}, func(progress int) {
-		if callProgress != nil {
-			callProgress(progress)
-		}
-	}); err != nil {
+		}); err != nil {
 		return err
 	}
 

bindings/go/pkg/whisper/context_test.go

@@ -88,6 +88,37 @@ func TestProcess(t *testing.T) {
 	context, err := model.NewContext()
 	assert.NoError(err)
 
-	err = context.Process(data, nil, nil)
+	err = context.Process(data, nil, nil, nil)
 	assert.NoError(err)
 }
+
+func TestDetectedLanguage(t *testing.T) {
+	assert := assert.New(t)
+
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Decode the WAV file - load the full buffer
+	dec := wav.NewDecoder(fh)
+	buf, err := dec.FullPCMBuffer()
+	assert.NoError(err)
+	assert.Equal(uint16(1), dec.NumChans)
+
+	data := buf.AsFloat32Buffer().Data
+
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	context, err := model.NewContext()
+	assert.NoError(err)
+
+	err = context.Process(data, nil, nil, nil)
+	assert.NoError(err)
+
+	expectedLanguage := "en"
+	actualLanguage := context.DetectedLanguage()
+	assert.Equal(expectedLanguage, actualLanguage)
+}

bindings/go/pkg/whisper/interface.go

@@ -16,6 +16,10 @@ type SegmentCallback func(Segment)
 // processing. It is called during the Process function
 type ProgressCallback func(int)
 
+// EncoderBeginCallback is the callback function for checking if we want to
+// continue processing. It is called during the Process function
+type EncoderBeginCallback func() bool
+
 // Model is the interface to a whisper model. Create a new model with the
 // function whisper.New(string)
 type Model interface {
@@ -31,12 +35,13 @@ type Model interface {
 	Languages() []string
 }
 
-// Context is the speach recognition context.
+// Context is the speech recognition context.
 type Context interface {
 	SetLanguage(string) error // Set the language to use for speech recognition, use "auto" for auto detect language.
 	SetTranslate(bool)        // Set translate flag
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language
+	DetectedLanguage() string // Get detected language
 
 	SetOffset(time.Duration)          // Set offset
 	SetDuration(time.Duration)        // Set duration
@@ -58,7 +63,7 @@ type Context interface {
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the
 	// callback function during processing.
-	Process([]float32, SegmentCallback, ProgressCallback) error
+	Process([]float32, EncoderBeginCallback, SegmentCallback, ProgressCallback) error
 
 	// After process is called, return segments until the end of the stream
 	// is reached, when io.EOF is returned.

bindings/go/whisper.go

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

bindings/java/build.gradle

@@ -25,13 +25,13 @@ sourceSets {
 }
 
 tasks.register('copyLibwhisperDynlib', Copy) {
-    from '../../build'
-    include 'libwhisper.dynlib'
+    from '../../build/src'
+    include 'libwhisper.dylib'
     into 'build/generated/resources/main/darwin'
 }
 
 tasks.register('copyLibwhisperSo', Copy) {
-    from '../../build'
+    from '../../build/src'
     include 'libwhisper.so'
     into 'build/generated/resources/main/linux-x86-64'
 }
@@ -55,7 +55,12 @@ java {
     withJavadocJar()
 }
 
+sourcesJar() {
+    dependsOn copyLibs
+}
+
 jar {
+    dependsOn copyLibs
     exclude '**/whisper_java.exp', '**/whisper_java.lib'
 }
 
@@ -67,6 +72,9 @@ tasks.withType(Test) {
     useJUnitPlatform()
 }
 
+test.dependsOn copyLibs
+processResources.dependsOn copyLibs
+
 dependencies {
     implementation "net.java.dev.jna:jna:5.13.0"
     testImplementation "org.junit.jupiter:junit-jupiter:5.9.2"

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

@@ -0,0 +1,24 @@
+package io.github.ggerganov.whispercpp;
+
+/**
+ * Presets for alignment heads in DTW token timestamps
+ */
+public class WhisperConstants {
+    // Alignment heads presets
+    public static final int WHISPER_AHEADS_NONE = 0;
+    public static final int WHISPER_AHEADS_TINY_EN = 1;
+    public static final int WHISPER_AHEADS_TINY = 2;
+    public static final int WHISPER_AHEADS_BASE_EN = 3;
+    public static final int WHISPER_AHEADS_BASE = 4;
+    public static final int WHISPER_AHEADS_SMALL_EN = 5;
+    public static final int WHISPER_AHEADS_SMALL = 6;
+    public static final int WHISPER_AHEADS_MEDIUM_EN = 7;
+    public static final int WHISPER_AHEADS_MEDIUM = 8;
+    public static final int WHISPER_AHEADS_LARGE_V1 = 9;
+    public static final int WHISPER_AHEADS_LARGE_V2 = 10;
+    public static final int WHISPER_AHEADS_LARGE_V3 = 11;
+    public static final int WHISPER_AHEADS_LARGE_V3_TURBO = 12;
+    public static final int WHISPER_AHEADS_CUSTOM = 13;
+    public static final int WHISPER_AHEADS_N_TOP_MOST = 14;
+    public static final int WHISPER_AHEADS_COUNT = 15;
+}

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

@@ -1,7 +1,9 @@
 package io.github.ggerganov.whispercpp;
 
+import com.sun.jna.NativeLong;
 import com.sun.jna.Structure;
 import com.sun.jna.ptr.PointerByReference;
+import com.sun.jna.Pointer;
 import io.github.ggerganov.whispercpp.ggml.GgmlType;
 import io.github.ggerganov.whispercpp.WhisperModel;
 import io.github.ggerganov.whispercpp.params.WhisperContextParams;
@@ -9,33 +11,26 @@ import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import java.util.List;
 
 public class WhisperContext extends Structure {
-    int t_load_us = 0;
-    int t_start_us = 0;
+    public NativeLong t_load_us;
+    public NativeLong t_start_us;
 
     /** weight type (FP32 / FP16 / QX) */
-    GgmlType wtype = GgmlType.GGML_TYPE_F16;
+    public GgmlType wtype = GgmlType.GGML_TYPE_F16;
     /** intermediate type (FP32 or FP16) */
-    GgmlType itype = GgmlType.GGML_TYPE_F16;
+    public GgmlType itype = GgmlType.GGML_TYPE_F16;
 
-//    WhisperModel model;
-    public PointerByReference model;
-//    whisper_vocab vocab;
-//    whisper_state * state = nullptr;
-    public PointerByReference vocab;
-    public PointerByReference state;
+    public WhisperContextParams.ByValue params;
+
+    public Pointer model;
+    public Pointer vocab;
+    public Pointer state;
 
     /** populated by whisper_init_from_file_with_params() */
-    String path_model;
-    WhisperContextParams params;
+    public Pointer path_model;
 
-//    public static class ByReference extends WhisperContext implements Structure.ByReference {
-//    }
-//
-//    public static class ByValue extends WhisperContext implements Structure.ByValue {
-//    }
-//
-//    @Override
-//    protected List<String> getFieldOrder() {
-//        return List.of("t_load_us", "t_start_us", "wtype", "itype", "model", "vocab", "state", "path_model");
-//    }
+    @Override
+    protected List<String> getFieldOrder() {
+        return List.of("t_load_us", "t_start_us", "wtype", "itype",
+                "params", "model", "vocab", "state", "path_model");
+    }
 }

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

@@ -43,11 +43,11 @@ public class WhisperCpp implements AutoCloseable {
      * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
      * @param params - params to use when initialising the context
      */
-    public void initContext(String modelPath, WhisperContextParams params) throws FileNotFoundException {
+    public void initContext(String modelPath, WhisperContextParams.ByValue params) throws FileNotFoundException {
         initContextImpl(modelPath, params);
     }
 
-    private void initContextImpl(String modelPath, WhisperContextParams params) throws FileNotFoundException {
+    private void initContextImpl(String modelPath, WhisperContextParams.ByValue params) throws FileNotFoundException {
         if (ctx != null) {
             lib.whisper_free(ctx);
         }
@@ -69,15 +69,13 @@ public class WhisperCpp implements AutoCloseable {
 
     /**
      * 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:
-     * - call `whisper_free_context_params()`
-     * - `Native.free(Pointer.nativeValue(pointer));`
+     * Returns a ByValue instance to ensure proper parameter passing to native code.
      */
-    public WhisperContextParams getContextDefaultParams() {
-        paramsPointer = lib.whisper_context_default_params_by_ref();
-        WhisperContextParams params = new WhisperContextParams(paramsPointer);
-        params.read();
-        return params;
+    public WhisperContextParams.ByValue getContextDefaultParams() {
+        WhisperContextParams.ByValue valueParams = new WhisperContextParams.ByValue(
+            lib.whisper_context_default_params_by_ref());
+        valueParams.read();
+        return valueParams;
     }
     
     /**
@@ -88,7 +86,7 @@ public class WhisperCpp implements AutoCloseable {
      *
      * @param strategy - GREEDY
      */
-    public WhisperFullParams getFullDefaultParams(WhisperSamplingStrategy strategy) {
+    public WhisperFullParams.ByValue getFullDefaultParams(WhisperSamplingStrategy strategy) {
         Pointer pointer;
 
         // whisper_full_default_params_by_ref allocates memory which we need to delete, so only create max 1 pointer for each strategy.
@@ -104,7 +102,7 @@ public class WhisperCpp implements AutoCloseable {
             pointer = beamParamsPointer;
         }
 
-        WhisperFullParams params = new WhisperFullParams(pointer);
+        WhisperFullParams.ByValue params = new WhisperFullParams.ByValue(pointer);
         params.read();
         return params;
     }
@@ -138,15 +136,21 @@ public class WhisperCpp implements AutoCloseable {
     }
 
     /**
-     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text.
+     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text.
      * Not thread safe for same context
      * Uses the specified decoding strategy to obtain the text.
      */
-    public String fullTranscribe(WhisperFullParams whisperParams, float[] audioData) throws IOException {
+    public String fullTranscribe(WhisperFullParams.ByValue whisperParams, float[] audioData) throws IOException {
         if (ctx == null) {
             throw new IllegalStateException("Model not initialised");
         }
 
+        /*
+        WhisperFullParams.ByValue valueParams = new WhisperFullParams.ByValue(
+            lib.whisper_full_default_params_by_ref(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH.ordinal()));
+        valueParams.read();
+        */
+
         if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
             throw new IOException("Failed to process audio");
         }
@@ -163,12 +167,17 @@ 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) {
+        WhisperFullParams.ByValue valueParams = new WhisperFullParams.ByValue(
+            lib.whisper_full_default_params_by_ref(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH.ordinal()));
+        valueParams.read();
+
+        if (lib.whisper_full(ctx, valueParams, audioData, audioData.length) != 0) {
             throw new IOException("Failed to process audio");
         }
 

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

@@ -38,7 +38,7 @@ public interface WhisperCppJnaLibrary extends Library {
      * @param params     Pointer to whisper_context_params
      * @return Whisper context on success, null on failure
      */
-    Pointer whisper_init_from_file_with_params(String path_model, WhisperContextParams params);
+    Pointer whisper_init_from_file_with_params(String path_model, WhisperContextParams.ByValue params);
 
     /**
      * Allocate (almost) all memory needed for the model by loading from a buffer.
@@ -180,12 +180,12 @@ public interface WhisperCppJnaLibrary extends Library {
     /**
      * @return the id of the specified language, returns -1 if not found.
      * Examples:
-     *   "de" -> 2
-     *   "german" -> 2
+     *   "de" -> 2
+     *   "german" -> 2
      */
     int whisper_lang_id(String lang);
 
-    /** @return the short string of the specified language id (e.g. 2 -> "de"), returns nullptr if not found */
+    /** @return the short string of the specified language id (e.g. 2 -> "de"), returns nullptr if not found */
     String whisper_lang_str(int id);
 
     /**
@@ -268,20 +268,21 @@ public interface WhisperCppJnaLibrary extends Library {
     void whisper_free_params(Pointer params);
 
     /**
-     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
      * Not thread safe for same context
      * Uses the specified decoding strategy to obtain the text.
      */
-    int whisper_full(Pointer ctx, WhisperFullParams params, final float[] samples, int n_samples);
+    int whisper_full(Pointer ctx, WhisperFullParams.ByValue params, final float[] samples, int n_samples);
 
-    int whisper_full_with_state(Pointer ctx, Pointer state, WhisperFullParams params, final float[] samples, int n_samples);
+    public int whisper_full_with_state(Pointer ctx, Pointer state, WhisperFullParams.ByValue params, float[] samples, int n_samples);
+    //int whisper_full_with_state(Pointer ctx, Pointer state, WhisperFullParams params, final float[] samples, int n_samples);
 
     // Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
     // Result is stored in the default state of the context
     // Not thread safe if executed in parallel on the same context.
     // It seems this approach can offer some speedup in some cases.
     // However, the transcription accuracy can be worse at the beginning and end of each chunk.
-    int whisper_full_parallel(Pointer ctx, WhisperFullParams params, final float[] samples, int n_samples, int n_processors);
+    int whisper_full_parallel(Pointer ctx, WhisperFullParams.ByValue params, final float[] samples, int n_samples, int n_processors);
 
     /**
      * Number of generated text segments.

bindings/java/src/main/java/io/github/ggerganov/whispercpp/callbacks/GgmlAbortCallback.java

@@ -0,0 +1,17 @@
+package io.github.ggerganov.whispercpp.callbacks;
+
+import com.sun.jna.Callback;
+
+/**
+ * Callback for aborting GGML computation
+ * Maps to the C typedef: bool (*ggml_abort_callback)(void * data)
+ */
+public interface GgmlAbortCallback extends Callback {
+    /**
+     * Return true to abort the computation, false to continue
+     *
+     * @param data User data passed to the callback
+     * @return true to abort, false to continue
+     */
+    boolean invoke(com.sun.jna.Pointer data);
+}

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

@@ -0,0 +1,30 @@
+package io.github.ggerganov.whispercpp.params;
+import com.sun.jna.*;
+import java.util.Arrays;
+import java.util.List;
+
+public class WhisperAhead extends Structure {
+
+    public int n_text_layer;
+
+    public int n_head;
+
+    public WhisperAhead() {
+        super();
+    }
+
+    public WhisperAhead(int textLayer, int head) {
+        super();
+        this.n_text_layer = textLayer;
+        this.n_head = head;
+    }
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("n_text_layer", "n_head");
+    }
+
+    public static class ByReference extends WhisperAhead implements Structure.ByReference {}
+
+    public static class ByValue extends WhisperAhead implements Structure.ByValue {}
+}

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

@@ -0,0 +1,41 @@
+package io.github.ggerganov.whispercpp.params;
+import com.sun.jna.*;
+import java.util.Arrays;
+import java.util.List;
+
+public class WhisperAheads extends Structure {
+    public NativeLong n_heads;
+
+    public Pointer heads;
+
+    public WhisperAheads() {
+        super();
+    }
+
+    /**
+     * Create alignment heads from an array of WhisperAhead objects
+     */
+    public void setHeads(WhisperAhead[] aheadsArray) {
+        this.n_heads = new NativeLong(aheadsArray.length);
+
+        int structSize = aheadsArray[0].size();
+        Memory mem = new Memory(structSize * aheadsArray.length);
+
+        for (int i = 0; i < aheadsArray.length; i++) {
+            aheadsArray[i].write();
+            byte[] buffer = aheadsArray[i].getPointer().getByteArray(0, structSize);
+            mem.write(i * structSize, buffer, 0, buffer.length);
+        }
+
+        this.heads = mem;
+    }
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("n_heads", "heads");
+    }
+
+    public static class ByReference extends WhisperAheads implements Structure.ByReference {}
+
+    public static class ByValue extends WhisperAheads implements Structure.ByValue {}
+}

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

@@ -1,7 +1,5 @@
 package io.github.ggerganov.whispercpp.params;
-
 import com.sun.jna.*;
-
 import java.util.Arrays;
 import java.util.List;
 
@@ -11,21 +9,73 @@ import java.util.List;
  * whisper_context_default_params()
  */
 public class WhisperContextParams extends Structure {
-
     public WhisperContextParams(Pointer p) {
         super(p);
     }
 
-    /** Use GPU for inference Number (default = true) */
+    public WhisperContextParams() {
+        super();
+    }
+
+    /** Use GPU for inference (default = true) */
     public CBool use_gpu;
 
-    /** Use GPU for inference Number (default = true) */
+    /** Use flash attention (default = false) */
+    public CBool flash_attn;
+
+    /** CUDA device to use (default = 0) */
+    public int gpu_device;
+
+    /** [EXPERIMENTAL] Enable token-level timestamps with DTW (default = false) */
+    public CBool dtw_token_timestamps;
+
+    /** [EXPERIMENTAL] Alignment heads preset for DTW */
+    public int dtw_aheads_preset;
+
+    /** Number of top layers to use for DTW when using WHISPER_AHEADS_N_TOP_MOST preset */
+    public int dtw_n_top;
+
+    public WhisperAheads.ByValue dtw_aheads;
+
+    /** DTW memory size (internal use) */
+    public NativeLong dtw_mem_size;
+
+    /** Use GPU for inference */
     public void useGpu(boolean enable) {
         use_gpu = enable ? CBool.TRUE : CBool.FALSE;
     }
 
+    /** Use flash attention */
+    public void useFlashAttn(boolean enable) {
+        flash_attn = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Enable DTW token-level timestamps */
+    public void enableDtwTokenTimestamps(boolean enable) {
+        dtw_token_timestamps = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Set DTW alignment heads preset */
+    public void setDtwAheadsPreset(int preset) {
+        dtw_aheads_preset = preset;
+    }
+
     @Override
     protected List<String> getFieldOrder() {
-        return Arrays.asList("use_gpu");
+        return Arrays.asList(
+            "use_gpu",
+            "flash_attn",
+            "gpu_device",
+            "dtw_token_timestamps",
+            "dtw_aheads_preset",
+            "dtw_n_top",
+            "dtw_aheads",
+            "dtw_mem_size"
+        );
+    }
+
+    public static class ByValue extends WhisperContextParams implements Structure.ByValue {
+        public ByValue() { super(); }
+        public ByValue(Pointer p) { super(p); }
     }
 }

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

@@ -5,6 +5,7 @@ import io.github.ggerganov.whispercpp.callbacks.WhisperEncoderBeginCallback;
 import io.github.ggerganov.whispercpp.callbacks.WhisperLogitsFilterCallback;
 import io.github.ggerganov.whispercpp.callbacks.WhisperNewSegmentCallback;
 import io.github.ggerganov.whispercpp.callbacks.WhisperProgressCallback;
+import io.github.ggerganov.whispercpp.callbacks.GgmlAbortCallback;
 
 import java.util.Arrays;
 import java.util.List;
@@ -16,10 +17,12 @@ import java.util.List;
  */
 public class WhisperFullParams extends Structure {
 
+    public WhisperFullParams() {
+        super();
+    }
+
     public WhisperFullParams(Pointer p) {
         super(p);
-//        super(p, ALIGN_MSVC);
-//        super(p, ALIGN_GNUC);
     }
 
     /** Sampling strategy for whisper_full() function. */
@@ -69,10 +72,10 @@ public class WhisperFullParams extends Structure {
         single_segment = single ? CBool.TRUE : CBool.FALSE;
     }
 
-    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
+    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
     public CBool print_special;
 
-    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
+    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
     public void printSpecial(boolean enable) {
         print_special = enable ? CBool.TRUE : CBool.FALSE;
     }
@@ -129,6 +132,14 @@ public class WhisperFullParams extends Structure {
     /** Maximum tokens per segment (0, default = no limit) */
     public int max_tokens;
 
+    /** [EXPERIMENTAL] Enable debug mode for extra info */
+    public CBool debug_mode;
+
+    /** Enable debug mode */
+    public void enableDebugMode(boolean enable) {
+        debug_mode = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
     /** Overwrite the audio context size (0 = use default). */
     public int audio_ctx;
 
@@ -274,6 +285,16 @@ public class WhisperFullParams extends Structure {
      */
     public Pointer encoder_begin_callback_user_data;
 
+    /** Callback used to abort GGML computation */
+    public Pointer abort_callback;
+
+    /** User data for the abort_callback */
+    public Pointer abort_callback_user_data;
+
+    public void setAbortCallback(GgmlAbortCallback callback) {
+        abort_callback = CallbackReference.getFunctionPointer(callback);
+    }
+
     /**
      * Callback by each decoder to filter obtained logits.
      * WhisperLogitsFilterCallback
@@ -310,17 +331,28 @@ public class WhisperFullParams extends Structure {
 
     @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",
-                "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",
-                "suppress_blank", "suppress_nst", "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",
+        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx",
+                "offset_ms", "duration_ms", "translate", "no_context",
+                "no_timestamps", "single_segment", "print_special",
+                "print_progress", "print_realtime", "print_timestamps",
+                "token_timestamps", "thold_pt", "thold_ptsum", "max_len",
+                "split_on_word", "max_tokens", "debug_mode", "audio_ctx", 
+                "tdrz_enable", "suppress_regex", "initial_prompt",
+                "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
+                "suppress_blank", "suppress_nst", "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",
+                "abort_callback", "abort_callback_user_data",
                 "logits_filter_callback", "logits_filter_callback_user_data",
                 "grammar_rules", "n_grammar_rules", "i_start_rule", "grammar_penalty");
     }
+
+    public static class ByValue extends WhisperFullParams implements Structure.ByValue {
+        public ByValue() { super(); }
+        public ByValue(Pointer p) { super(p); }
+    }
+
 }

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

@@ -76,7 +76,7 @@ class WhisperCppTest {
         float[] floats = new float[b.length / 2];
 
         //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
-        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+        WhisperFullParams.ByValue 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";

bindings/javascript/README.md

@@ -33,6 +33,9 @@ mkdir build-em && cd build-em
 emcmake cmake .. && make -j
 
 # run test
+node ../tests/test-whisper.js
+
+# For Node.js versions prior to v16.4.0, experimental features need to be enabled:
 node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js
 
 # publish npm package

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.7.4",
+  "version": "1.7.5",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/README.md

@@ -24,14 +24,15 @@ require "whisper"
 
 whisper = Whisper::Context.new("base")
 
-params = Whisper::Params.new
-params.language = "en"
-params.offset = 10_000
-params.duration = 60_000
-params.max_text_tokens = 300
-params.translate = true
-params.print_timestamps = false
-params.initial_prompt = "Initial prompt here."
+params = Whisper::Params.new(
+  language: "en",
+  offset: 10_000,
+  duration: 60_000,
+  max_text_tokens: 300,
+  translate: true,
+  print_timestamps: false,
+  initial_prompt: "Initial prompt here."
+)
 
 whisper.transcribe("path/to/audio.wav", params) do |whole_text|
   puts whole_text
@@ -113,18 +114,18 @@ def format_time(time_ms)
   "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
 end
 
-whisper.transcribe("path/to/audio.wav", params)
-
-whisper.each_segment.with_index do |segment, index|
-  line = "[%{nth}: %{st} --> %{ed}] %{text}" % {
-    nth: index + 1,
-    st: format_time(segment.start_time),
-    ed: format_time(segment.end_time),
-    text: segment.text
-  }
-  line << " (speaker turned)" if segment.speaker_next_turn?
-  puts line
-end
+whisper
+  .transcribe("path/to/audio.wav", params)
+  .each_segment.with_index do |segment, index|
+    line = "[%{nth}: %{st} --> %{ed}] %{text}" % {
+      nth: index + 1,
+      st: format_time(segment.start_time),
+      ed: format_time(segment.end_time),
+      text: segment.text
+    }
+    line << " (speaker turned)" if segment.speaker_next_turn?
+    puts line
+  end
 
 ```
 
@@ -215,10 +216,11 @@ reader = WaveFile::Reader.new("path/to/audio.wav", WaveFile::Format.new(:mono, :
 samples = reader.enum_for(:each_buffer).map(&:samples).flatten
 
 whisper = Whisper::Context.new("base")
-whisper.full(Whisper::Params.new, samples)
-whisper.each_segment do |segment|
-  puts segment.text
-end
+whisper
+  .full(Whisper::Params.new, samples)
+  .each_segment do |segment|
+    puts segment.text
+  end
 ```
 
 The second argument `samples` may be an array, an object with `length` and `each` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.

bindings/ruby/Rakefile

@@ -18,9 +18,11 @@ EXTSOURCES.each do |src|
 end
 
 CLEAN.include SOURCES
-CLEAN.include FileList["ext/*.o", "ext/*.metal", "ext/whisper.{so,bundle,dll}"]
+CLEAN.include FileList["ext/**/*.o", "ext/**/*.metal", "ext/**/*.tmp", "ext/whisper.{so,bundle,dll}"]
 
-task build: ["ext/Makefile", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp", "whispercpp.gemspec"]
+SRC = FileList["ext/*.{c,cpp,h}"]
+
+task build: SOURCES
 
 directory "pkg"
 CLOBBER.include "pkg"
@@ -29,14 +31,14 @@ LIB_NAME = "whisper".ext(RbConfig::CONFIG["DLEXT"])
 SO_FILE = File.join("ext", LIB_NAME)
 LIB_FILE = File.join("lib", LIB_NAME)
 
-file "ext/Makefile" => ["ext/extconf.rb", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp"] + SOURCES do |t|
-  Dir.chdir "ext" do
+file "ext/Makefile" => SRC + ["ext/extconf.rb"] + SOURCES do |t|
+  chdir "ext" do
     ruby "extconf.rb"
   end
 end
 
 file SO_FILE => "ext/Makefile" do |t|
-  Dir.chdir "ext" do
+  chdir "ext" do
     sh "make"
   end
 end
@@ -54,7 +56,7 @@ end
 
 TEST_MEMORY_VIEW = "tests/jfk_reader/jfk_reader.#{RbConfig::CONFIG['DLEXT']}"
 file TEST_MEMORY_VIEW => "tests/jfk_reader/jfk_reader.c" do |t|
-  Dir.chdir "tests/jfk_reader" do
+  chdir "tests/jfk_reader" do
     ruby "extconf.rb"
     sh "make"
   end

bindings/ruby/ext/.gitignore

@@ -4,10 +4,8 @@ whisper.bundle
 whisper.dll
 scripts/get-flags.mk
 *.o
-*.c
-*.cpp
-*.h
-*.m
-*.metal
-!ruby_whisper.cpp
-!ruby_whisper.h
+/*/**/*.c
+/*/**/*.cpp
+/*/**/*.h
+/*/**/*.m
+/*/**/*.metal

bindings/ruby/ext/cpu.mk

@@ -1,5 +1,7 @@
 ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
 	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/src/ggml-cpu/unary-ops.cpp \
+	ggml/src/ggml-cpu/binary-ops.cpp \
 	ggml/include/ggml-backend.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-alloc.h \

bindings/ruby/ext/extconf.rb

@@ -35,7 +35,7 @@ if $GGML_METAL
   $GGML_METAL_EMBED_LIBRARY = true
 end
 
-$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples'
+$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples -DGGML_USE_CPU'
 $MK_CFLAGS   = '-std=c11   -fPIC'
 $MK_CXXFLAGS = '-std=c++17 -fPIC'
 $MK_NVCCFLAGS = '-std=c++17'
@@ -168,13 +168,24 @@ $OBJ_GGML <<
   'ggml/src/ggml-cpu/ggml-cpu-aarch64.o' <<
   'ggml/src/ggml-cpu/ggml-cpu-hbm.o' <<
   'ggml/src/ggml-cpu/ggml-cpu-quants.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-traits.o'
+  'ggml/src/ggml-cpu/ggml-cpu-traits.o' <<
+  'ggml/src/ggml-cpu/unary-ops.o' <<
+  'ggml/src/ggml-cpu/binary-ops.o'
 
 $OBJ_WHISPER <<
-  'src/whisper.o'
+  'src/whisper.o' <<
+  'examples/common.o' <<
+  'examples/common-whisper.o'
 
 $objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
-$objs << "ruby_whisper.o"
+$objs <<
+  "ruby_whisper.o" <<
+  "ruby_whisper_context.o" <<
+  "ruby_whisper_transcribe.o" <<
+  "ruby_whisper_params.o" <<
+  "ruby_whisper_error.o" <<
+  "ruby_whisper_segment.o" <<
+  "ruby_whisper_model.o"
 
 $CPPFLAGS  = "#{$MK_CPPFLAGS} #{$CPPFLAGS}"
 $CFLAGS    = "#{$CPPFLAGS} #{$MK_CFLAGS} #{$GF_CFLAGS} #{$CFLAGS}"

bindings/ruby/ext/ruby_whisper.c

@@ -0,0 +1,164 @@
+#include <ruby.h>
+#include <ruby/memory_view.h>
+#include "ruby_whisper.h"
+
+VALUE mWhisper;
+VALUE cContext;
+VALUE cParams;
+VALUE eError;
+
+VALUE cSegment;
+VALUE cModel;
+
+ID id_to_s;
+ID id_call;
+ID id___method__;
+ID id_to_enum;
+ID id_length;
+ID id_next;
+ID id_new;
+ID id_to_path;
+ID id_URI;
+ID id_pre_converted_models;
+
+static bool is_log_callback_finalized = false;
+
+// High level API
+extern VALUE ruby_whisper_segment_allocate(VALUE klass);
+
+extern void init_ruby_whisper_context(VALUE *mWhisper);
+extern void init_ruby_whisper_params(VALUE *mWhisper);
+extern void init_ruby_whisper_error(VALUE *mWhisper);
+extern void init_ruby_whisper_segment(VALUE *mWhisper, VALUE *cSegment);
+extern void init_ruby_whisper_model(VALUE *mWhisper);
+extern void register_callbacks(ruby_whisper_params *rwp, VALUE *context);
+
+/*
+ * call-seq:
+ *   lang_max_id -> Integer
+ */
+static VALUE ruby_whisper_s_lang_max_id(VALUE self) {
+  return INT2NUM(whisper_lang_max_id());
+}
+
+/*
+ * call-seq:
+ *   lang_id(lang_name) -> Integer
+ */
+static VALUE ruby_whisper_s_lang_id(VALUE self, VALUE lang) {
+  const char * lang_str = StringValueCStr(lang);
+  const int id = whisper_lang_id(lang_str);
+  if (-1 == id) {
+    rb_raise(rb_eArgError, "language not found: %s", lang_str);
+  }
+  return INT2NUM(id);
+}
+
+/*
+ * call-seq:
+ *   lang_str(lang_id) -> String
+ */
+static VALUE ruby_whisper_s_lang_str(VALUE self, VALUE id) {
+  const int lang_id = NUM2INT(id);
+  const char * str = whisper_lang_str(lang_id);
+  if (NULL == str) {
+    rb_raise(rb_eIndexError, "id %d outside of language id", lang_id);
+  }
+  return rb_str_new2(str);
+}
+
+/*
+ * call-seq:
+ *   lang_str(lang_id) -> String
+ */
+static VALUE ruby_whisper_s_lang_str_full(VALUE self, VALUE id) {
+  const int lang_id = NUM2INT(id);
+  const char * str_full = whisper_lang_str_full(lang_id);
+  if (NULL == str_full) {
+    rb_raise(rb_eIndexError, "id %d outside of language id", lang_id);
+  }
+  return rb_str_new2(str_full);
+}
+
+static VALUE ruby_whisper_s_finalize_log_callback(VALUE self, VALUE id) {
+  is_log_callback_finalized = true;
+  return Qnil;
+}
+
+static void
+ruby_whisper_log_callback(enum ggml_log_level level, const char * buffer, void * user_data) {
+  if (is_log_callback_finalized) {
+    return;
+  }
+  VALUE log_callback = rb_iv_get(mWhisper, "log_callback");
+  VALUE udata = rb_iv_get(mWhisper, "user_data");
+  rb_funcall(log_callback, id_call, 3, INT2NUM(level), rb_str_new2(buffer), udata);
+}
+
+/*
+ * call-seq:
+ *   log_set ->(level, buffer, user_data) { ... }, user_data -> nil
+ */
+static VALUE ruby_whisper_s_log_set(VALUE self, VALUE log_callback, VALUE user_data) {
+  VALUE old_callback = rb_iv_get(self, "log_callback");
+  if (!NIL_P(old_callback)) {
+    rb_undefine_finalizer(old_callback);
+  }
+
+  rb_iv_set(self, "log_callback", log_callback);
+  rb_iv_set(self, "user_data", user_data);
+
+  VALUE finalize_log_callback = rb_funcall(mWhisper, rb_intern("method"), 1, rb_str_new2("finalize_log_callback"));
+  rb_define_finalizer(log_callback, finalize_log_callback);
+
+  whisper_log_set(ruby_whisper_log_callback, NULL);
+
+  return Qnil;
+}
+
+static void rb_whisper_model_mark(ruby_whisper_model *rwm) {
+  rb_gc_mark(rwm->context);
+}
+
+static VALUE ruby_whisper_model_allocate(VALUE klass) {
+  ruby_whisper_model *rwm;
+  rwm = ALLOC(ruby_whisper_model);
+  return Data_Wrap_Struct(klass, rb_whisper_model_mark, RUBY_DEFAULT_FREE, rwm);
+}
+
+void Init_whisper() {
+  id_to_s = rb_intern("to_s");
+  id_call = rb_intern("call");
+  id___method__ = rb_intern("__method__");
+  id_to_enum = rb_intern("to_enum");
+  id_length = rb_intern("length");
+  id_next = rb_intern("next");
+  id_new = rb_intern("new");
+  id_to_path = rb_intern("to_path");
+  id_URI = rb_intern("URI");
+  id_pre_converted_models = rb_intern("pre_converted_models");
+
+  mWhisper = rb_define_module("Whisper");
+
+  rb_define_const(mWhisper, "LOG_LEVEL_NONE", INT2NUM(GGML_LOG_LEVEL_NONE));
+  rb_define_const(mWhisper, "LOG_LEVEL_INFO", INT2NUM(GGML_LOG_LEVEL_INFO));
+  rb_define_const(mWhisper, "LOG_LEVEL_WARN", INT2NUM(GGML_LOG_LEVEL_WARN));
+  rb_define_const(mWhisper, "LOG_LEVEL_ERROR", INT2NUM(GGML_LOG_LEVEL_ERROR));
+  rb_define_const(mWhisper, "LOG_LEVEL_DEBUG", INT2NUM(GGML_LOG_LEVEL_DEBUG));
+  rb_define_const(mWhisper, "LOG_LEVEL_CONT", INT2NUM(GGML_LOG_LEVEL_CONT));
+
+  rb_define_singleton_method(mWhisper, "lang_max_id", ruby_whisper_s_lang_max_id, 0);
+  rb_define_singleton_method(mWhisper, "lang_id", ruby_whisper_s_lang_id, 1);
+  rb_define_singleton_method(mWhisper, "lang_str", ruby_whisper_s_lang_str, 1);
+  rb_define_singleton_method(mWhisper, "lang_str_full", ruby_whisper_s_lang_str_full, 1);
+  rb_define_singleton_method(mWhisper, "log_set", ruby_whisper_s_log_set, 2);
+  rb_define_private_method(rb_singleton_class(mWhisper), "finalize_log_callback", ruby_whisper_s_finalize_log_callback, 1);
+
+  init_ruby_whisper_context(&mWhisper);
+  init_ruby_whisper_params(&mWhisper);
+  init_ruby_whisper_error(&mWhisper);
+  init_ruby_whisper_segment(&mWhisper, &cContext);
+  init_ruby_whisper_model(&mWhisper);
+
+  rb_require("whisper/model/uri");
+}

bindings/ruby/ext/ruby_whisper.h

@@ -22,4 +22,13 @@ typedef struct {
   ruby_whisper_callback_container *abort_callback_container;
 } ruby_whisper_params;
 
+typedef struct {
+  VALUE context;
+  int index;
+} ruby_whisper_segment;
+
+typedef struct {
+  VALUE context;
+} ruby_whisper_model;
+
 #endif

bindings/ruby/ext/ruby_whisper_context.c

@@ -0,0 +1,613 @@
+#include <ruby.h>
+#include <ruby/memory_view.h>
+#include "ruby_whisper.h"
+
+extern ID id_to_s;
+extern ID id___method__;
+extern ID id_to_enum;
+extern ID id_length;
+extern ID id_next;
+extern ID id_new;
+extern ID id_to_path;
+extern ID id_URI;
+extern ID id_pre_converted_models;
+
+extern VALUE cContext;
+extern VALUE eError;
+extern VALUE cModel;
+
+extern VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self);
+extern VALUE rb_whisper_model_initialize(VALUE context);
+extern VALUE rb_whisper_segment_initialize(VALUE context, int index);
+extern void register_callbacks(ruby_whisper_params *rwp, VALUE *context);
+
+static void
+ruby_whisper_free(ruby_whisper *rw)
+{
+  if (rw->context) {
+    whisper_free(rw->context);
+    rw->context = NULL;
+  }
+}
+
+void
+rb_whisper_mark(ruby_whisper *rw)
+{
+  // call rb_gc_mark on any ruby references in rw
+}
+
+void
+rb_whisper_free(ruby_whisper *rw)
+{
+  ruby_whisper_free(rw);
+  free(rw);
+}
+
+static VALUE
+ruby_whisper_allocate(VALUE klass)
+{
+  ruby_whisper *rw;
+  rw = ALLOC(ruby_whisper);
+  rw->context = NULL;
+  return Data_Wrap_Struct(klass, rb_whisper_mark, rb_whisper_free, rw);
+}
+
+/*
+ * call-seq:
+ *   new("base.en") -> Whisper::Context
+ *   new("path/to/model.bin") -> Whisper::Context
+ *   new(Whisper::Model::URI.new("https://example.net/uri/of/model.bin")) -> Whisper::Context
+ */
+static VALUE
+ruby_whisper_initialize(int argc, VALUE *argv, VALUE self)
+{
+  ruby_whisper *rw;
+  VALUE whisper_model_file_path;
+
+  // TODO: we can support init from buffer here too maybe another ruby object to expose
+  rb_scan_args(argc, argv, "01", &whisper_model_file_path);
+  Data_Get_Struct(self, ruby_whisper, rw);
+
+  VALUE pre_converted_models = rb_funcall(cModel, id_pre_converted_models, 0);
+  VALUE pre_converted_model = rb_hash_aref(pre_converted_models, whisper_model_file_path);
+  if (!NIL_P(pre_converted_model)) {
+    whisper_model_file_path = pre_converted_model;
+  }
+  if (TYPE(whisper_model_file_path) == T_STRING) {
+    const char * whisper_model_file_path_str = StringValueCStr(whisper_model_file_path);
+    if (strncmp("http://", whisper_model_file_path_str, 7) == 0 || strncmp("https://", whisper_model_file_path_str, 8) == 0) {
+      VALUE uri_class = rb_const_get(cModel, id_URI);
+      whisper_model_file_path = rb_class_new_instance(1, &whisper_model_file_path, uri_class);
+    }
+  }
+  if (rb_obj_is_kind_of(whisper_model_file_path, rb_path2class("URI::HTTP"))) {
+    VALUE uri_class = rb_const_get(cModel, id_URI);
+    whisper_model_file_path = rb_class_new_instance(1, &whisper_model_file_path, uri_class);
+  }
+  if (rb_respond_to(whisper_model_file_path, id_to_path)) {
+    whisper_model_file_path = rb_funcall(whisper_model_file_path, id_to_path, 0);
+  }
+  if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
+    rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
+  }
+  rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), whisper_context_default_params());
+  if (rw->context == NULL) {
+    rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
+  }
+  return self;
+}
+
+/*
+ * call-seq:
+ *   model_n_vocab -> Integer
+ */
+VALUE ruby_whisper_model_n_vocab(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_vocab(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_ctx -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_ctx(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_state -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_state(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_head -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_head(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_audio_layer -> Integer
+ */
+VALUE ruby_whisper_model_n_audio_layer(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_ctx -> Integer
+ */
+VALUE ruby_whisper_model_n_text_ctx(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_state -> Integer
+ */
+VALUE ruby_whisper_model_n_text_state(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_head -> Integer
+ */
+VALUE ruby_whisper_model_n_text_head(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_text_layer -> Integer
+ */
+VALUE ruby_whisper_model_n_text_layer(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_n_mels -> Integer
+ */
+VALUE ruby_whisper_model_n_mels(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_mels(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_ftype -> Integer
+ */
+VALUE ruby_whisper_model_ftype(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_model_ftype(rw->context));
+}
+
+/*
+ * call-seq:
+ *   model_type -> String
+ */
+VALUE ruby_whisper_model_type(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return rb_str_new2(whisper_model_type_readable(rw->context));
+}
+
+/*
+ * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+ * Not thread safe for same context
+ * Uses the specified decoding strategy to obtain the text.
+ *
+ * call-seq:
+ *   full(params, samples, n_samples) -> nil
+ *   full(params, samples) -> nil
+ *
+ * The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
+ */
+VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self)
+{
+  if (argc < 2 || argc > 3) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  if (argc == 3) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (memory_view_available_p) {
+      if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+        view.obj = Qnil;
+        rb_raise(rb_eArgError, "unable to get a memory view");
+      }
+      n_samples = view.byte_size / view.item_size;
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p)  {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // TODO: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError appropriately
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  register_callbacks(rwp, &self);
+  const int result = whisper_full(rw->context, rwp->params, c_samples, n_samples);
+  if (0 == result) {
+    return self;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
+/*
+ * Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+ * Result is stored in the default state of the context
+ * Not thread safe if executed in parallel on the same context.
+ * It seems this approach can offer some speedup in some cases.
+ * However, the transcription accuracy can be worse at the beginning and end of each chunk.
+ *
+ * call-seq:
+ *   full_parallel(params, samples) -> nil
+ *   full_parallel(params, samples, n_samples) -> nil
+ *   full_parallel(params, samples, n_samples, n_processors) -> nil
+ *   full_parallel(params, samples, nil, n_processors) -> nil
+ */
+static VALUE
+ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self)
+{
+  if (argc < 2 || argc > 4) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  int n_processors;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  switch (argc) {
+  case 2:
+    n_processors = 1;
+    break;
+  case 3:
+    n_processors = 1;
+    break;
+  case 4:
+    n_processors = NUM2INT(argv[3]);
+    break;
+  }
+  if (argc >= 3 && !NIL_P(argv[2])) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else if (memory_view_available_p) {
+    if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+      view.obj = Qnil;
+      rb_raise(rb_eArgError, "unable to get a memory view");
+    }
+    n_samples = view.byte_size / view.item_size;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p) {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // FIXME: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  register_callbacks(rwp, &self);
+  const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
+  if (0 == result) {
+    return self;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
+/*
+ * Number of segments.
+ *
+ * call-seq:
+ *   full_n_segments -> Integer
+ */
+static VALUE
+ruby_whisper_full_n_segments(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_full_n_segments(rw->context));
+}
+
+/*
+ * Language ID, which can be converted to string by Whisper.lang_str and Whisper.lang_str_full.
+ *
+ * call-seq:
+ *   full_lang_id -> Integer
+ */
+static VALUE
+ruby_whisper_full_lang_id(VALUE self)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  return INT2NUM(whisper_full_lang_id(rw->context));
+}
+
+static int ruby_whisper_full_check_segment_index(const ruby_whisper * rw, const VALUE i_segment)
+{
+  const int c_i_segment = NUM2INT(i_segment);
+  if (c_i_segment < 0 || c_i_segment >= whisper_full_n_segments(rw->context)) {
+    rb_raise(rb_eIndexError, "segment index %d out of range", c_i_segment);
+  }
+  return c_i_segment;
+}
+
+/*
+ * Start time of a segment indexed by +segment_index+ in centiseconds (10 times milliseconds).
+ *
+ *   full_get_segment_t0(3) # => 1668 (16680 ms)
+ *
+ * call-seq:
+ *   full_get_segment_t0(segment_index) -> Integer
+ */
+static VALUE
+ruby_whisper_full_get_segment_t0(VALUE self, VALUE i_segment)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const int64_t t0 = whisper_full_get_segment_t0(rw->context, c_i_segment);
+  return INT2NUM(t0);
+}
+
+/*
+ * End time of a segment indexed by +segment_index+ in centiseconds (10 times milliseconds).
+ *
+ *   full_get_segment_t1(3) # => 1668 (16680 ms)
+ *
+ * call-seq:
+ *   full_get_segment_t1(segment_index) -> Integer
+ */
+static VALUE
+ruby_whisper_full_get_segment_t1(VALUE self, VALUE i_segment)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const int64_t t1 = whisper_full_get_segment_t1(rw->context, c_i_segment);
+  return INT2NUM(t1);
+}
+
+/*
+ * Whether the next segment indexed by +segment_index+ is predicated as a speaker turn.
+ *
+ *   full_get_segment_speacker_turn_next(3) # => true
+ *
+ * call-seq:
+ *   full_get_segment_speacker_turn_next(segment_index) -> bool
+ */
+static VALUE
+ruby_whisper_full_get_segment_speaker_turn_next(VALUE self, VALUE i_segment)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const bool speaker_turn_next = whisper_full_get_segment_speaker_turn_next(rw->context, c_i_segment);
+  return speaker_turn_next ? Qtrue : Qfalse;
+}
+
+/*
+ * Text of a segment indexed by +segment_index+.
+ *
+ *   full_get_segment_text(3) # => "ask not what your country can do for you, ..."
+ *
+ * call-seq:
+ *   full_get_segment_text(segment_index) -> String
+ */
+static VALUE
+ruby_whisper_full_get_segment_text(VALUE self, VALUE i_segment)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const char * text = whisper_full_get_segment_text(rw->context, c_i_segment);
+  return rb_str_new2(text);
+}
+
+/*
+ * call-seq:
+ *   full_get_segment_no_speech_prob(segment_index) -> Float
+ */
+static VALUE
+ruby_whisper_full_get_segment_no_speech_prob(VALUE self, VALUE i_segment)
+{
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const float no_speech_prob = whisper_full_get_segment_no_speech_prob(rw->context, c_i_segment);
+  return DBL2NUM(no_speech_prob);
+}
+
+// High level API
+
+static VALUE
+ruby_whisper_full_get_segment(VALUE self, VALUE i_segment)
+{
+  return rb_whisper_segment_initialize(self, NUM2INT(i_segment));
+}
+
+/*
+ * Yields each Whisper::Segment:
+ *
+ *   whisper.transcribe("path/to/audio.wav", params)
+ *   whisper.each_segment do |segment|
+ *     puts segment.text
+ *   end
+ *
+ * Returns an Enumerator if no block given:
+ *
+ *   whisper.transcribe("path/to/audio.wav", params)
+ *   enum = whisper.each_segment
+ *   enum.to_a # => [#<Whisper::Segment>, ...]
+ *
+ * call-seq:
+ *   each_segment {|segment| ... }
+ *   each_segment -> Enumerator
+ */
+static VALUE
+ruby_whisper_each_segment(VALUE self)
+{
+  if (!rb_block_given_p()) {
+    const VALUE method_name = rb_funcall(self, id___method__, 0);
+    return rb_funcall(self, id_to_enum, 1, method_name);
+  }
+
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+
+  const int n_segments = whisper_full_n_segments(rw->context);
+  for (int i = 0; i < n_segments; ++i) {
+    rb_yield(rb_whisper_segment_initialize(self, i));
+  }
+
+  return self;
+}
+
+/*
+ * call-seq:
+ *   model -> Whisper::Model
+ */
+static VALUE
+ruby_whisper_get_model(VALUE self)
+{
+  return rb_whisper_model_initialize(self);
+}
+
+void
+init_ruby_whisper_context(VALUE *mWhisper)
+{
+  cContext = rb_define_class_under(*mWhisper, "Context", rb_cObject);
+
+  rb_define_alloc_func(cContext, ruby_whisper_allocate);
+  rb_define_method(cContext, "initialize", ruby_whisper_initialize, -1);
+
+  rb_define_method(cContext, "transcribe", ruby_whisper_transcribe, -1);
+  rb_define_method(cContext, "model_n_vocab", ruby_whisper_model_n_vocab, 0);
+  rb_define_method(cContext, "model_n_audio_ctx", ruby_whisper_model_n_audio_ctx, 0);
+  rb_define_method(cContext, "model_n_audio_state", ruby_whisper_model_n_audio_state, 0);
+  rb_define_method(cContext, "model_n_audio_head", ruby_whisper_model_n_audio_head, 0);
+  rb_define_method(cContext, "model_n_audio_layer", ruby_whisper_model_n_audio_layer, 0);
+  rb_define_method(cContext, "model_n_text_ctx", ruby_whisper_model_n_text_ctx, 0);
+  rb_define_method(cContext, "model_n_text_state", ruby_whisper_model_n_text_state, 0);
+  rb_define_method(cContext, "model_n_text_head", ruby_whisper_model_n_text_head, 0);
+  rb_define_method(cContext, "model_n_text_layer", ruby_whisper_model_n_text_layer, 0);
+  rb_define_method(cContext, "model_n_mels", ruby_whisper_model_n_mels, 0);
+  rb_define_method(cContext, "model_ftype", ruby_whisper_model_ftype, 0);
+  rb_define_method(cContext, "model_type", ruby_whisper_model_type, 0);
+  rb_define_method(cContext, "full_n_segments", ruby_whisper_full_n_segments, 0);
+  rb_define_method(cContext, "full_lang_id", ruby_whisper_full_lang_id, 0);
+  rb_define_method(cContext, "full_get_segment_t0", ruby_whisper_full_get_segment_t0, 1);
+  rb_define_method(cContext, "full_get_segment_t1", ruby_whisper_full_get_segment_t1, 1);
+  rb_define_method(cContext, "full_get_segment_speaker_turn_next", ruby_whisper_full_get_segment_speaker_turn_next, 1);
+  rb_define_method(cContext, "full_get_segment_text", ruby_whisper_full_get_segment_text, 1);
+  rb_define_method(cContext, "full_get_segment_no_speech_prob", ruby_whisper_full_get_segment_no_speech_prob, 1);
+  rb_define_method(cContext, "full", ruby_whisper_full, -1);
+  rb_define_method(cContext, "full_parallel", ruby_whisper_full_parallel, -1);
+
+  // High leve
+  rb_define_method(cContext, "full_get_segment", ruby_whisper_full_get_segment, 1);
+  rb_define_method(cContext, "each_segment", ruby_whisper_each_segment, 0);
+
+  rb_define_method(cContext, "model", ruby_whisper_get_model, 0);
+}

bindings/ruby/ext/ruby_whisper_error.c

@@ -0,0 +1,52 @@
+#include <ruby.h>
+
+extern VALUE eError;
+
+VALUE ruby_whisper_error_initialize(VALUE self, VALUE code)
+{
+  const int c_code = NUM2INT(code);
+  const char *raw_message;
+  switch (c_code) {
+  case -2:
+    raw_message = "failed to compute log mel spectrogram";
+    break;
+  case -3:
+    raw_message = "failed to auto-detect language";
+    break;
+  case -4:
+    raw_message = "too many decoders requested";
+    break;
+  case -5:
+    raw_message = "audio_ctx is larger than the maximum allowed";
+    break;
+  case -6:
+    raw_message = "failed to encode";
+    break;
+  case -7:
+    raw_message = "whisper_kv_cache_init() failed for self-attention cache";
+    break;
+  case -8:
+    raw_message = "failed to decode";
+    break;
+  case -9:
+    raw_message = "failed to decode";
+    break;
+  default:
+    raw_message = "unknown error";
+    break;
+  }
+  const VALUE message = rb_str_new2(raw_message);
+  rb_call_super(1, &message);
+  rb_iv_set(self, "@code", code);
+
+  return self;
+}
+
+void
+init_ruby_whisper_error(VALUE *mWhisper)
+{
+  eError = rb_define_class_under(*mWhisper, "Error", rb_eStandardError);
+
+  rb_define_attr(eError, "code", true, false);
+  rb_define_method(eError, "initialize", ruby_whisper_error_initialize, 1);
+}

bindings/ruby/ext/ruby_whisper_model.c

@@ -0,0 +1,210 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+
+extern VALUE cModel;
+
+static void rb_whisper_model_mark(ruby_whisper_model *rwm) {
+  rb_gc_mark(rwm->context);
+}
+
+static VALUE ruby_whisper_model_allocate(VALUE klass) {
+  ruby_whisper_model *rwm;
+  rwm = ALLOC(ruby_whisper_model);
+  return Data_Wrap_Struct(klass, rb_whisper_model_mark, RUBY_DEFAULT_FREE, rwm);
+}
+
+VALUE rb_whisper_model_initialize(VALUE context) {
+  ruby_whisper_model *rwm;
+  const VALUE model = ruby_whisper_model_allocate(cModel);
+  Data_Get_Struct(model, ruby_whisper_model, rwm);
+  rwm->context = context;
+  return model;
+};
+
+/*
+ * call-seq:
+ *   n_vocab -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_vocab(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_vocab(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_ctx -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_audio_ctx(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_state -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_audio_state(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_head -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_audio_head(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_audio_layer -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_audio_layer(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_audio_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_ctx -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_text_ctx(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_ctx(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_state -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_text_state(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_state(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_head -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_text_head(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_head(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_text_layer -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_text_layer(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_text_layer(rw->context));
+}
+
+/*
+ * call-seq:
+ *   n_mels -> Integer
+ */
+static VALUE
+ruby_whisper_model_n_mels(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_n_mels(rw->context));
+}
+
+/*
+ * call-seq:
+ *   ftype -> Integer
+ */
+static VALUE
+ruby_whisper_model_ftype(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return INT2NUM(whisper_model_ftype(rw->context));
+}
+
+/*
+ * call-seq:
+ *   type -> String
+ */
+static VALUE
+ruby_whisper_model_type(VALUE self)
+{
+  ruby_whisper_model *rwm;
+  Data_Get_Struct(self, ruby_whisper_model, rwm);
+  ruby_whisper *rw;
+  Data_Get_Struct(rwm->context, ruby_whisper, rw);
+  return rb_str_new2(whisper_model_type_readable(rw->context));
+}
+
+void
+init_ruby_whisper_model(VALUE *mWhisper)
+{
+  cModel = rb_define_class_under(*mWhisper, "Model", rb_cObject);
+
+  rb_define_alloc_func(cModel, ruby_whisper_model_allocate);
+  rb_define_method(cModel, "n_vocab", ruby_whisper_model_n_vocab, 0);
+  rb_define_method(cModel, "n_audio_ctx", ruby_whisper_model_n_audio_ctx, 0);
+  rb_define_method(cModel, "n_audio_state", ruby_whisper_model_n_audio_state, 0);
+  rb_define_method(cModel, "n_audio_head", ruby_whisper_model_n_audio_head, 0);
+  rb_define_method(cModel, "n_audio_layer", ruby_whisper_model_n_audio_layer, 0);
+  rb_define_method(cModel, "n_text_ctx", ruby_whisper_model_n_text_ctx, 0);
+  rb_define_method(cModel, "n_text_state", ruby_whisper_model_n_text_state, 0);
+  rb_define_method(cModel, "n_text_head", ruby_whisper_model_n_text_head, 0);
+  rb_define_method(cModel, "n_text_layer", ruby_whisper_model_n_text_layer, 0);
+  rb_define_method(cModel, "n_mels", ruby_whisper_model_n_mels, 0);
+  rb_define_method(cModel, "ftype", ruby_whisper_model_ftype, 0);
+  rb_define_method(cModel, "type", ruby_whisper_model_type, 0);
+}

bindings/ruby/ext/ruby_whisper_segment.c

@@ -0,0 +1,123 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+
+extern VALUE cSegment;
+
+static void
+rb_whisper_segment_mark(ruby_whisper_segment *rws)
+{
+  rb_gc_mark(rws->context);
+}
+
+VALUE
+ruby_whisper_segment_allocate(VALUE klass)
+{
+  ruby_whisper_segment *rws;
+  rws = ALLOC(ruby_whisper_segment);
+  return Data_Wrap_Struct(klass, rb_whisper_segment_mark, RUBY_DEFAULT_FREE, rws);
+}
+
+VALUE
+rb_whisper_segment_initialize(VALUE context, int index)
+{
+  ruby_whisper_segment *rws;
+  const VALUE segment = ruby_whisper_segment_allocate(cSegment);
+  Data_Get_Struct(segment, ruby_whisper_segment, rws);
+  rws->context = context;
+  rws->index = index;
+  return segment;
+};
+
+/*
+ * Start time in milliseconds.
+ *
+ * call-seq:
+ *   start_time -> Integer
+ */
+static VALUE
+ruby_whisper_segment_get_start_time(VALUE self)
+{
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  const int64_t t0 = whisper_full_get_segment_t0(rw->context, rws->index);
+  // able to multiply 10 without overflow because to_timestamp() in whisper.cpp does it
+  return INT2NUM(t0 * 10);
+}
+
+/*
+ * End time in milliseconds.
+ *
+ * call-seq:
+ *   end_time -> Integer
+ */
+static VALUE
+ruby_whisper_segment_get_end_time(VALUE self)
+{
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  const int64_t t1 = whisper_full_get_segment_t1(rw->context, rws->index);
+  // able to multiply 10 without overflow because to_timestamp() in whisper.cpp does it
+  return INT2NUM(t1 * 10);
+}
+
+/*
+ * Whether the next segment is predicted as a speaker turn.
+ *
+ * call-seq:
+ *   speaker_turn_next? -> bool
+ */
+static VALUE
+ruby_whisper_segment_get_speaker_turn_next(VALUE self)
+{
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  return whisper_full_get_segment_speaker_turn_next(rw->context, rws->index) ? Qtrue : Qfalse;
+}
+
+/*
+ * call-seq:
+ *   text -> String
+ */
+static VALUE
+ruby_whisper_segment_get_text(VALUE self)
+{
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  const char * text = whisper_full_get_segment_text(rw->context, rws->index);
+  return rb_str_new2(text);
+}
+
+/*
+ * call-seq:
+ *   no_speech_prob -> Float
+ */
+static VALUE
+ruby_whisper_segment_get_no_speech_prob(VALUE self)
+{
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  return DBL2NUM(whisper_full_get_segment_no_speech_prob(rw->context, rws->index));
+}
+
+void
+init_ruby_whisper_segment(VALUE *mWhisper, VALUE *cContext)
+{
+  cSegment  = rb_define_class_under(*mWhisper, "Segment", rb_cObject);
+
+  rb_define_alloc_func(cSegment, ruby_whisper_segment_allocate);
+  rb_define_method(cSegment, "start_time", ruby_whisper_segment_get_start_time, 0);
+  rb_define_method(cSegment, "end_time", ruby_whisper_segment_get_end_time, 0);
+  rb_define_method(cSegment, "speaker_next_turn?", ruby_whisper_segment_get_speaker_turn_next, 0);
+  rb_define_method(cSegment, "text", ruby_whisper_segment_get_text, 0);
+  rb_define_method(cSegment, "no_speech_prob", ruby_whisper_segment_get_no_speech_prob, 0);
+}

bindings/ruby/ext/ruby_whisper_transcribe.cpp

@@ -0,0 +1,83 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+#include "common-whisper.h"
+#include <string>
+#include <vector>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern ID id_to_s;
+extern ID id_call;
+
+extern void
+register_callbacks(ruby_whisper_params * rwp, VALUE * self);
+
+/*
+ * transcribe a single file
+ * can emit to a block results
+ *
+ *   params = Whisper::Params.new
+ *   params.duration = 60_000
+ *   whisper.transcribe "path/to/audio.wav", params do |text|
+ *     puts text
+ *   end
+ *
+ * call-seq:
+ *   transcribe(path_to_audio, params) {|text| ...}
+ **/
+VALUE
+ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  VALUE wave_file_path, blk, params;
+
+  rb_scan_args(argc, argv, "02&", &wave_file_path, &params, &blk);
+  Data_Get_Struct(self, ruby_whisper, rw);
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+
+  if (!rb_respond_to(wave_file_path, id_to_s)) {
+    rb_raise(rb_eRuntimeError, "Expected file path to wave file");
+  }
+
+  std::string fname_inp = StringValueCStr(wave_file_path);
+
+  std::vector<float> pcmf32; // mono-channel F32 PCM
+  std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+  if (!read_audio_data(fname_inp, pcmf32, pcmf32s, rwp->diarize)) {
+    fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
+    return self;
+  }
+  {
+    static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+      bool is_aborted = *(bool*)user_data;
+      return !is_aborted;
+    };
+    rwp->params.encoder_begin_callback_user_data = &is_aborted;
+  }
+
+  register_callbacks(rwp, &self);
+
+  if (whisper_full_parallel(rw->context, rwp->params, pcmf32.data(), pcmf32.size(), 1) != 0) {
+    fprintf(stderr, "failed to process audio\n");
+    return self;
+  }
+  const int n_segments = whisper_full_n_segments(rw->context);
+  VALUE output = rb_str_new2("");
+  for (int i = 0; i < n_segments; ++i) {
+    const char * text = whisper_full_get_segment_text(rw->context, i);
+    output = rb_str_concat(output, rb_str_new2(text));
+  }
+  VALUE idCall = id_call;
+  if (blk != Qnil) {
+    rb_funcall(blk, idCall, 1, output);
+  }
+  return self;
+}
+#ifdef __cplusplus
+}
+#endif

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

@@ -65,6 +65,13 @@ module Whisper
             end
           end
         end
+      rescue => err
+        if cache_path.exist?
+          warn err
+        # Use cache file
+        else
+          raise
+        end
       end
 
       def download(response)

bindings/ruby/sig/whisper.rbs

@@ -20,13 +20,12 @@ module Whisper
   def self.lang_id: (string name) -> Integer
   def self.lang_str: (Integer id) -> String
   def self.lang_str_full: (Integer id) -> String
-  def self.log_set=: (log_callback) -> log_callback
-  def self.finalize_log_callback: (void) -> void # Second argument of ObjectSpace.define_finalizer
+  def self.log_set: (log_callback, Object? user_data) -> log_callback
 
   class Context
-    def initialize: (string | _ToPath | ::URI::HTTP ) -> void
-    def transcribe: (string, Params) -> void
-                  | (string, Params) { (String) -> void } -> void
+    def self.new: (string | _ToPath | ::URI::HTTP) -> instance
+    def transcribe: (string, Params) -> self
+                  | (string, Params) { (String) -> void } -> self
     def model_n_vocab: () -> Integer
     def model_n_audio_ctx: () -> Integer
     def model_n_audio_state: () -> Integer
@@ -35,6 +34,10 @@ module Whisper
     def model_n_mels: () -> Integer
     def model_ftype: () -> Integer
     def model_type: () -> String
+    def each_segment: { (Segment) -> void } -> void
+                    | () -> Enumerator[Segment]
+    def model: () -> Model
+    def full_get_segment: (Integer nth) -> Segment
     def full_n_segments: () -> Integer
     def full_lang_id: () -> Integer
     def full_get_segment_t0: (Integer) -> Integer
@@ -42,18 +45,46 @@ module Whisper
     def full_get_segment_speaker_turn_next: (Integer) -> (true | false)
     def full_get_segment_text: (Integer) -> String
     def full_get_segment_no_speech_prob: (Integer) -> Float
-    def full: (Params, Array[Float], ?Integer) -> void
-            | (Params, _Samples, ?Integer) -> void
-    def full_parallel: (Params, Array[Float], ?Integer) -> void
-                     | (Params, _Samples, ?Integer) -> void
-                     | (Params, _Samples, ?Integer?, Integer) -> void
-    def each_segment: { (Segment) -> void } -> void
-                    | () -> Enumerator[Segment]
-    def model: () -> Model
+    def full: (Params, Array[Float] samples, ?Integer n_samples) -> self
+            | (Params, _Samples, ?Integer n_samples) -> self
+    def full_parallel: (Params, Array[Float], ?Integer n_samples) -> self
+                     | (Params, _Samples, ?Integer n_samples) -> self
+                     | (Params, _Samples, ?Integer? n_samples, Integer n_processors) -> self
   end
 
   class Params
-    def initialize: () -> void
+    def self.new: (
+      ?language: string,
+      ?translate: boolish,
+      ?no_context: boolish,
+      ?single_segment: boolish,
+      ?print_special: boolish,
+      ?print_progress: boolish,
+      ?print_realtime: boolish,
+      ?print_timestamps: boolish,
+      ?suppress_blank: boolish,
+      ?suppress_nst: boolish,
+      ?token_timestamps: boolish,
+      ?split_on_word: boolish,
+      ?initial_prompt: string | nil,
+      ?diarize: boolish,
+      ?offset: Integer,
+      ?duration: Integer,
+      ?max_text_tokens: Integer,
+      ?temperature: Float,
+      ?max_initial_ts: Float,
+      ?length_penalty: Float,
+      ?temperature_inc: Float,
+      ?entropy_thold: Float,
+      ?logprob_thold: Float,
+      ?no_speech_thold: Float,
+      ?new_segment_callback: new_segment_callback,
+      ?new_segment_callback_user_data: Object,
+      ?progress_callback: progress_callback,
+      ?progress_callback_user_data: Object,
+      ?abort_callback: abort_callback,
+      ?abort_callback_user_data: Object
+    ) -> instance
     def language=: (String) -> String # TODO: Enumerate lang names
     def language: () -> String
     def translate=: (boolish) -> boolish
@@ -79,7 +110,7 @@ module Whisper
     def split_on_word=: (boolish) -> boolish
     def split_on_word: () -> (true | false)
     def initial_prompt=: (_ToS) -> _ToS
-    def initial_prompt: () -> String
+    def initial_prompt: () -> (String | nil)
     def diarize=: (boolish) -> boolish
     def diarize: () -> (true | false)
     def offset=: (Integer) -> Integer
@@ -103,19 +134,25 @@ module Whisper
     def no_speech_thold=: (Float) -> Float
     def no_speech_thold: () -> Float
     def new_segment_callback=: (new_segment_callback) -> new_segment_callback
+    def new_segment_callback: () -> (new_segment_callback | nil)
     def new_segment_callback_user_data=: (Object) -> Object
+    def new_segment_callback_user_data: () -> Object
     def progress_callback=: (progress_callback) -> progress_callback
+    def progress_callback: () -> (progress_callback | nil)
     def progress_callback_user_data=: (Object) -> Object
+    def progress_callback_user_data: () -> Object
     def abort_callback=: (abort_callback) -> abort_callback
+    def abort_callback: () -> (abort_callback | nil)
     def abort_callback_user_data=: (Object) -> Object
+    def abort_callback_user_data: () -> Object
     def on_new_segment: { (Segment) -> void } -> void
-    def on_progress: { (Integer) -> void } -> void
-    def abort_on: { (Object) -> boolish } -> void
+    def on_progress: { (Integer progress) -> void } -> void
+    def abort_on: { (Object user_data) -> boolish } -> void
   end
 
   class Model
     def self.pre_converted_models: () -> Hash[String, Model::URI]
-    def initialize: () -> void
+    def self.new: () -> instance
     def n_vocab: () -> Integer
     def n_audio_ctx: () -> Integer
     def n_audio_state: () -> Integer
@@ -130,14 +167,13 @@ module Whisper
     def type: () -> String
 
     class URI
-      def initialize: (string | ::URI::HTTP) -> void
+      def self.new: (string | ::URI::HTTP) -> self
       def to_path: -> String
       def clear_cache: -> void
     end
   end
 
   class Segment
-    def initialize: () -> void
     def start_time: () -> Integer
     def end_time: () -> Integer
     def speaker_next_turn?: () -> (true | false)
@@ -148,6 +184,6 @@ module Whisper
   class Error < StandardError
     attr_reader code: Integer
 
-    def initialize: (Integer) -> void
+    def self.new: (Integer code) -> instance
   end
 end

bindings/ruby/tests/test_callback.rb

@@ -25,7 +25,7 @@ class TestCallback < TestBase
         assert start_time >= 0
         assert_kind_of Integer, end_time
         assert end_time > 0
-        assert_match /ask not what your country can do for you, ask what you can do for your country/, text if i_segment == 0
+        assert_match(/ask not what your country can do for you, ask what you can do for your country/, text) if i_segment == 0
       end
     }
 
@@ -145,9 +145,9 @@ class TestCallback < TestBase
 
   def test_abort_on
     do_abort = false
-    aborted_from_callback = false
+    _aborted_from_callback = false
     @params.on_new_segment do |segment|
-      do_abort = true if segment.text.match? /ask/
+      do_abort = true if segment.text.match?(/ask/)
     end
     i = 0
     @params.abort_on do

bindings/ruby/tests/test_error.rb

@@ -4,7 +4,7 @@ class TestError < TestBase
   def test_error
     error = Whisper::Error.new(-2)
     assert_equal "failed to compute log mel spectrogram", error.message
-    assert_equal -2, error.code
+    assert_equal(-2, error.code)
   end
 
   def test_unknown_error
@@ -14,7 +14,7 @@ class TestError < TestBase
 
   def test_non_int_code
     assert_raise TypeError do
-      error = Whisper::Error.new("non int")
+      _error = Whisper::Error.new("non int")
     end
   end
 end

bindings/ruby/tests/test_params.rb

@@ -1,6 +1,39 @@
 require_relative "helper"
 
 class TestParams < TestBase
+  PARAM_NAMES = [
+    :language,
+    :translate,
+    :no_context,
+    :single_segment,
+    :print_special,
+    :print_progress,
+    :print_realtime,
+    :print_timestamps,
+    :suppress_blank,
+    :suppress_nst,
+    :token_timestamps,
+    :split_on_word,
+    :initial_prompt,
+    :diarize,
+    :offset,
+    :duration,
+    :max_text_tokens,
+    :temperature,
+    :max_initial_ts,
+    :length_penalty,
+    :temperature_inc,
+    :entropy_thold,
+    :logprob_thold,
+    :no_speech_thold,
+    :new_segment_callback,
+    :new_segment_callback_user_data,
+    :progress_callback,
+    :progress_callback_user_data,
+    :abort_callback,
+    :abort_callback_user_data,
+  ]
+
   def setup
     @params  = Whisper::Params.new
   end
@@ -129,7 +162,7 @@ class TestParams < TestBase
   end
 
   def test_length_penalty
-    assert_equal -1.0, @params.length_penalty
+    assert_equal(-1.0, @params.length_penalty)
     @params.length_penalty = 0.5
     assert_equal 0.5, @params.length_penalty
   end
@@ -147,9 +180,9 @@ class TestParams < TestBase
   end
 
   def test_logprob_thold
-    assert_in_delta -1.0, @params.logprob_thold
+    assert_in_delta(-1.0, @params.logprob_thold)
     @params.logprob_thold = -0.5
-    assert_in_delta -0.5, @params.logprob_thold
+    assert_in_delta(-0.5, @params.logprob_thold)
   end
 
   def test_no_speech_thold
@@ -157,4 +190,57 @@ class TestParams < TestBase
     @params.no_speech_thold = 0.2
     assert_in_delta 0.2, @params.no_speech_thold
   end
+
+  def test_new_with_kw_args
+    params = Whisper::Params.new(language: "es")
+    assert_equal "es", params.language
+    assert_equal 1.0, params.max_initial_ts
+  end
+
+  def test_new_with_kw_args_non_existent
+    assert_raise ArgumentError do
+      Whisper::Params.new(non_existent: "value")
+    end
+  end
+
+  def test_new_with_kw_args_wrong_type
+    assert_raise TypeError do
+      Whisper::Params.new(language: 3)
+    end
+  end
+
+  data(PARAM_NAMES.collect {|param| [param, param]}.to_h)
+  def test_new_with_kw_args_default_values(param)
+    default_value = @params.send(param)
+    value = case [param, default_value]
+            in [*, true | false]
+              !default_value
+            in [*, Integer | Float]
+              default_value + 1
+            in [:language, *]
+              "es"
+            in [:initial_prompt, *]
+              "Initial prompt"
+            in [/_callback\Z/, *]
+              proc {}
+            in [/_user_data\Z/, *]
+              Object.new
+            end
+    params = Whisper::Params.new(param => value)
+    if Float === value
+      assert_in_delta value, params.send(param)
+    else
+      assert_equal value, params.send(param)
+    end
+
+    PARAM_NAMES.reject {|name| name == param}.each do |name|
+      expected = @params.send(name)
+      actual = params.send(name)
+      if Float === expected
+        assert_in_delta expected, actual
+      else
+        assert_equal expected, actual
+      end
+    end
+  end
 end

bindings/ruby/tests/test_segment.rb

@@ -49,13 +49,13 @@ class TestSegment < TestBase
       if index == 0
         seg = segment
         assert_equal 0, segment.start_time
-        assert_match /ask not what your country can do for you, ask what you can do for your country/, segment.text
+        assert_match(/ask not what your country can do for you, ask what you can do for your country/, segment.text)
       end
       index += 1
     end
     whisper.transcribe(AUDIO, params)
     assert_equal 0, seg.start_time
-    assert_match /ask not what your country can do for you, ask what you can do for your country/, seg.text
+    assert_match(/ask not what your country can do for you, ask what you can do for your country/, seg.text)
   end
 
   def test_on_new_segment_twice

bindings/ruby/tests/test_whisper.rb

@@ -16,7 +16,7 @@ class TestWhisper < TestBase
     params.print_timestamps = false
 
     @whisper.transcribe(AUDIO, params) {|text|
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, text
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, text)
     }
   end
 
@@ -29,6 +29,12 @@ class TestWhisper < TestBase
       assert_equal 0, whisper.full_lang_id
     end
 
+    def test_full_get_segment
+      segment = whisper.full_get_segment(0)
+      assert_equal 0, segment.start_time
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, segment.text)
+    end
+
     def test_full_get_segment_t0
       assert_equal 0, whisper.full_get_segment_t0(0)
       assert_raise IndexError do
@@ -53,7 +59,7 @@ class TestWhisper < TestBase
     end
 
     def test_full_get_segment_text
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, whisper.full_get_segment_text(0)
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, whisper.full_get_segment_text(0))
     end
 
     def test_full_get_segment_no_speech_prob
@@ -128,14 +134,14 @@ class TestWhisper < TestBase
       @whisper.full(@params, @samples, @samples.length)
 
       assert_equal 1, @whisper.full_n_segments
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text)
     end
 
     def test_full_without_length
       @whisper.full(@params, @samples)
 
       assert_equal 1, @whisper.full_n_segments
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text)
     end
 
     def test_full_enumerator
@@ -143,7 +149,7 @@ class TestWhisper < TestBase
       @whisper.full(@params, samples, @samples.length)
 
       assert_equal 1, @whisper.full_n_segments
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text)
     end
 
     def test_full_enumerator_without_length
@@ -165,26 +171,28 @@ class TestWhisper < TestBase
       @whisper.full(@params, samples)
 
       assert_equal 1, @whisper.full_n_segments
-      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+      assert_match(/ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text)
     end
 
     def test_full_parallel
-      @whisper.full_parallel(@params, @samples, @samples.length, Etc.nprocessors)
+      nprocessors = 2
+      @whisper.full_parallel(@params, @samples, @samples.length, nprocessors)
 
-      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      assert_equal nprocessors, @whisper.full_n_segments
       text = @whisper.each_segment.collect(&:text).join
-      assert_match /ask what you can do/i, text
-      assert_match /for your country/i, text
+      assert_match(/ask what you can do/i, text)
+      assert_match(/for your country/i, text)
     end
 
     def test_full_parallel_with_memory_view
+      nprocessors = 2
       samples = JFKReader.new(AUDIO)
-      @whisper.full_parallel(@params, samples, nil, Etc.nprocessors)
+      @whisper.full_parallel(@params, samples, nil, nprocessors)
 
-      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      assert_equal nprocessors, @whisper.full_n_segments
       text = @whisper.each_segment.collect(&:text).join
-      assert_match /ask what you can do/i, text
-      assert_match /for your country/i, text
+      assert_match(/ask what you can do/i, text)
+      assert_match(/for your country/i, text)
     end
 
     def test_full_parallel_without_length_and_n_processors
@@ -192,17 +200,18 @@ class TestWhisper < TestBase
 
       assert_equal 1, @whisper.full_n_segments
       text = @whisper.each_segment.collect(&:text).join
-      assert_match /ask what you can do/i, text
-      assert_match /for your country/i, text
+      assert_match(/ask what you can do/i, text)
+      assert_match(/for your country/i, text)
     end
 
     def test_full_parallel_without_length
-      @whisper.full_parallel(@params, @samples, nil, Etc.nprocessors)
+      nprocessors = 2
+      @whisper.full_parallel(@params, @samples, nil, nprocessors)
 
-      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      assert_equal nprocessors, @whisper.full_n_segments
       text = @whisper.each_segment.collect(&:text).join
-      assert_match /ask what you can do/i, text
-      assert_match /for your country/i, text
+      assert_match(/ask what you can do/i, text)
+      assert_match(/for your country/i, text)
     end
 
     def test_full_parallel_without_n_processors
@@ -210,8 +219,8 @@ class TestWhisper < TestBase
 
       assert_equal 1, @whisper.full_n_segments
       text = @whisper.each_segment.collect(&:text).join
-      assert_match /ask what you can do/i, text
-      assert_match /for your country/i, text
+      assert_match(/ask what you can do/i, text)
+      assert_match(/for your country/i, text)
     end
   end
 end

build-xcframework.sh

@@ -0,0 +1,535 @@
+#!/bin/bash
+#
+# Options
+IOS_MIN_OS_VERSION=16.4
+MACOS_MIN_OS_VERSION=13.3
+VISIONOS_MIN_OS_VERSION=1.0
+TVOS_MIN_OS_VERSION=16.4
+
+BUILD_SHARED_LIBS=OFF
+WHISPER_BUILD_EXAMPLES=OFF
+WHISPER_BUILD_TESTS=OFF
+WHISPER_BUILD_SERVER=OFF
+GGML_METAL=ON
+GGML_METAL_EMBED_LIBRARY=ON
+GGML_BLAS_DEFAULT=ON
+GGML_METAL_USE_BF16=ON
+GGML_OPENMP=OFF
+
+COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
+COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
+
+# Common options for all builds
+COMMON_CMAKE_ARGS=(
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_REQUIRED=NO
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGN_IDENTITY=""
+    -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_ALLOWED=NO
+    -DCMAKE_XCODE_ATTRIBUTE_DEBUG_INFORMATION_FORMAT="dwarf-with-dsym"
+    -DCMAKE_XCODE_ATTRIBUTE_GCC_GENERATE_DEBUGGING_SYMBOLS=YES
+    -DCMAKE_XCODE_ATTRIBUTE_COPY_PHASE_STRIP=NO
+    -DCMAKE_XCODE_ATTRIBUTE_STRIP_INSTALLED_PRODUCT=NO
+    -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+    -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
+    -DWHISPER_BUILD_EXAMPLES=${WHISPER_BUILD_EXAMPLES}
+    -DWHISPER_BUILD_TESTS=${WHISPER_BUILD_TESTS}
+    -DWHISPER_BUILD_SERVER=${WHISPER_BUILD_SERVER}
+    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
+    -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
+    -DGGML_METAL=${GGML_METAL}
+    -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16}
+    -DGGML_NATIVE=OFF
+    -DGGML_OPENMP=${GGML_OPENMP}
+)
+
+check_required_tool() {
+    local tool=$1
+    local install_message=$2
+
+    if ! command -v $tool &> /dev/null; then
+        echo "Error: $tool is required but not found."
+        echo "$install_message"
+        exit 1
+    fi
+}
+echo "Checking for required tools..."
+check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
+check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
+check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+
+set -e
+
+## Clean up previous builds
+rm -rf build-apple
+rm -rf build-ios-sim
+rm -rf build-ios-device
+rm -rf build-macos
+rm -rf build-visionos
+rm -rf build-visionos-sim
+rm -rf build-tvos-sim
+rm -rf build-tvos-device
+
+# Setup the xcframework build directory structure
+setup_framework_structure() {
+    local build_dir=$1
+    local min_os_version=$2
+    local platform=$3  # "ios", "macos", "visionos", or "tvos"
+    local framework_name="whisper"
+
+    echo "Creating ${platform}-style framework structure for ${build_dir}"
+
+    if [[ "$platform" == "macos" ]]; then
+        # macOS versioned structure uses versioned directories
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Headers
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Modules
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources
+
+        # Create symbolic links
+        ln -sf A ${build_dir}/framework/${framework_name}.framework/Versions/Current
+        ln -sf Versions/Current/Headers ${build_dir}/framework/${framework_name}.framework/Headers
+        ln -sf Versions/Current/Modules ${build_dir}/framework/${framework_name}.framework/Modules
+        ln -sf Versions/Current/Resources ${build_dir}/framework/${framework_name}.framework/Resources
+        ln -sf Versions/Current/${framework_name} ${build_dir}/framework/${framework_name}.framework/${framework_name}
+
+        # Set header and module paths
+        local header_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Headers/
+        local module_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Modules/
+    else
+        # iOS/VisionOS/tvOS use a flat structure
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Headers
+        mkdir -p ${build_dir}/framework/${framework_name}.framework/Modules
+
+        # Remove any existing structure to ensure clean build
+        rm -rf ${build_dir}/framework/${framework_name}.framework/Versions
+
+        # Set header and module paths
+        local header_path=${build_dir}/framework/${framework_name}.framework/Headers/
+        local module_path=${build_dir}/framework/${framework_name}.framework/Modules/
+    fi
+
+    # Copy all required headers (common for all platforms)
+    cp include/whisper.h           ${header_path}
+    cp ggml/include/ggml.h         ${header_path}
+    cp ggml/include/ggml-alloc.h   ${header_path}
+    cp ggml/include/ggml-backend.h ${header_path}
+    cp ggml/include/ggml-metal.h   ${header_path}
+    cp ggml/include/ggml-cpu.h     ${header_path}
+    cp ggml/include/ggml-blas.h    ${header_path}
+    cp ggml/include/gguf.h         ${header_path}
+
+    # Create module map (common for all platforms)
+    cat > ${module_path}module.modulemap << EOF
+framework module whisper {
+    header "whisper.h"
+    header "ggml.h"
+    header "ggml-alloc.h"
+    header "ggml-backend.h"
+    header "ggml-metal.h"
+    header "ggml-cpu.h"
+    header "ggml-blas.h"
+    header "gguf.h"
+
+    link "c++"
+    link framework "Accelerate"
+    link framework "Metal"
+    link framework "Foundation"
+
+    export *
+}
+EOF
+
+    # Platform-specific settings for Info.plist
+    local platform_name=""
+    local sdk_name=""
+    local supported_platform=""
+
+    case "$platform" in
+        "ios")
+            platform_name="iphoneos"
+            sdk_name="iphoneos${min_os_version}"
+            supported_platform="iPhoneOS"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family='    <key>UIDeviceFamily</key>
+    <array>
+        <integer>1</integer>
+        <integer>2</integer>
+    </array>'
+            ;;
+        "macos")
+            platform_name="macosx"
+            sdk_name="macosx${min_os_version}"
+            supported_platform="MacOSX"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/Info.plist"
+            local device_family=""
+            ;;
+        "visionos")
+            platform_name="xros"
+            sdk_name="xros${min_os_version}"
+            supported_platform="XRPlatform"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family=""
+            ;;
+        "tvos")
+            platform_name="appletvos"
+            sdk_name="appletvos${min_os_version}"
+            supported_platform="AppleTVOS"
+            local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist"
+            local device_family='    <key>UIDeviceFamily</key>
+    <array>
+        <integer>3</integer>
+    </array>'
+            ;;
+    esac
+
+    # Create Info.plist
+    cat > ${plist_path} << EOF
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>CFBundleDevelopmentRegion</key>
+    <string>en</string>
+    <key>CFBundleExecutable</key>
+    <string>whisper</string>
+    <key>CFBundleIdentifier</key>
+    <string>org.ggml.whisper</string>
+    <key>CFBundleInfoDictionaryVersion</key>
+    <string>6.0</string>
+    <key>CFBundleName</key>
+    <string>whisper</string>
+    <key>CFBundlePackageType</key>
+    <string>FMWK</string>
+    <key>CFBundleShortVersionString</key>
+    <string>1.0</string>
+    <key>CFBundleVersion</key>
+    <string>1</string>
+    <key>MinimumOSVersion</key>
+    <string>${min_os_version}</string>
+    <key>CFBundleSupportedPlatforms</key>
+    <array>
+        <string>${supported_platform}</string>
+    </array>${device_family}
+    <key>DTPlatformName</key>
+    <string>${platform_name}</string>
+    <key>DTSDKName</key>
+    <string>${sdk_name}</string>
+</dict>
+</plist>
+EOF
+}
+
+# Create dynamic libraries from static libraries.
+combine_static_libraries() {
+    local build_dir="$1"
+    local release_dir="$2"
+    local platform="$3"  # "ios", "macos", "visionos", or "tvos"
+    local is_simulator="$4"
+    local base_dir="$(pwd)"
+    local framework_name="whisper"
+
+    # Determine output path based on platform
+    local output_lib=""
+    if [[ "$platform" == "macos" ]]; then
+        # macOS uses versioned structure
+        output_lib="${build_dir}/framework/${framework_name}.framework/Versions/A/${framework_name}"
+    else
+        # iOS, visionOS, and tvOS use a directory flat structure
+        output_lib="${build_dir}/framework/${framework_name}.framework/${framework_name}"
+    fi
+
+    local libs=(
+        "${base_dir}/${build_dir}/src/${release_dir}/libwhisper.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-base.a"
+        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a"
+        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
+        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
+    )
+    if [[ "$platform" == "macos" || "$platform" == "ios" ]]; then
+        echo "Adding libwhisper.coreml library to the build."
+        libs+=(
+            "${base_dir}/${build_dir}/src/${release_dir}/libwhisper.coreml.a"
+        )
+    fi
+
+    # Create temporary directory for processing
+    local temp_dir="${base_dir}/${build_dir}/temp"
+    echo "Creating temporary directory: ${temp_dir}"
+    mkdir -p "${temp_dir}"
+
+    # Since we have multiple architectures libtool will find object files that do not
+    # match the target architecture. We suppress these warnings.
+    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+
+    # Determine SDK, architectures, and install_name based on platform and simulator flag.
+    local sdk=""
+    local archs=""
+    local min_version_flag=""
+    local install_name=""
+    local frameworks="-framework Foundation -framework Metal -framework Accelerate"
+
+    case "$platform" in
+        "ios")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="iphonesimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mios-simulator-version-min=${IOS_MIN_OS_VERSION}"
+            else
+                sdk="iphoneos"
+                archs="arm64"
+                min_version_flag="-mios-version-min=${IOS_MIN_OS_VERSION}"
+            fi
+            install_name="@rpath/whisper.framework/whisper"
+            frameworks+=" -framework CoreML"
+            ;;
+        "macos")
+            sdk="macosx"
+            archs="arm64 x86_64"
+            min_version_flag="-mmacosx-version-min=${MACOS_MIN_OS_VERSION}"
+            install_name="@rpath/whisper.framework/Versions/Current/whisper"
+            frameworks+=" -framework CoreML"
+            ;;
+        "visionos")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="xrsimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}-simulator"
+            else
+                sdk="xros"
+                archs="arm64"
+                min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}"
+            fi
+            # Use flat structure for visionOS, same as iOS
+            install_name="@rpath/whisper.framework/whisper"
+            ;;
+        "tvos")
+            if [[ "$is_simulator" == "true" ]]; then
+                sdk="appletvsimulator"
+                archs="arm64 x86_64"
+                min_version_flag="-mtvos-simulator-version-min=${TVOS_MIN_OS_VERSION}"
+            else
+                sdk="appletvos"
+                archs="arm64"
+                min_version_flag="-mtvos-version-min=${TVOS_MIN_OS_VERSION}"
+            fi
+            install_name="@rpath/whisper.framework/whisper"
+            ;;
+    esac
+
+    # Build architecture flags
+    local arch_flags=""
+    for arch in $archs; do
+        arch_flags+=" -arch $arch"
+    done
+
+    # Create dynamic library
+    echo "Creating dynamic library for ${platform}."
+    xcrun -sdk $sdk clang++ -dynamiclib \
+        -isysroot $(xcrun --sdk $sdk --show-sdk-path) \
+        $arch_flags \
+        $min_version_flag \
+        -Wl,-force_load,"${temp_dir}/combined.a" \
+        $frameworks \
+        -install_name "$install_name" \
+        -o "${base_dir}/${output_lib}"
+
+    # Platform-specific post-processing for device builds
+    if [[ "$is_simulator" == "false" ]]; then
+        if command -v vtool &>/dev/null; then
+            case "$platform" in
+                "ios")
+                    echo "Marking binary as a framework binary for iOS..."
+                    vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+                "visionos")
+                    echo "Marking binary as a framework binary for visionOS..."
+                    vtool -set-build-version xros ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+                "tvos")
+                    echo "Marking binary as a framework binary for tvOS..."
+                    vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \
+                        -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}"
+                    ;;
+            esac
+        else
+            echo "Warning: vtool not found. Binary may not pass App Store validation."
+        fi
+    fi
+
+    echo "Creating properly formatted dSYM..."
+    # Create a separate directory for dSYMs for all platforms
+    mkdir -p "${base_dir}/${build_dir}/dSYMs"
+
+    # iOS and visionOS style dSYM (flat structure)
+    if [[ "$platform" == "ios" || "$platform" == "visionos" || "$platform" == "tvos" ]]; then
+        # Generate dSYM in the dSYMs directory
+        xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/whisper.dSYM"
+
+        # Create a copy of the binary that will be stripped
+        cp "${base_dir}/${output_lib}" "${temp_dir}/binary_to_strip"
+
+        # Strip debug symbols from the copy
+        xcrun strip -S "${temp_dir}/binary_to_strip" -o "${temp_dir}/stripped_lib"
+
+        # Replace the original with the stripped version
+        mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}"
+    else
+        # macOS style dSYM
+        # First strip debug info to a separate file
+        xcrun strip -S "${base_dir}/${output_lib}" -o "${temp_dir}/stripped_lib"
+
+        # Generate dSYM in the dSYMs directory
+        xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/whisper.dSYM"
+
+        # Replace original binary with stripped version
+        mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}"
+    fi
+
+    # Remove any automatically generated dSYM files in the framework structure as they will
+    # otherwise case Invalid Bundle Structure validation errors.
+    if [ -d "${base_dir}/${output_lib}.dSYM" ]; then
+        echo "Removing generated dSYM file in framework structure: ${base_dir}/${output_lib}.dSYM"
+        rm -rf "${base_dir}/${output_lib}.dSYM"
+    fi
+
+    # Clean up
+    rm -rf "${temp_dir}"
+}
+
+echo "Building for iOS simulator..."
+cmake -B build-ios-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DIOS=ON \
+    -DCMAKE_SYSTEM_NAME=iOS \
+    -DCMAKE_OSX_SYSROOT=iphonesimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DWHISPER_COREML="ON" \
+    -DWHISPER_COREML_ALLOW_FALLBACK="ON" \
+    -S .
+cmake --build build-ios-sim --config Release -- -quiet
+
+echo "Building for iOS devices..."
+cmake -B build-ios-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_SYSROOT=iphoneos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DWHISPER_COREML="ON" \
+    -DWHISPER_COREML_ALLOW_FALLBACK="ON" \
+    -S .
+cmake --build build-ios-device --config Release -- -quiet
+
+echo "Building for macOS..."
+cmake -B build-macos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -DWHISPER_COREML="ON" \
+    -DWHISPER_COREML_ALLOW_FALLBACK="ON" \
+    -S .
+cmake --build build-macos --config Release -- -quiet
+
+echo "Building for visionOS..."
+cmake -B build-visionos -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xros \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-visionos --config Release -- -quiet
+
+echo "Building for visionOS simulator..."
+cmake -B build-visionos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DCMAKE_SYSTEM_NAME=visionOS \
+    -DCMAKE_OSX_SYSROOT=xrsimulator \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-visionos-sim --config Release -- -quiet
+
+# Add tvOS builds (might need the same u_int definitions as watchOS and visionOS)
+echo "Building for tvOS simulator..."
+cmake -B build-tvos-sim -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvsimulator \
+    -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-tvos-sim --config Release -- -quiet
+
+echo "Building for tvOS devices..."
+cmake -B build-tvos-device -G Xcode \
+    "${COMMON_CMAKE_ARGS[@]}" \
+    -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=tvOS \
+    -DCMAKE_OSX_SYSROOT=appletvos \
+    -DCMAKE_OSX_ARCHITECTURES="arm64" \
+    -DGGML_METAL=ON \
+    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
+    -S .
+cmake --build build-tvos-device --config Release -- -quiet
+
+# Setup frameworks and copy binaries and headers
+echo "Setting up framework structures..."
+setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios"
+setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos"
+setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos"
+setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos"
+setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos"
+
+# Create dynamic libraries from static libraries
+echo "Creating dynamic libraries from static libraries..."
+combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true"
+combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false"
+combine_static_libraries "build-macos" "Release" "macos" "false"
+combine_static_libraries "build-visionos" "Release-xros" "visionos" "false"
+combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true"
+combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true"
+combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
+
+# Create XCFramework with correct debug symbols paths
+echo "Creating XCFramework..."
+xcodebuild -create-xcframework \
+    -framework $(pwd)/build-ios-sim/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-ios-sim/dSYMs/whisper.dSYM \
+    -framework $(pwd)/build-ios-device/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-ios-device/dSYMs/whisper.dSYM \
+    -framework $(pwd)/build-macos/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-macos/dSYMS/whisper.dSYM \
+    -framework $(pwd)/build-visionos/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-visionos/dSYMs/whisper.dSYM \
+    -framework $(pwd)/build-visionos-sim/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-visionos-sim/dSYMs/whisper.dSYM \
+    -framework $(pwd)/build-tvos-device/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-tvos-device/dSYMs/whisper.dSYM \
+    -framework $(pwd)/build-tvos-sim/framework/whisper.framework \
+    -debug-symbols $(pwd)/build-tvos-sim/dSYMs/whisper.dSYM \
+    -output $(pwd)/build-apple/whisper.xcframework

ci/README.md

@@ -0,0 +1,41 @@
+# CI
+
+In addition to [Github Actions](https://github.com/ggerganov/whisper.cpp/actions) `whisper.cpp` uses a custom CI framework:
+
+https://github.com/ggml-org/ci
+
+It monitors the `master` branch for new commits and runs the
+[ci/run.sh](https://github.com/ggerganov/whisper.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
+to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
+to cover various hardware architectures, including GPU and Apple Silicon instances.
+
+Collaborators can optionally trigger the CI run by adding the `ggml-ci` keyword to their commit message.
+Only the branches of this repo are monitored for this keyword.
+
+It is a good practice, before publishing changes to execute the full CI locally on your machine:
+
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with CUDA support
+GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```
+
+## Environment Variables
+
+The CI script supports several environment variables to control the build:
+
+| Variable | Description |
+|----------|-------------|
+| `GG_BUILD_CUDA` | Enable NVIDIA CUDA GPU acceleration |
+| `GG_BUILD_SYCL` | Enable Intel SYCL acceleration |
+| `GG_BUILD_VULKAN` | Enable Vulkan GPU acceleration |
+| `GG_BUILD_METAL` | Enable Metal acceleration on Apple Silicon |
+| `GG_BUILD_BLAS` | Enable BLAS CPU acceleration |
+| `GG_BUILD_OPENVINO` | Enable OpenVINO support |
+| `GG_BUILD_COREML` | Enable Core ML support for Apple Neural Engine |
+| `GG_BUILD_LOW_PERF` | Limit tests for low-performance hardware |
+| `GG_BUILD_TEST_MODELS` | Comma-separated list of models to test (e.g. "tiny.en,tiny,base,medium", defaults to all models unless `GG_BUILD_LOW_PERF` is set) |

ci/run.sh

@@ -0,0 +1,336 @@
+#!/bin/bash
+#
+# sample usage:
+#
+# mkdir tmp
+#
+# # CPU-only build
+# bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with CUDA support
+# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with SYCL support
+# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+if [ -z "$2" ]; then
+    echo "usage: $0 <output-dir> <mnt-dir>"
+    exit 1
+fi
+
+mkdir -p "$1"
+mkdir -p "$2"
+
+OUT=$(realpath "$1")
+MNT=$(realpath "$2")
+
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
+
+sd=`dirname $0`
+cd $sd/../
+SRC=`pwd`
+
+ALL_MODELS=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" "large-v3-turbo" )
+BENCH_N_THREADS=4
+BENCH_ENCODER_ONLY=0
+BENCH_FLASH_ATTN=0
+
+# check for user-specified models first. if not specified, use fast models
+if [ ! -z ${GG_BUILD_TEST_MODELS} ]; then
+    IFS=',' read -r -a MODELS <<< "${GG_BUILD_TEST_MODELS}"
+else
+    if [ ! -z ${GG_BUILD_LOW_PERF} ]; then
+        MODELS=( "tiny" "base" "small" )
+    else
+        MODELS=("${ALL_MODELS[@]}")
+    fi
+fi
+
+CMAKE_EXTRA="-DWHISPER_FATAL_WARNINGS=ON"
+
+if [ ! -z ${GG_BUILD_CUDA} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+fi
+
+if [ ! -z ${GG_BUILD_SYCL} ]; then
+    if [ -z ${ONEAPI_ROOT} ]; then
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
+        echo "source /opt/intel/oneapi/setvars.sh"
+        exit 1
+    fi
+
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
+fi
+
+if [ ! -z ${GG_BUILD_OPENVINO} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DWHISPER_OPENVINO=ON"
+fi
+
+if [ ! -z ${GG_BUILD_METAL} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
+fi
+
+if [ ! -z ${GG_BUILD_VULKAN} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=ON"
+fi
+
+if [ ! -z ${GG_BUILD_BLAS} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_BLAS=ON"
+fi
+
+if [ ! -z ${GG_BUILD_COREML} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DWHISPER_COREML=ON"
+fi
+
+## helpers
+
+# download a file if it does not exist or if it is outdated
+function gg_wget {
+    local out=$1
+    local url=$2
+
+    local cwd=`pwd`
+
+    mkdir -p $out
+    cd $out
+
+    # should not re-download if file is the same
+    wget -nv -N $url
+
+    cd $cwd
+}
+
+function gg_download_model {
+    local model_name=$1
+    local model_file="$MNT/models/ggml-${model_name}.bin"
+
+    if [ ! -f ${model_file} ]; then
+        local cwd=`pwd`
+        mkdir -p "$MNT/models"
+        cd "$MNT/models"
+        bash "$cwd/models/download-ggml-model.sh" ${model_name} .
+        cd "$cwd"
+    fi
+}
+
+function gg_printf {
+    printf -- "$@" >> $OUT/README.md
+}
+
+# Helper function to check command exit status
+function gg_check_last_command_status {
+    local exit_file=$1
+    local command_name=$2
+
+    local exit_status=$?
+    echo "$exit_status" > "$exit_file"
+
+    if [ $exit_status -ne 0 ]; then
+        echo "Error: Command $command_name failed with exit status $exit_status"
+        return 1
+    fi
+
+    return 0
+}
+
+# Usage: gg_run <test_name> [additional_args...]
+#
+# Parameters:
+#   test_name       - Name of the test to run (calls gg_run_<test_name>)
+#   additional_args - Any additional arguments to pass to the test function (first argument is appended to the log filename)
+function gg_run {
+    ci=$1
+
+    if [ $# -gt 1 ]; then
+        ci="${ci}_${2}"
+    fi
+
+    set -o pipefail
+    set -x
+
+    gg_run_$1 "$@" | tee $OUT/$ci.log
+    cur=$?
+    echo "$cur" > $OUT/$ci.exit
+
+    set +x
+    set +o pipefail
+
+    gg_sum_$1 "$@"
+
+    ret=$((ret | cur))
+}
+
+function gg_check_build_requirements {
+    if ! command -v cmake &> /dev/null; then
+        gg_printf 'cmake not found, please install'
+    fi
+
+    if ! command -v make &> /dev/null; then
+        gg_printf 'make not found, please install'
+    fi
+}
+
+## ci
+
+function gg_run_ctest {
+    mode=$2
+
+    cd ${SRC}
+    
+    rm -rf build-ci-${mode} && mkdir build-ci-${mode} && cd build-ci-${mode}
+
+    set -e
+
+    gg_check_build_requirements
+
+    (time cmake -DCMAKE_BUILD_TYPE=${mode} ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    set +e
+}
+
+function gg_sum_ctest {
+    mode=$2
+
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in '${mode}' mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+function gg_run_bench {
+    cd ${SRC}
+
+    # set flash attention flag if enabled
+    fattn=""
+    if [ "$BENCH_FLASH_ATTN" -eq 1 ]; then
+        fattn="-fa"
+    fi
+
+    # run memcpy benchmark if not encoder-only mode
+    if [ "$BENCH_ENCODER_ONLY" -eq 0 ]; then
+        echo "Running memcpy benchmark"
+        (time ./build-ci-release/bin/whisper-bench -w 1 -t $BENCH_N_THREADS 2>&1) | tee -a $OUT/${ci}-memcpy.log
+        gg_check_last_command_status "$OUT/${ci}-memcpy.exit" "memcpy benchmark"
+        
+        echo "Running ggml_mul_mat benchmark with $BENCH_N_THREADS threads"
+        (time ./build-ci-release/bin/whisper-bench -w 2 -t $BENCH_N_THREADS 2>&1) | tee -a $OUT/${ci}-mul_mat.log
+        gg_check_last_command_status "$OUT/${ci}-mul_mat.exit" "ggml_mul_mat benchmark"
+    fi
+
+    echo "Running benchmark for all models"
+
+    # generate header for the benchmark table
+    {
+        printf "| %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "Config" "Model" "Th" "FA" "Enc." "Dec." "Bch5" "PP" "Commit"
+        printf "| %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---"
+    } | tee -a $OUT/${ci}-models-table.log
+
+    # run benchmark for each model
+    for model in "${MODELS[@]}"; do
+        echo "Benchmarking model: $model"
+
+        # run the benchmark and capture output
+        output=$(./build-ci-release/bin/whisper-bench -m $MNT/models/ggml-$model.bin -t $BENCH_N_THREADS $fattn 2>&1)
+        ret=$?
+
+        # save the raw output
+        echo "$output" > $OUT/${ci}-bench-$model.log
+
+        if [ $ret -eq 0 ]; then
+            # parse the benchmark results
+            encode_time=$(echo "$output" | grep "encode time" | awk '{print $11}')
+            decode_time=$(echo "$output" | grep "decode time" | awk '{print $11}')
+            batchd_time=$(echo "$output" | grep "batchd time" | awk '{print $11}')
+            prompt_time=$(echo "$output" | grep "prompt time" | awk '{print $11}')
+            system_info=$(echo "$output" | grep "system_info")
+            actual_threads=$(echo "$output" | grep "system_info" | awk '{print $4}')
+
+            # determine configuration
+            config=""
+            if [[ $system_info == *"AVX2 = 1"* ]]; then
+                config="$config AVX2"
+            fi
+            if [[ $system_info == *"NEON = 1"* ]]; then
+                config="$config NEON"
+            fi
+            if [[ $system_info == *"BLAS = 1"* ]]; then
+                config="$config BLAS"
+            fi
+            if [[ $system_info == *"COREML = 1"* ]]; then
+                config="$config COREML"
+            fi
+            if [[ $system_info == *"CUDA = 1"* ]]; then
+                config="$config CUDA"
+            fi
+            if [[ $system_info == *"METAL = 1"* ]]; then
+                config="$config METAL"
+            fi
+
+            # get commit hash
+            commit=$(git rev-parse --short HEAD)
+
+            # add row to benchmark table
+            printf "| %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" \
+                "$config" "$model" "$actual_threads" "$BENCH_FLASH_ATTN" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit" \
+                | tee -a $OUT/${ci}-models-table.log
+        else
+            echo "Benchmark failed for model: $model" | tee -a $OUT/${ci}-bench-errors.log
+        fi
+    done
+}
+
+function gg_sum_bench {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Whisper Benchmark Results\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+
+    # show memcpy and ggml_mul_mat benchmark results if available
+    if [ "$BENCH_ENCODER_ONLY" -eq 0 ]; then
+        if [ -f "$OUT/${ci}-memcpy.log" ]; then
+            gg_printf '#### memcpy Benchmark\n\n'
+            gg_printf '```\n%s\n```\n\n' "$(cat $OUT/${ci}-memcpy.log)"
+        fi
+
+        if [ -f "$OUT/${ci}-mul_mat.log" ]; then
+            gg_printf '#### ggml_mul_mat Benchmark\n\n'
+            gg_printf '```\n%s\n```\n\n' "$(cat $OUT/${ci}-mul_mat.log)"
+        fi
+    fi
+
+    # show model benchmark results
+    gg_printf '#### Model Benchmarks\n\n'
+    if [ -f "$OUT/${ci}-models-table.log" ]; then
+        gg_printf '%s\n\n' "$(cat $OUT/${ci}-models-table.log)"
+    else
+        gg_printf 'No model benchmark results available.\n\n'
+    fi
+
+    # show any errors that occurred
+    if [ -f "$OUT/${ci}-bench-errors.log" ]; then
+        gg_printf '#### Benchmark Errors\n\n'
+        gg_printf '```\n%s\n```\n\n' "$(cat $OUT/${ci}-bench-errors.log)"
+    fi
+}
+
+ret=0
+
+for model in "${MODELS[@]}"; do
+    test $ret -eq 0 && gg_download_model ${model}
+done
+if [ -z ${GG_BUILD_SYCL}]; then
+    test $ret -eq 0 && gg_run ctest debug
+fi
+test $ret -eq 0 && gg_run ctest release
+
+test $ret -eq 0 && gg_run bench
+
+exit $ret

close-issue.yml

@@ -0,0 +1,28 @@
+name: Close inactive issues
+on:
+  schedule:
+    - cron: "42 0 * * *"
+
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  issues: write
+
+jobs:
+  close-issues:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v5
+        with:
+          exempt-issue-labels: "refactor,help wanted,good first issue,research,bug,roadmap"
+          days-before-issue-stale: 30
+          days-before-issue-close: 14
+          stale-issue-label: "stale"
+          close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+          operations-per-run: 10000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}

cmake/build-info.cmake

@@ -42,6 +42,8 @@ endif()
 if(MSVC)
     set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
     set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/utf-8>")
 else()
     execute_process(
         COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}

examples/CMakeLists.txt

@@ -14,10 +14,6 @@ if (WHISPER_SDL2)
     message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
 endif()
 
-if (WHISPER_CLBLAST)
-    find_package(CLBlast REQUIRED)
-endif()
-
 # common
 
 set(TARGET common)
@@ -56,6 +52,8 @@ add_library(${TARGET} STATIC
     common.cpp
     common-ggml.h
     common-ggml.cpp
+    common-whisper.h
+    common-whisper.cpp
     grammar-parser.h
     grammar-parser.cpp
     ${COMMON_SOURCES_FFMPEG}
@@ -63,7 +61,7 @@ add_library(${TARGET} STATIC
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
+target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS} ${CMAKE_DL_LIBS})
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 set_target_properties(${TARGET} PROPERTIES FOLDER "libs")

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

@@ -18,6 +18,7 @@ const whisperParamsMock = {
   translate: true,
   no_timestamps: false,
   audio_ctx: 0,
+  max_len: 0,
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -1,5 +1,6 @@
 #include "napi.h"
 #include "common.h"
+#include "common-whisper.h"
 
 #include "whisper.h"
 
@@ -127,192 +128,227 @@ 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");
-        return 2;
-    }
-
-    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
-        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
-        exit(0);
-    }
-
-    // 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 * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
-
-    if (ctx == nullptr) {
-        fprintf(stderr, "error: failed to initialize whisper context\n");
-        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];
-
-        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;
-        }
-
-        // 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) {
-                    params.language = "en";
-                    params.translate = false;
-                    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",
-                    __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);
-
-            fprintf(stderr, "\n");
-        }
-
-        // run the inference
-        {
-            whisper_full_params wparams = whisper_full_default_params(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;
-            wparams.print_timestamps = !params.no_timestamps;
-            wparams.print_special    = params.print_special;
-            wparams.translate        = params.translate;
-            wparams.language         = params.language.c_str();
-            wparams.n_threads        = params.n_threads;
-            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
-            wparams.offset_ms        = params.offset_t_ms;
-            wparams.duration_ms      = params.duration_ms;
-
-            wparams.token_timestamps = params.output_wts || params.max_len > 0;
-            wparams.thold_pt         = params.word_thold;
-            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.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
-            if (!wparams.print_realtime) {
-                wparams.new_segment_callback           = whisper_print_segment_callback;
-                wparams.new_segment_callback_user_data = &user_data;
-            }
-
-            // example for abort mechanism
-            // in this example, we do not abort the processing, but we could if the flag is set to true
-            // the callback is called before every encoder run - if it returns false, the processing is aborted
-            {
-                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
-
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
-                    bool is_aborted = *(bool*)user_data;
-                    return !is_aborted;
-                };
-                wparams.encoder_begin_callback_user_data = &is_aborted;
-            }
-
-            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
-                fprintf(stderr, "failed to process audio\n");
-                return 10;
-            }
-        }
-    }
-
-    const int n_segments = whisper_full_n_segments(ctx);
-    result.resize(n_segments);
-    for (int i = 0; i < n_segments; ++i) {
-        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);
-
-        result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
-        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
-        result[i].emplace_back(text);
-    }
-
-    whisper_print_timings(ctx);
-    whisper_free(ctx);
-
-    return 0;
-}
-
-class Worker : public Napi::AsyncWorker {
+class ProgressWorker : public Napi::AsyncWorker {
  public:
-  Worker(Napi::Function& callback, whisper_params params)
-      : Napi::AsyncWorker(callback), params(params) {}
-
-  void Execute() override {
-    run(params, result);
-  }
-
-  void OnOK() override {
-    Napi::HandleScope scope(Env());
-    Napi::Object res = Napi::Array::New(Env(), result.size());
-    for (uint64_t i = 0; i < result.size(); ++i) {
-      Napi::Object tmp = Napi::Array::New(Env(), 3);
-      for (uint64_t j = 0; j < 3; ++j) {
-        tmp[j] = Napi::String::New(Env(), result[i][j]);
-      }
-      res[i] = tmp;
+    ProgressWorker(Napi::Function& callback, whisper_params params, Napi::Function progress_callback, Napi::Env env)
+        : Napi::AsyncWorker(callback), params(params), env(env) {
+        // Create thread-safe function
+        if (!progress_callback.IsEmpty()) {
+            tsfn = Napi::ThreadSafeFunction::New(
+                env,
+                progress_callback,
+                "Progress Callback",
+                0,
+                1
+            );
+        }
+    }
+
+    ~ProgressWorker() {
+        if (tsfn) {
+            // Make sure to release the thread-safe function on destruction
+            tsfn.Release();
+        }
+    }
+
+    void Execute() override {
+        // Use custom run function with progress callback support
+        run_with_progress(params, result);
+    }
+
+    void OnOK() override {
+        Napi::HandleScope scope(Env());
+        Napi::Object res = Napi::Array::New(Env(), result.size());
+        for (uint64_t i = 0; i < result.size(); ++i) {
+            Napi::Object tmp = Napi::Array::New(Env(), 3);
+            for (uint64_t j = 0; j < 3; ++j) {
+                tmp[j] = Napi::String::New(Env(), result[i][j]);
+            }
+            res[i] = tmp;
+        }
+        Callback().Call({Env().Null(), res});
+    }
+
+    // Progress callback function - using thread-safe function
+    void OnProgress(int progress) {
+        if (tsfn) {
+            // Use thread-safe function to call JavaScript callback
+            auto callback = [progress](Napi::Env env, Napi::Function jsCallback) {
+                jsCallback.Call({Napi::Number::New(env, progress)});
+            };
+            
+            tsfn.BlockingCall(callback);
+        }
     }
-    Callback().Call({Env().Null(), res});
-  }
 
  private:
-  whisper_params params;
-  std::vector<std::vector<std::string>> result;
+    whisper_params params;
+    std::vector<std::vector<std::string>> result;
+    Napi::Env env;
+    Napi::ThreadSafeFunction tsfn;
+
+    // Custom run function with progress callback support
+    int run_with_progress(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");
+            return 2;
+        }
+
+        if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+            fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+            exit(0);
+        }
+
+        // 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 * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+        if (ctx == nullptr) {
+            fprintf(stderr, "error: failed to initialize whisper context\n");
+            return 3;
+        }
+
+        // If params.pcmf32 provides, set params.fname_inp as "buffer"
+        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];
+
+            std::vector<float> pcmf32; // mono-channel F32 PCM
+            std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+            // If params.pcmf32 is empty, read input audio file
+            if (params.pcmf32.empty()) {
+                if (!::read_audio_data(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+                    fprintf(stderr, "error: failed to read audio file '%s'\n", fname_inp.c_str());
+                    continue;
+                }
+            } else {
+                pcmf32 = params.pcmf32;
+            }
+
+            // Print system info
+            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 processing info
+            if (!params.no_prints) {
+                fprintf(stderr, "\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__);
+                    }
+                }
+                fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d, audio_ctx = %d ...\n",
+                        __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
+                        params.n_threads, params.n_processors,
+                        params.language.c_str(),
+                        params.translate ? "translate" : "transcribe",
+                        params.no_timestamps ? 0 : 1,
+                        params.audio_ctx);
+
+                fprintf(stderr, "\n");
+            }
+
+            // Run inference
+            {
+                whisper_full_params wparams = whisper_full_default_params(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;
+                wparams.print_timestamps = !params.no_timestamps;
+                wparams.print_special    = params.print_special;
+                wparams.translate        = params.translate;
+                wparams.language         = params.language.c_str();
+                wparams.n_threads        = params.n_threads;
+                wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+                wparams.offset_ms        = params.offset_t_ms;
+                wparams.duration_ms      = params.duration_ms;
+
+                wparams.token_timestamps = params.output_wts || params.max_len > 0;
+                wparams.thold_pt         = params.word_thold;
+                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.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 for each new segment
+                if (!wparams.print_realtime) {
+                    wparams.new_segment_callback           = whisper_print_segment_callback;
+                    wparams.new_segment_callback_user_data = &user_data;
+                }
+
+                // Set progress callback
+                wparams.progress_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+                    ProgressWorker* worker = static_cast<ProgressWorker*>(user_data);
+                    worker->OnProgress(progress);
+                };
+                wparams.progress_callback_user_data = this;
+
+                // Abort mechanism example
+                {
+                    static bool is_aborted = false; // Note: this should be atomic to avoid data races
+
+                    wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                        bool is_aborted = *(bool*)user_data;
+                        return !is_aborted;
+                    };
+                    wparams.encoder_begin_callback_user_data = &is_aborted;
+                }
+
+                if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                    fprintf(stderr, "failed to process audio\n");
+                    return 10;
+                }
+            }
+    }
+
+        const int n_segments = whisper_full_n_segments(ctx);
+        result.resize(n_segments);
+        for (int i = 0; i < n_segments; ++i) {
+            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);
+
+            result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
+            result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
+            result[i].emplace_back(text);
+        }
+
+        whisper_print_timings(ctx);
+        whisper_free(ctx);
+
+        return 0;
+    }
 };
 
-
-
 Napi::Value whisper(const Napi::CallbackInfo& info) {
   Napi::Env env = info.Env();
   if (info.Length() <= 0 || !info[0].IsObject()) {
@@ -330,6 +366,24 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   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>();
+  int32_t max_len = whisper_params.Get("max_len").As<Napi::Number>();
+  
+  // support prompt
+  std::string prompt = "";
+  if (whisper_params.Has("prompt") && whisper_params.Get("prompt").IsString()) {
+    prompt = whisper_params.Get("prompt").As<Napi::String>();
+  }
+  
+  // Add support for print_progress
+  bool print_progress = false;
+  if (whisper_params.Has("print_progress")) {
+    print_progress = whisper_params.Get("print_progress").As<Napi::Boolean>();
+  }
+  // Add support for progress_callback
+  Napi::Function progress_callback;
+  if (whisper_params.Has("progress_callback") && whisper_params.Get("progress_callback").IsFunction()) {
+    progress_callback = whisper_params.Get("progress_callback").As<Napi::Function>();
+  }
 
   Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
   std::vector<float> pcmf32_vec;
@@ -352,9 +406,13 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   params.audio_ctx = audio_ctx;
   params.pcmf32 = pcmf32_vec;
   params.comma_in_time = comma_in_time;
+  params.max_len = max_len;
+  params.print_progress = print_progress;
+  params.prompt = prompt;
 
   Napi::Function callback = info[1].As<Napi::Function>();
-  Worker* worker = new Worker(callback, params);
+  // Create a new Worker class with progress callback support
+  ProgressWorker* worker = new ProgressWorker(callback, params, progress_callback, env);
   worker->Queue();
   return env.Undefined();
 }

examples/addon.node/index.js

@@ -18,6 +18,10 @@ const whisperParams = {
   translate: true,
   no_timestamps: false,
   audio_ctx: 0,
+  max_len: 0,
+  progress_callback: (progress) => {
+      console.log(`progress: ${progress}%`);
+    }
 };
 
 const arguments = process.argv.slice(2);

examples/bench.wasm/README.md

@@ -2,7 +2,7 @@
 
 Benchmark the performance of whisper.cpp in the browser using WebAssembly
 
-Link: https://whisper.ggerganov.com/bench/
+Link: https://ggerganov.github.io/whisper.cpp/bench.wasm
 
 Terminal version: [examples/bench](/examples/bench)
 
@@ -15,7 +15,17 @@ cd whisper.cpp
 mkdir build-em && cd build-em
 emcmake cmake ..
 make -j
+```
+The example can then be started by running a local HTTP server:
+```console
+python3 examples/server.py
+```
+And then opening a browser to the following URL:
+http://localhost:8000/bench.wasm
 
+To run the example in a different server, you need to copy the following files
+to the server's HTTP path:
+```
 # copy the produced page to your HTTP path
 cp bin/bench.wasm/*       /path/to/html/
 cp bin/libbench.worker.js /path/to/html/

examples/bench.wasm/index-tmpl.html

@@ -24,6 +24,8 @@
                 overflow-x: scroll;
             }
         </style>
+        <script src="../coi-serviceworker.js"></script>
+        <link rel="icon" href="data:,">
     </head>
     <body>
         <div id="main-container">
@@ -36,11 +38,10 @@
             <br><br>
 
             <b>More examples:</b>
-                <a href="https://whisper.ggerganov.com/">main</a> |
-                <a href="https://whisper.ggerganov.com/bench">bench</a> |
-                <a href="https://whisper.ggerganov.com/stream">stream</a> |
-                <a href="https://whisper.ggerganov.com/command">command</a> |
-                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+                <a href="../">main</a> |
+                <a href="../bench.wasm/">bench</a> |
+                <a href="../stream.wasm">stream</a> |
+                <a href="../command.wasm/">command</a> |
 
             <br><br>
 

examples/bench/bench.cpp

@@ -50,11 +50,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
-    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
     fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
+    fprintf(stderr, "  -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, "\n");
 }
 

examples/cli/cli.cpp

@@ -1,4 +1,5 @@
 #include "common.h"
+#include "common-whisper.h"
 
 #include "whisper.h"
 #include "grammar-parser.h"
@@ -6,12 +7,18 @@
 #include <cmath>
 #include <fstream>
 #include <cstdio>
-#include <regex>
 #include <string>
 #include <thread>
 #include <vector>
 #include <cstring>
 
+#if defined(_WIN32)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -194,7 +201,8 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
 
 static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
-    fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
+    fprintf(stderr, "usage: %s [options] file0 file1 ...\n", argv[0]);
+    fprintf(stderr, "supported audio formats: flac, mp3, ogg, wav\n");
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
@@ -239,7 +247,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     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, "  -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, "  -f FNAME,  --file FNAME        [%-7s] input audio 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");
@@ -916,6 +924,13 @@ static bool output_lrc(struct whisper_context * ctx, const char * fname, const w
 static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
 
 int main(int argc, char ** argv) {
+#if defined(_WIN32)
+    // Set the console output code page to UTF-8, while command line arguments
+    // are still encoded in the system's code page. In this way, we can print
+    // non-ASCII characters to the console, and access files with non-ASCII paths.
+    SetConsoleOutputCP(CP_UTF8);
+#endif
+
     whisper_params params;
 
     // If the only argument starts with "@", read arguments line-by-line
@@ -1057,8 +1072,8 @@ int main(int argc, char ** argv) {
         std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+        if (!::read_audio_data(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read audio file '%s'\n", fname_inp.c_str());
             continue;
         }
 

examples/coi-serviceworker.js

@@ -0,0 +1,146 @@
+/*! coi-serviceworker v0.1.7 - Guido Zuidhof and contributors, licensed under MIT */
+let coepCredentialless = false;
+if (typeof window === 'undefined') {
+    self.addEventListener("install", () => self.skipWaiting());
+    self.addEventListener("activate", (event) => event.waitUntil(self.clients.claim()));
+
+    self.addEventListener("message", (ev) => {
+        if (!ev.data) {
+            return;
+        } else if (ev.data.type === "deregister") {
+            self.registration
+                .unregister()
+                .then(() => {
+                    return self.clients.matchAll();
+                })
+                .then(clients => {
+                    clients.forEach((client) => client.navigate(client.url));
+                });
+        } else if (ev.data.type === "coepCredentialless") {
+            coepCredentialless = ev.data.value;
+        }
+    });
+
+    self.addEventListener("fetch", function (event) {
+        const r = event.request;
+        if (r.cache === "only-if-cached" && r.mode !== "same-origin") {
+            return;
+        }
+
+        const request = (coepCredentialless && r.mode === "no-cors")
+            ? new Request(r, {
+                credentials: "omit",
+            })
+            : r;
+        event.respondWith(
+            fetch(request)
+                .then((response) => {
+                    if (response.status === 0) {
+                        return response;
+                    }
+
+                    const newHeaders = new Headers(response.headers);
+                    newHeaders.set("Cross-Origin-Embedder-Policy",
+                        coepCredentialless ? "credentialless" : "require-corp"
+                    );
+                    if (!coepCredentialless) {
+                        newHeaders.set("Cross-Origin-Resource-Policy", "cross-origin");
+                    }
+                    newHeaders.set("Cross-Origin-Opener-Policy", "same-origin");
+
+                    return new Response(response.body, {
+                        status: response.status,
+                        statusText: response.statusText,
+                        headers: newHeaders,
+                    });
+                })
+                .catch((e) => console.error(e))
+        );
+    });
+
+} else {
+    (() => {
+        const reloadedBySelf = window.sessionStorage.getItem("coiReloadedBySelf");
+        window.sessionStorage.removeItem("coiReloadedBySelf");
+        const coepDegrading = (reloadedBySelf == "coepdegrade");
+
+        // You can customize the behavior of this script through a global `coi` variable.
+        const coi = {
+            shouldRegister: () => !reloadedBySelf,
+            shouldDeregister: () => false,
+            coepCredentialless: () => true,
+            coepDegrade: () => true,
+            doReload: () => window.location.reload(),
+            quiet: false,
+            ...window.coi
+        };
+
+        const n = navigator;
+        const controlling = n.serviceWorker && n.serviceWorker.controller;
+
+        // Record the failure if the page is served by serviceWorker.
+        if (controlling && !window.crossOriginIsolated) {
+            window.sessionStorage.setItem("coiCoepHasFailed", "true");
+        }
+        const coepHasFailed = window.sessionStorage.getItem("coiCoepHasFailed");
+
+        if (controlling) {
+            // Reload only on the first failure.
+            const reloadToDegrade = coi.coepDegrade() && !(
+                coepDegrading || window.crossOriginIsolated
+            );
+            n.serviceWorker.controller.postMessage({
+                type: "coepCredentialless",
+                value: (reloadToDegrade || coepHasFailed && coi.coepDegrade())
+                    ? false
+                    : coi.coepCredentialless(),
+            });
+            if (reloadToDegrade) {
+                !coi.quiet && console.log("Reloading page to degrade COEP.");
+                window.sessionStorage.setItem("coiReloadedBySelf", "coepdegrade");
+                coi.doReload("coepdegrade");
+            }
+
+            if (coi.shouldDeregister()) {
+                n.serviceWorker.controller.postMessage({ type: "deregister" });
+            }
+        }
+
+        // If we're already coi: do nothing. Perhaps it's due to this script doing its job, or COOP/COEP are
+        // already set from the origin server. Also if the browser has no notion of crossOriginIsolated, just give up here.
+        if (window.crossOriginIsolated !== false || !coi.shouldRegister()) return;
+
+        if (!window.isSecureContext) {
+            !coi.quiet && console.log("COOP/COEP Service Worker not registered, a secure context is required.");
+            return;
+        }
+
+        // In some environments (e.g. Firefox private mode) this won't be available
+        if (!n.serviceWorker) {
+            !coi.quiet && console.error("COOP/COEP Service Worker not registered, perhaps due to private mode.");
+            return;
+        }
+
+        n.serviceWorker.register(window.document.currentScript.src).then(
+            (registration) => {
+                !coi.quiet && console.log("COOP/COEP Service Worker registered", registration.scope);
+
+                registration.addEventListener("updatefound", () => {
+                    !coi.quiet && console.log("Reloading page to make use of updated COOP/COEP Service Worker.");
+                    window.sessionStorage.setItem("coiReloadedBySelf", "updatefound");
+                    coi.doReload();
+                });
+
+                // If the registration is active, but it's not controlling the page
+                if (registration.active && !n.serviceWorker.controller) {
+                    !coi.quiet && console.log("Reloading page to make use of COOP/COEP Service Worker.");
+                    window.sessionStorage.setItem("coiReloadedBySelf", "notcontrolling");
+                    coi.doReload();
+                }
+            },
+            (err) => {
+                !coi.quiet && console.error("COOP/COEP Service Worker failed to register:", err);
+            }
+        );
+    })();
+}

examples/command.wasm/README.md

@@ -3,7 +3,7 @@
 This is a basic Voice Assistant example that accepts voice commands from the microphone.
 It runs in fully in the browser via WebAseembly.
 
-Online demo: https://whisper.ggerganov.com/command/
+Online demo: https://ggerganov.github.io/whisper.cpp/command.wasm
 
 Terminal version: [examples/command](/examples/command)
 
@@ -15,9 +15,18 @@ git clone https://github.com/ggerganov/whisper.cpp
 cd whisper.cpp
 mkdir build-em && cd build-em
 emcmake cmake ..
-make -j
+make -j libcommand
+```
+The example can then be started by running a local HTTP server:
+```console
+python3 examples/server.py
+```
+And then opening a browser to the following URL:
+http://localhost:8000/command.wasm/
 
-# copy the produced page to your HTTP path
+To run the example in a different server, you need to copy the following files
+to the server's HTTP path:
+```
 cp bin/command.wasm/*       /path/to/html/
 cp bin/libcommand.worker.js /path/to/html/
 ```

examples/command.wasm/index-tmpl.html

@@ -24,6 +24,8 @@
                 overflow-x: scroll;
             }
         </style>
+        <script src="../coi-serviceworker.js"></script>
+        <link rel="icon" href="data:,">
     </head>
     <body>
         <div id="main-container">
@@ -36,11 +38,10 @@
             <br><br>
 
             <b>More examples:</b>
-                <a href="https://whisper.ggerganov.com/">main</a> |
-                <a href="https://whisper.ggerganov.com/bench">bench</a> |
-                <a href="https://whisper.ggerganov.com/stream">stream</a> |
-                <a href="https://whisper.ggerganov.com/command">command</a> |
-                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+                <a href="../">main</a> |
+                <a href="../bench.wasm/">bench</a> |
+                <a href="../stream.wasm">stream</a> |
+                <a href="../command.wasm/">command</a> |
 
             <br><br>
 

examples/command/command.cpp

@@ -11,16 +11,15 @@
 #include "whisper.h"
 #include "grammar-parser.h"
 
-#include <sstream>
-#include <cassert>
+#include <algorithm>
+#include <chrono>
 #include <cstdio>
 #include <fstream>
-#include <mutex>
-#include <regex>
+#include <map>
+#include <sstream>
 #include <string>
 #include <thread>
 #include <vector>
-#include <map>
 
 // command-line parameters
 struct whisper_params {

examples/common-sdl.cpp

@@ -159,15 +159,11 @@ void audio_async::callback(uint8_t * stream, int len) {
 
             memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
             memcpy(&m_audio[0], stream + n0 * sizeof(float), (n_samples - n0) * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
         } else {
             memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
         }
+        m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
+        m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
     }
 }
 

examples/common-whisper.cpp

@@ -0,0 +1,172 @@
+#define _USE_MATH_DEFINES // for M_PI
+
+#include "common-whisper.h"
+
+#include "common.h"
+
+#include "whisper.h"
+
+// third-party utilities
+// use your favorite implementations
+#define STB_VORBIS_HEADER_ONLY
+#include "stb_vorbis.c"    /* Enables Vorbis decoding. */
+
+#ifdef _WIN32
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#endif
+
+#define MA_NO_DEVICE_IO
+#define MA_NO_THREADING
+#define MA_NO_ENCODING
+#define MA_NO_GENERATION
+#define MA_NO_RESOURCE_MANAGER
+#define MA_NO_NODE_GRAPH
+#define MINIAUDIO_IMPLEMENTATION
+#include "miniaudio.h"
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#ifdef _WIN32
+#include <fcntl.h>
+#include <io.h>
+#endif
+
+#include <cstring>
+#include <fstream>
+
+#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
+
+bool read_audio_data(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
+    std::vector<uint8_t> audio_data; // used for pipe input from stdin or ffmpeg decoding output
+
+    ma_result result;
+    ma_decoder_config decoder_config;
+    ma_decoder decoder;
+
+    decoder_config = ma_decoder_config_init(ma_format_f32, stereo ? 2 : 1, WHISPER_SAMPLE_RATE);
+
+    if (fname == "-") {
+		#ifdef _WIN32
+		_setmode(_fileno(stdin), _O_BINARY);
+		#endif
+
+		uint8_t buf[1024];
+		while (true)
+		{
+			const size_t n = fread(buf, 1, sizeof(buf), stdin);
+			if (n == 0) {
+				break;
+			}
+			audio_data.insert(audio_data.end(), buf, buf + n);
+		}
+
+		if ((result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
+
+			fprintf(stderr, "Error: failed to open audio data from stdin (%s)\n", ma_result_description(result));
+
+			return false;
+		}
+
+		fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, audio_data.size());
+    }
+    else if (((result = ma_decoder_init_file(fname.c_str(), &decoder_config, &decoder)) != MA_SUCCESS)) {
+#if defined(WHISPER_FFMPEG)
+		if (ffmpeg_decode_audio(fname, audio_data) != 0) {
+			fprintf(stderr, "error: failed to ffmpeg decode '%s'\n", fname.c_str());
+
+			return false;
+		}
+
+		if ((result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
+			fprintf(stderr, "error: failed to read audio data as wav (%s)\n", ma_result_description(result));
+
+			return false;
+		}
+#else
+		if ((result = ma_decoder_init_memory(fname.c_str(), fname.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
+			fprintf(stderr, "error: failed to read audio data as wav (%s)\n", ma_result_description(result));
+
+			return false;
+		}
+#endif
+    }
+
+    ma_uint64 frame_count;
+    ma_uint64 frames_read;
+
+    if ((result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count)) != MA_SUCCESS) {
+		fprintf(stderr, "error: failed to retrieve the length of the audio data (%s)\n", ma_result_description(result));
+
+		return false;
+    }
+
+    pcmf32.resize(stereo ? frame_count*2 : frame_count);
+
+    if ((result = ma_decoder_read_pcm_frames(&decoder, pcmf32.data(), frame_count, &frames_read)) != MA_SUCCESS) {
+		fprintf(stderr, "error: failed to read the frames of the audio data (%s)\n", ma_result_description(result));
+
+		return false;
+    }
+
+    if (stereo) {
+		pcmf32s.resize(2);
+		pcmf32s[0].resize(frame_count);
+		pcmf32s[1].resize(frame_count);
+		for (uint64_t i = 0; i < frame_count; i++) {
+			pcmf32s[0][i] = pcmf32[2*i];
+			pcmf32s[1][i] = pcmf32[2*i + 1];
+		}
+    }
+
+    ma_decoder_uninit(&decoder);
+
+    return true;
+}
+
+//  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 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;
+}
+
+#undef STB_VORBIS_HEADER_ONLY
+#include "stb_vorbis.c"

examples/common-whisper.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <cstdint>
+
+// 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_audio_data(
+        const std::string & fname,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
+// 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);
+
+// 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/common.cpp

@@ -2,33 +2,18 @@
 
 #include "common.h"
 
-// third-party utilities
-// use your favorite implementations
-#define DR_WAV_IMPLEMENTATION
-#include "dr_wav.h"
-
 #include <cmath>
+#include <codecvt>
 #include <cstring>
 #include <fstream>
-#include <regex>
 #include <locale>
-#include <codecvt>
+#include <regex>
 #include <sstream>
 
 #if defined(_MSC_VER)
 #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
 static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
@@ -262,17 +247,6 @@ std::map<std::string, int32_t> json_parse(const std::string & fname) {
     return result;
 }
 
-std::string convert_to_utf8(const std::wstring & input) {
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
-    return converter.to_bytes(input);
-}
-
-
-std::wstring convert_to_wstring(const std::string & input) {
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
-    return converter.from_bytes(input);
-}
-
 void gpt_split_words(std::string str, std::vector<std::string>& words) {
     const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
     const std::regex re(pattern);
@@ -624,129 +598,6 @@ 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
-
-    if (fname == "-") {
-        {
-            #ifdef _WIN32
-            _setmode(_fileno(stdin), _O_BINARY);
-            #endif
-
-            uint8_t buf[1024];
-            while (true)
-            {
-                const size_t n = fread(buf, 1, sizeof(buf), stdin);
-                if (n == 0) {
-                    break;
-                }
-                wav_data.insert(wav_data.end(), buf, buf + n);
-            }
-        }
-
-        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-            fprintf(stderr, "error: failed to open WAV file from stdin\n");
-            return false;
-        }
-
-        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;
-    }
-
-    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
-
-    std::vector<int16_t> pcm16;
-    pcm16.resize(n*wav.channels);
-    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
-    drwav_uninit(&wav);
-
-    // convert to mono, float
-    pcmf32.resize(n);
-    if (wav.channels == 1) {
-        for (uint64_t i = 0; i < n; i++) {
-            pcmf32[i] = float(pcm16[i])/32768.0f;
-        }
-    } else {
-        for (uint64_t i = 0; i < n; i++) {
-            pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-        }
-    }
-
-    if (stereo) {
-        // convert to stereo, float
-        pcmf32s.resize(2);
-
-        pcmf32s[0].resize(n);
-        pcmf32s[1].resize(n);
-        for (uint64_t i = 0; i < n; i++) {
-            pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
-            pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
-        }
-    }
-
-    return true;
-}
-
 void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
     const float rc = 1.0f / (2.0f * M_PI * cutoff);
     const float dt = 1.0f / sample_rate;
@@ -822,90 +673,7 @@ float similarity(const std::string & s0, const std::string & s1) {
     return 1.0f - (dist / std::max(s0.size(), s1.size()));
 }
 
-bool sam_params_parse(int argc, char ** argv, sam_params & params) {
-    for (int i = 1; i < argc; i++) {
-        std::string arg = argv[i];
-
-        if (arg == "-s" || arg == "--seed") {
-            params.seed = std::stoi(argv[++i]);
-        } 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 == "-i" || arg == "--inp") {
-            params.fname_inp = argv[++i];
-        } else if (arg == "-o" || arg == "--out") {
-            params.fname_out = argv[++i];
-        } else if (arg == "-h" || arg == "--help") {
-            sam_print_usage(argc, argv, params);
-            exit(0);
-        } else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            sam_print_usage(argc, argv, params);
-            exit(0);
-        }
-    }
-
-    return true;
-}
-
-void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -m FNAME, --model FNAME\n");
-    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
-    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
-    fprintf(stderr, "  -o FNAME, --out FNAME\n");
-    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);
+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

@@ -11,8 +11,6 @@
 #include <fstream>
 #include <sstream>
 
-#define COMMON_SAMPLE_RATE 16000
-
 //
 // GPT CLI argument parsing
 //
@@ -136,19 +134,6 @@ 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(
-        const std::string & fname,
-        std::vector<float> & pcmf32,
-        std::vector<std::vector<float>> & pcmf32s,
-        bool stereo);
-
 // Write PCM data into WAV audio file
 class wav_writer {
 private:
@@ -266,23 +251,6 @@ bool vad_simple(
 // compute similarity between two strings using Levenshtein distance
 float similarity(const std::string & s0, const std::string & s1);
 
-//
-// SAM argument parsing
-//
-
-struct sam_params {
-    int32_t seed      = -1; // RNG seed
-    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-
-    std::string model     = "models/sam-vit-b/ggml-model-f16.bin"; // model path
-    std::string fname_inp = "img.jpg";
-    std::string fname_out = "img.out";
-};
-
-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
 //
@@ -330,14 +298,5 @@ const std::vector<std::string> k_colors = {
 // 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);
+bool is_file_exist(const char * filename);

examples/deprecation-warning/CMakeLists.txt

@@ -1,4 +1,6 @@
 add_executable(main ./deprecation-warning.cpp)
 add_executable(bench ./deprecation-warning.cpp)
-add_executable(stream ./deprecation-warning.cpp)
-add_executable(command ./deprecation-warning.cpp)
+if (WHISPER_SDL2)
+    add_executable(stream ./deprecation-warning.cpp)
+    add_executable(command ./deprecation-warning.cpp)
+endif()

examples/generate-karaoke.sh

@@ -41,20 +41,17 @@ fi
 # record some raw audio
 sox -d rec.wav
 
-# resample to 16kHz
-ffmpeg -y -i ./rec.wav -ar 16000 -ac 1 -c:a pcm_s16le ./rec16.wav > /dev/null 2>&1
-
 # run Whisper
 echo "Processing ..."
-${executable} -m models/ggml-base.en.bin rec16.wav -owts > /dev/null 2>&1
+${executable} -m models/ggml-base.en.bin rec.wav -owts > /dev/null 2>&1
 
 # generate Karaoke video
 echo "Generating video ..."
-source rec16.wav.wts > /dev/null 2>&1
+source rec.wav.wts > /dev/null 2>&1
 
 # play the video
 echo "Playing ./rec16.wav.mp4 ..."
-ffplay -loglevel 0 -autoexit ./rec16.wav.mp4
+ffplay -loglevel 0 -autoexit ./rec.wav.mp4
 
 echo "Done"
 exit 0

examples/lsp/lsp.cpp

@@ -3,14 +3,15 @@
 #include "whisper.h"
 #include "json.hpp"
 
-#include <iostream>
 #include <cassert>
+#include <chrono>
 #include <cstdio>
+#include <deque>
+#include <iostream>
+#include <set>
 #include <string>
 #include <thread>
 #include <vector>
-#include <deque>
-#include <set>
 
 using json = nlohmann::json;
 

examples/server.py

@@ -0,0 +1,39 @@
+import http.server
+import socketserver
+import os
+from pathlib import Path
+
+SCRIPT_DIR = Path(__file__).parent.absolute()
+DIRECTORY = os.path.join(SCRIPT_DIR, "../build-em/bin")
+DIRECTORY = os.path.abspath(DIRECTORY)
+
+class CustomHTTPRequestHandler(http.server.SimpleHTTPRequestHandler):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, directory=DIRECTORY, **kwargs)
+
+    def do_GET(self):
+        # If requesting a worker file from any subdirectory
+        if '.worker.js' in self.path:
+            worker_file = os.path.basename(self.path)
+            worker_path = os.path.join(DIRECTORY, worker_file)
+
+            if os.path.exists(worker_path):
+                self.path = '/' + worker_file
+
+        return super().do_GET()
+
+    def end_headers(self):
+        # Add required headers for SharedArrayBuffer
+        self.send_header("Cross-Origin-Opener-Policy", "same-origin")
+        self.send_header("Cross-Origin-Embedder-Policy", "require-corp")
+        self.send_header("Access-Control-Allow-Origin", "*");
+        super().end_headers()
+
+PORT = 8000
+
+with socketserver.TCPServer(("", PORT), CustomHTTPRequestHandler) as httpd:
+    print(f"Serving directory '{DIRECTORY}' at http://localhost:{PORT}")
+    try:
+        httpd.serve_forever()
+    except KeyboardInterrupt:
+        print("\nServer stopped.")

examples/server/server.cpp

@@ -1,17 +1,18 @@
 #include "common.h"
+#include "common-whisper.h"
 
 #include "whisper.h"
 #include "httplib.h"
 #include "json.hpp"
 
+#include <chrono>
 #include <cmath>
-#include <fstream>
 #include <cstdio>
+#include <fstream>
+#include <sstream>
 #include <string>
 #include <thread>
 #include <vector>
-#include <cstring>
-#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -223,6 +224,24 @@ void check_ffmpeg_availibility() {
     }
 }
 
+std::string generate_temp_filename(const std::string &prefix, const std::string &extension) {
+    auto now = std::chrono::system_clock::now();
+    auto now_time_t = std::chrono::system_clock::to_time_t(now);
+
+    static std::mt19937 rng{std::random_device{}()};
+    std::uniform_int_distribution<long long> dist(0, 1e9);
+
+    std::stringstream ss;
+    ss << prefix
+       << "-"
+       << std::put_time(std::localtime(&now_time_t), "%Y%m%d-%H%M%S")
+       << "-"
+       << dist(rng)
+       << extension;
+
+    return ss.str();
+}
+
 bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
     std::ostringstream cmd_stream;
     std::string converted_filename_temp = temp_filename + "_temp.wav";
@@ -692,9 +711,7 @@ int main(int argc, char ** argv) {
         if (sparams.ffmpeg_converter) {
             // if file is not wav, convert to wav
             // write to temporary file
-            //const std::string temp_filename_base = std::tmpnam(nullptr);
-            const std::string temp_filename_base = "whisper-server-tmp"; // TODO: this is a hack, remove when the mutext is removed
-            const std::string temp_filename = temp_filename_base + ".wav";
+            const std::string temp_filename = generate_temp_filename("whisper-server", ".wav");
             std::ofstream temp_file{temp_filename, std::ios::binary};
             temp_file << audio_file.content;
             temp_file.close();
@@ -706,8 +723,8 @@ int main(int argc, char ** argv) {
                 return;
             }
 
-            // read wav content into pcmf32
-            if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize))
+            // read audio content into pcmf32
+            if (!::read_audio_data(temp_filename, pcmf32, pcmf32s, params.diarize))
             {
                 fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
                 const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
@@ -718,10 +735,10 @@ int main(int argc, char ** argv) {
             // remove temp file
             std::remove(temp_filename.c_str());
         } else {
-            if (!::read_wav(audio_file.content, pcmf32, pcmf32s, params.diarize))
+            if (!::read_audio_data(audio_file.content, pcmf32, pcmf32s, params.diarize))
             {
-                fprintf(stderr, "error: failed to read WAV file\n");
-                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+                fprintf(stderr, "error: failed to read audio data\n");
+                const std::string error_resp = "{\"error\":\"failed to read audio data\"}";
                 res.set_content(error_resp, "application/json");
                 return;
             }
@@ -1007,6 +1024,11 @@ int main(int argc, char ** argv) {
         // check if the model is in the file system
     });
 
+    svr.Get(sparams.request_path + "/health", [&](const Request &, Response &res){
+        const std::string health_response = "{\"status\":\"ok\"}";
+        res.set_content(health_response, "application/json");
+    });
+
     svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) {
         const char fmt[] = "500 Internal Server Error\n%s";
         char buf[BUFSIZ];

examples/stream.wasm/README.md

@@ -13,7 +13,17 @@ cd whisper.cpp
 mkdir build-em && cd build-em
 emcmake cmake ..
 make -j
+```
+The example can then be started by running a local HTTP server:
+```console
+python3 examples/server.py
+```
+And then opening a browser to the following URL:
+http://localhost:8000/stream.wasm
 
+To run the example in a different server, you need to copy the following files
+to the server's HTTP path:
+```
 # copy the produced page to your HTTP path
 cp bin/stream.wasm/*       /path/to/html/
 cp bin/libstream.worker.js /path/to/html/

examples/stream.wasm/index-tmpl.html

@@ -24,6 +24,8 @@
                 overflow-x: scroll;
             }
         </style>
+        <script src="../coi-serviceworker.js"></script>
+        <link rel="icon" href="data:,">
     </head>
     <body>
         <div id="main-container">
@@ -36,11 +38,10 @@
             <br><br>
 
             <b>More examples:</b>
-                <a href="https://whisper.ggerganov.com/">main</a> |
-                <a href="https://whisper.ggerganov.com/bench">bench</a> |
-                <a href="https://whisper.ggerganov.com/stream">stream</a> |
-                <a href="https://whisper.ggerganov.com/command">command</a> |
-                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+                <a href="../">main</a> |
+                <a href="../bench.wasm/">bench</a> |
+                <a href="../stream.wasm">stream</a> |
+                <a href="../command.wasm/">command</a> |
 
             <br><br>
 

examples/stream/stream.cpp

@@ -4,15 +4,15 @@
 //
 #include "common-sdl.h"
 #include "common.h"
+#include "common-whisper.h"
 #include "whisper.h"
 
-#include <cassert>
+#include <chrono>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <thread>
 #include <vector>
-#include <fstream>
-
 
 // command-line parameters
 struct whisper_params {
@@ -23,6 +23,7 @@ struct whisper_params {
     int32_t capture_id = -1;
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
+    int32_t beam_size  = -1;
 
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
@@ -59,6 +60,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (arg == "-c"    || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
         else if (arg == "-mt"   || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
         else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")     { params.beam_size     = 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 == "-tr"   || arg == "--translate")     { params.translate     = true; }
@@ -96,6 +98,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                              params.capture_id);
     fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",       params.max_tokens);
     fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -bs N,    --beam-size N   [%-7d] beam size for beam search\n",                      params.beam_size);
     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, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
@@ -241,6 +244,11 @@ int main(int argc, char ** argv) {
 
         if (!use_vad) {
             while (true) {
+                // handle Ctrl + C
+                is_running = sdl_poll_events();
+                if (!is_running) {
+                    break;
+                }
                 audio.get(params.step_ms, pcmf32_new);
 
                 if ((int) pcmf32_new.size() > 2*n_samples_step) {
@@ -298,7 +306,7 @@ int main(int argc, char ** argv) {
 
         // run the inference
         {
-            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+            whisper_full_params wparams = whisper_full_default_params(params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY);
 
             wparams.print_progress   = false;
             wparams.print_special    = params.print_special;
@@ -309,6 +317,7 @@ int main(int argc, char ** argv) {
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();
             wparams.n_threads        = params.n_threads;
+            wparams.beam_search.beam_size = params.beam_size;
 
             wparams.audio_ctx        = params.audio_ctx;
 

examples/talk-llama/CMakeLists.txt

@@ -1,18 +1,31 @@
 if (WHISPER_SDL2)
+    set(CMAKE_CXX_STANDARD 17)
+    set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
     set(TARGET whisper-talk-llama)
     add_executable(${TARGET} talk-llama.cpp
         llama.cpp
-        llama-vocab.cpp
+        llama-adapter.cpp
+        llama-arch.cpp
+        llama-batch.cpp
+        llama-chat.cpp
+        llama-context.cpp
+        llama-cparams.cpp
         llama-grammar.cpp
+        llama-hparams.cpp
+        llama-impl.cpp
+        llama-kv-cache.cpp
+        llama-mmap.cpp
+        llama-model-loader.cpp
+        llama-model.cpp
+        llama-quant.cpp
         llama-sampling.cpp
+        llama-vocab.cpp
         unicode.cpp
         unicode-data.cpp)
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
 
-    if (WHISPER_CLBLAST)
-        set(CLBLAST_LIBNAME clblast)
-    endif ()
-    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CLBLAST_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 
     if(WIN32)
         # It requires Windows 8.1 or later for PrefetchVirtualMemory

examples/talk-llama/llama-adapter.cpp

@@ -0,0 +1,347 @@
+#include "llama-adapter.h"
+
+#include "llama-impl.h"
+#include "llama-mmap.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <map>
+#include <cassert>
+#include <stdexcept>
+
+// vec
+
+struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
+    if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
+        return nullptr;
+    }
+
+    return tensors[il];
+}
+
+struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+    ggml_tensor * layer_dir = tensor_for(il);
+    if (layer_dir != nullptr) {
+        cur = ggml_add(ctx, cur, layer_dir);
+    }
+
+    return cur;
+}
+
+bool llama_adapter_cvec::init(const llama_model & model) {
+    const auto & hparams = model.hparams;
+
+    GGML_ASSERT(tensors.empty());
+    GGML_ASSERT(ctxs.empty());
+    GGML_ASSERT(bufs.empty());
+
+    // create a context for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+
+            ggml_context * ctx = ggml_init(params);
+            if (!ctx) {
+                return nullptr;
+            }
+
+            ctx_map[buft] = ctx;
+            ctxs.emplace_back(ctx);
+
+            return ctx;
+        }
+
+        return it->second;
+    };
+
+    // make tensors
+    tensors.reserve(hparams.n_layer);
+    tensors.push_back(nullptr); // there's never a tensor for layer 0
+    for (size_t il = 1; il < hparams.n_layer; il++) {
+        ggml_backend_buffer_type_t buft = model.select_buft(il);
+        ggml_context * ctx = ctx_for_buft(buft);
+        if (!ctx) {
+            LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
+            return false;
+        }
+        ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
+        tensors.push_back(tensor);
+    }
+
+    // allocate tensors / buffers and zero
+    bufs.reserve(ctx_map.size());
+    for (auto it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx = it.second;
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        if (!buf) {
+            LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__);
+            return false;
+        }
+        ggml_backend_buffer_clear(buf, 0);
+        bufs.emplace_back(buf);
+    }
+
+    return true;
+}
+
+int32_t llama_adapter_cvec::apply(
+        const llama_model & model,
+        const float * data,
+        size_t len,
+        int32_t n_embd,
+        int32_t il_start,
+        int32_t il_end) {
+    const auto & hparams = model.hparams;
+
+    if (data == nullptr) {
+        // disable the current control vector (but leave allocated for later)
+        layer_start = -1;
+        layer_end   = -1;
+        return 0;
+    }
+
+    if (n_embd != (int) hparams.n_embd) {
+        LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
+        return 1;
+    }
+
+    if (tensors.empty()) {
+        if (!init(model)) {
+            return 1;
+        }
+    }
+
+    layer_start = il_start;
+    layer_end   = il_end;
+
+    for (size_t il = 1; il < hparams.n_layer; il++) {
+        assert(tensors[il] != nullptr);
+
+        const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
+        if (off + n_embd <= len) {
+            ggml_backend_tensor_set(tensors[il], data + off, 0, n_embd * ggml_element_size(tensors[il]));
+        }
+    }
+
+    return 0;
+}
+
+// lora
+
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) {
+    const std::string name(w->name);
+
+    const auto pos = ab_map.find(name);
+    if (pos != ab_map.end()) {
+        return &pos->second;
+    }
+
+    return nullptr;
+}
+
+static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) {
+    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
+
+    ggml_context * ctx_init;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ true,
+        /* .ctx      = */ &ctx_init,
+    };
+
+    gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) };
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
+    }
+
+    ggml_context_ptr ctx { ctx_init };
+
+    // check metadata
+    {
+        auto get_kv_str = [&](const std::string & key) -> std::string {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
+        };
+        auto get_kv_f32 = [&](const std::string & key) -> float {
+            int id = gguf_find_key(ctx_gguf.get(), key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
+        };
+        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
+
+        auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE));
+        if (general_type != "adapter") {
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
+
+        auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
+        auto general_arch = llm_arch_from_string(general_arch_str);
+        if (general_arch != model.arch) {
+            throw std::runtime_error("model arch and LoRA arch mismatch");
+        }
+
+        auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE));
+        if (adapter_type != "lora") {
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
+        }
+
+        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+    }
+
+    int n_tensors = gguf_get_n_tensors(ctx_gguf.get());
+
+    // contexts for each buffer type
+    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            // add a new context
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * buft_ctx = ggml_init(params);
+            if (!buft_ctx) {
+                return nullptr;
+            }
+            ctx_map[buft] = buft_ctx;
+            adapter.ctxs.emplace_back(buft_ctx);
+            return buft_ctx;
+        };
+        return it->second;
+    };
+
+    // bundle lora_a and lora_b into pairs
+    std::map<std::string, llama_adapter_lora_weight> ab_map;
+    auto str_endswith = [](const std::string & str, const std::string & suffix) {
+        return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+    };
+
+    for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) {
+        std::string name(cur->name);
+        if (str_endswith(name, ".lora_a")) {
+            replace_all(name, ".lora_a", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_adapter_lora_weight(cur, nullptr);
+            } else {
+                ab_map[name].a = cur;
+            }
+        } else if (str_endswith(name, ".lora_b")) {
+            replace_all(name, ".lora_b", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_adapter_lora_weight(nullptr, cur);
+            } else {
+                ab_map[name].b = cur;
+            }
+        } else if (str_endswith(name, "_norm.weight")) {
+            // TODO: add support for norm vector
+            // for now, we don't really care because most adapters still work fine without it
+            continue;
+        } else {
+            throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
+        }
+    }
+
+    // add tensors
+    for (auto & it : ab_map) {
+        const std::string & name = it.first;
+        llama_adapter_lora_weight & w = it.second;
+        bool is_token_embd = str_endswith(name, "token_embd.weight");
+
+        if (!w.a || !w.b) {
+            throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
+        }
+
+        // device buft and device ctx
+        const auto * model_tensor = model.get_tensor(name.c_str());
+        if (!model_tensor) {
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
+        }
+
+        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        // validate tensor shape
+        if (is_token_embd) {
+            // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
+            if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) {
+                throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
+            }
+        } else {
+            if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
+                throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
+            }
+            if (w.a->ne[1] != w.b->ne[0]) {
+                throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+            }
+        }
+
+        // save tensor to adapter
+        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_set_name(tensor_a, w.a->name);
+        ggml_set_name(tensor_b, w.b->name);
+        adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
+    }
+
+    // allocate tensors / buffers and zero
+    {
+        adapter.ctxs.reserve(ctx_map.size());
+        adapter.bufs.reserve(ctx_map.size());
+        for (auto & it : ctx_map) {
+            ggml_backend_buffer_type_t buft = it.first;
+            ggml_context * ctx_dev = it.second;
+            ggml_backend_buffer_ptr buf { ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft) };
+            if (!buf) {
+                throw std::runtime_error("failed to allocate buffer for lora adapter\n");
+            }
+            LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get())/1024.0/1024.0);
+            adapter.bufs.emplace_back(std::move(buf));
+        }
+    }
+
+    // set tensor data
+    {
+        llama_file gguf_file(path_lora, "rb");
+        std::vector<uint8_t> read_buf;
+        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+            size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
+            size_t size = ggml_nbytes(orig);
+            read_buf.resize(size);
+            gguf_file.seek(offs, SEEK_SET);
+            gguf_file.read_raw(read_buf.data(), size);
+            ggml_backend_tensor_set(dev, read_buf.data(), 0, size);
+        };
+        for (auto & it : adapter.ab_map) {
+            auto orig = ab_map[it.first];
+            auto dev  = it.second;
+            set_tensor(orig.a, dev.a);
+            set_tensor(orig.b, dev.b);
+        }
+    }
+
+    LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
+}
+
+struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) {
+    struct llama_adapter_lora * adapter = new llama_adapter_lora();
+
+    try {
+        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
+        return adapter;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
+
+        delete adapter;
+    }
+
+    return nullptr;
+}
+
+void llama_adapter_lora_free(struct llama_adapter_lora * adapter) {
+    delete adapter;
+}

examples/talk-llama/llama-adapter.h

@@ -0,0 +1,74 @@
+#pragma once
+
+#include "llama.h"
+
+#include "ggml-cpp.h"
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+// TODO: pimpl
+
+//
+// llama_adapter_cvec
+//
+
+struct llama_adapter_cvec {
+    struct ggml_tensor * tensor_for(int il) const;
+
+    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+
+    int32_t apply(
+            const llama_model & model,
+            const float * data,
+            size_t len,
+            int32_t n_embd,
+            int32_t il_start,
+            int32_t il_end);
+
+private:
+    bool init(const llama_model & model);
+
+    int32_t layer_start = -1;
+    int32_t layer_end   = -1;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    std::vector<struct ggml_tensor *> tensors; // per layer
+};
+
+//
+// llama_adapter_lora
+//
+
+struct llama_adapter_lora_weight {
+    struct ggml_tensor * a = nullptr;
+    struct ggml_tensor * b = nullptr;
+
+    // get actual scale based on rank and alpha
+    float get_scale(float alpha, float adapter_scale) const {
+        const float rank  = (float) b->ne[0];
+        const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
+        return scale;
+    }
+
+    llama_adapter_lora_weight() = default;
+    llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+};
+
+struct llama_adapter_lora {
+    // map tensor name to lora_a_b
+    std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    float alpha;
+
+    llama_adapter_lora() = default;
+    ~llama_adapter_lora() = default;
+
+    llama_adapter_lora_weight * get_weight(struct ggml_tensor * w);
+};

examples/talk-llama/llama-arch.h

@@ -0,0 +1,402 @@
+#pragma once
+
+#include "ggml.h" // ggml_op
+
+#include <string>
+
+//
+// gguf constants (sync with gguf.py)
+//
+
+enum llm_arch {
+    LLM_ARCH_LLAMA,
+    LLM_ARCH_DECI,
+    LLM_ARCH_FALCON,
+    LLM_ARCH_BAICHUAN,
+    LLM_ARCH_GROK,
+    LLM_ARCH_GPT2,
+    LLM_ARCH_GPTJ,
+    LLM_ARCH_GPTNEOX,
+    LLM_ARCH_MPT,
+    LLM_ARCH_STARCODER,
+    LLM_ARCH_REFACT,
+    LLM_ARCH_BERT,
+    LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_JINA_BERT_V2,
+    LLM_ARCH_BLOOM,
+    LLM_ARCH_STABLELM,
+    LLM_ARCH_QWEN,
+    LLM_ARCH_QWEN2,
+    LLM_ARCH_QWEN2MOE,
+    LLM_ARCH_QWEN2VL,
+    LLM_ARCH_PHI2,
+    LLM_ARCH_PHI3,
+    LLM_ARCH_PHIMOE,
+    LLM_ARCH_PLAMO,
+    LLM_ARCH_CODESHELL,
+    LLM_ARCH_ORION,
+    LLM_ARCH_INTERNLM2,
+    LLM_ARCH_MINICPM,
+    LLM_ARCH_MINICPM3,
+    LLM_ARCH_GEMMA,
+    LLM_ARCH_GEMMA2,
+    LLM_ARCH_STARCODER2,
+    LLM_ARCH_MAMBA,
+    LLM_ARCH_XVERSE,
+    LLM_ARCH_COMMAND_R,
+    LLM_ARCH_COHERE2,
+    LLM_ARCH_DBRX,
+    LLM_ARCH_OLMO,
+    LLM_ARCH_OLMO2,
+    LLM_ARCH_OLMOE,
+    LLM_ARCH_OPENELM,
+    LLM_ARCH_ARCTIC,
+    LLM_ARCH_DEEPSEEK,
+    LLM_ARCH_DEEPSEEK2,
+    LLM_ARCH_CHATGLM,
+    LLM_ARCH_BITNET,
+    LLM_ARCH_T5,
+    LLM_ARCH_T5ENCODER,
+    LLM_ARCH_JAIS,
+    LLM_ARCH_NEMOTRON,
+    LLM_ARCH_EXAONE,
+    LLM_ARCH_RWKV6,
+    LLM_ARCH_RWKV6QWEN2,
+    LLM_ARCH_GRANITE,
+    LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_CHAMELEON,
+    LLM_ARCH_WAVTOKENIZER_DEC,
+    LLM_ARCH_UNKNOWN,
+};
+
+enum llm_kv {
+    LLM_KV_GENERAL_TYPE,
+    LLM_KV_GENERAL_ARCHITECTURE,
+    LLM_KV_GENERAL_QUANTIZATION_VERSION,
+    LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_NAME,
+    LLM_KV_GENERAL_AUTHOR,
+    LLM_KV_GENERAL_VERSION,
+    LLM_KV_GENERAL_URL,
+    LLM_KV_GENERAL_DESCRIPTION,
+    LLM_KV_GENERAL_LICENSE,
+    LLM_KV_GENERAL_SOURCE_URL,
+    LLM_KV_GENERAL_SOURCE_HF_REPO,
+
+    LLM_KV_VOCAB_SIZE,
+    LLM_KV_CONTEXT_LENGTH,
+    LLM_KV_EMBEDDING_LENGTH,
+    LLM_KV_FEATURES_LENGTH,
+    LLM_KV_BLOCK_COUNT,
+    LLM_KV_LEADING_DENSE_BLOCK_COUNT,
+    LLM_KV_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH,
+    LLM_KV_USE_PARALLEL_RESIDUAL,
+    LLM_KV_TENSOR_DATA_LAYOUT,
+    LLM_KV_EXPERT_COUNT,
+    LLM_KV_EXPERT_USED_COUNT,
+    LLM_KV_EXPERT_SHARED_COUNT,
+    LLM_KV_EXPERT_WEIGHTS_SCALE,
+    LLM_KV_EXPERT_WEIGHTS_NORM,
+    LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_POOLING_TYPE,
+    LLM_KV_LOGIT_SCALE,
+    LLM_KV_DECODER_START_TOKEN_ID,
+    LLM_KV_ATTN_LOGIT_SOFTCAPPING,
+    LLM_KV_FINAL_LOGIT_SOFTCAPPING,
+    LLM_KV_SWIN_NORM,
+    LLM_KV_RESCALE_EVERY_N_LAYERS,
+    LLM_KV_TIME_MIX_EXTRA_DIM,
+    LLM_KV_TIME_DECAY_EXTRA_DIM,
+    LLM_KV_RESIDUAL_SCALE,
+    LLM_KV_EMBEDDING_SCALE,
+    LLM_KV_TOKEN_SHIFT_COUNT,
+
+    LLM_KV_ATTENTION_HEAD_COUNT,
+    LLM_KV_ATTENTION_HEAD_COUNT_KV,
+    LLM_KV_ATTENTION_MAX_ALIBI_BIAS,
+    LLM_KV_ATTENTION_CLAMP_KQV,
+    LLM_KV_ATTENTION_KEY_LENGTH,
+    LLM_KV_ATTENTION_VALUE_LENGTH,
+    LLM_KV_ATTENTION_LAYERNORM_EPS,
+    LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
+    LLM_KV_ATTENTION_GROUPNORM_EPS,
+    LLM_KV_ATTENTION_GROUPNORM_GROUPS,
+    LLM_KV_ATTENTION_CAUSAL,
+    LLM_KV_ATTENTION_Q_LORA_RANK,
+    LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
+    LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SCALE,
+
+    LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_SECTIONS,
+    LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_SCALE_LINEAR,
+    LLM_KV_ROPE_SCALING_TYPE,
+    LLM_KV_ROPE_SCALING_FACTOR,
+    LLM_KV_ROPE_SCALING_ATTN_FACTOR,
+    LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
+    LLM_KV_ROPE_SCALING_FINETUNED,
+    LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
+
+    LLM_KV_SPLIT_NO,
+    LLM_KV_SPLIT_COUNT,
+    LLM_KV_SPLIT_TENSORS_COUNT,
+
+    LLM_KV_SSM_INNER_SIZE,
+    LLM_KV_SSM_CONV_KERNEL,
+    LLM_KV_SSM_STATE_SIZE,
+    LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_DT_B_C_RMS,
+
+    LLM_KV_WKV_HEAD_SIZE,
+
+    LLM_KV_TOKENIZER_MODEL,
+    LLM_KV_TOKENIZER_PRE,
+    LLM_KV_TOKENIZER_LIST,
+    LLM_KV_TOKENIZER_TOKEN_TYPE,
+    LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,
+    LLM_KV_TOKENIZER_SCORES,
+    LLM_KV_TOKENIZER_MERGES,
+    LLM_KV_TOKENIZER_BOS_ID,
+    LLM_KV_TOKENIZER_EOS_ID,
+    LLM_KV_TOKENIZER_EOT_ID,
+    LLM_KV_TOKENIZER_EOM_ID,
+    LLM_KV_TOKENIZER_UNK_ID,
+    LLM_KV_TOKENIZER_SEP_ID,
+    LLM_KV_TOKENIZER_PAD_ID,
+    LLM_KV_TOKENIZER_CLS_ID,
+    LLM_KV_TOKENIZER_MASK_ID,
+    LLM_KV_TOKENIZER_ADD_BOS,
+    LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_PREFIX,
+    LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
+    LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
+    LLM_KV_TOKENIZER_HF_JSON,
+    LLM_KV_TOKENIZER_RWKV,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
+    LLM_KV_TOKENIZER_FIM_PRE_ID,
+    LLM_KV_TOKENIZER_FIM_SUF_ID,
+    LLM_KV_TOKENIZER_FIM_MID_ID,
+    LLM_KV_TOKENIZER_FIM_PAD_ID,
+    LLM_KV_TOKENIZER_FIM_REP_ID,
+    LLM_KV_TOKENIZER_FIM_SEP_ID,
+
+    LLM_KV_ADAPTER_TYPE,
+    LLM_KV_ADAPTER_LORA_ALPHA,
+
+    LLM_KV_POSNET_EMBEDDING_LENGTH,
+    LLM_KV_POSNET_BLOCK_COUNT,
+
+    LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
+    LLM_KV_CONVNEXT_BLOCK_COUNT,
+
+    // deprecated:
+    LLM_KV_TOKENIZER_PREFIX_ID,
+    LLM_KV_TOKENIZER_SUFFIX_ID,
+    LLM_KV_TOKENIZER_MIDDLE_ID,
+};
+
+enum llm_tensor {
+    LLM_TENSOR_TOKEN_EMBD,
+    LLM_TENSOR_TOKEN_EMBD_NORM,
+    LLM_TENSOR_TOKEN_TYPES,
+    LLM_TENSOR_POS_EMBD,
+    LLM_TENSOR_OUTPUT,
+    LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_ROPE_FREQS,
+    LLM_TENSOR_ROPE_FACTORS_LONG,
+    LLM_TENSOR_ROPE_FACTORS_SHORT,
+    LLM_TENSOR_ATTN_Q,
+    LLM_TENSOR_ATTN_K,
+    LLM_TENSOR_ATTN_V,
+    LLM_TENSOR_ATTN_QKV,
+    LLM_TENSOR_ATTN_OUT,
+    LLM_TENSOR_ATTN_NORM,
+    LLM_TENSOR_ATTN_NORM_2,
+    LLM_TENSOR_ATTN_OUT_NORM,
+    LLM_TENSOR_ATTN_POST_NORM,
+    LLM_TENSOR_ATTN_ROT_EMBD,
+    LLM_TENSOR_FFN_GATE_INP,
+    LLM_TENSOR_FFN_GATE_INP_SHEXP,
+    LLM_TENSOR_FFN_NORM,
+    LLM_TENSOR_FFN_POST_NORM,
+    LLM_TENSOR_FFN_GATE,
+    LLM_TENSOR_FFN_DOWN,
+    LLM_TENSOR_FFN_UP,
+    LLM_TENSOR_FFN_ACT,
+    LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
+    LLM_TENSOR_FFN_GATE_EXP,
+    LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_NORM_EXPS,
+    LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
+    LLM_TENSOR_FFN_GATE_EXPS,
+    LLM_TENSOR_FFN_UP_EXPS,
+    LLM_TENSOR_FFN_DOWN_SHEXP,
+    LLM_TENSOR_FFN_GATE_SHEXP,
+    LLM_TENSOR_FFN_UP_SHEXP,
+    LLM_TENSOR_FFN_EXP_PROBS_B,
+    LLM_TENSOR_ATTN_Q_NORM,
+    LLM_TENSOR_ATTN_K_NORM,
+    LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_SSM_IN,
+    LLM_TENSOR_SSM_CONV1D,
+    LLM_TENSOR_SSM_X,
+    LLM_TENSOR_SSM_DT,
+    LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W1,
+    LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_LERP_X,
+    LLM_TENSOR_TIME_MIX_LERP_W,
+    LLM_TENSOR_TIME_MIX_LERP_K,
+    LLM_TENSOR_TIME_MIX_LERP_V,
+    LLM_TENSOR_TIME_MIX_LERP_R,
+    LLM_TENSOR_TIME_MIX_LERP_G,
+    LLM_TENSOR_TIME_MIX_LERP_FUSED,
+    LLM_TENSOR_TIME_MIX_FIRST,
+    LLM_TENSOR_TIME_MIX_DECAY,
+    LLM_TENSOR_TIME_MIX_DECAY_W1,
+    LLM_TENSOR_TIME_MIX_DECAY_W2,
+    LLM_TENSOR_TIME_MIX_KEY,
+    LLM_TENSOR_TIME_MIX_VALUE,
+    LLM_TENSOR_TIME_MIX_RECEPTANCE,
+    LLM_TENSOR_TIME_MIX_GATE,
+    LLM_TENSOR_TIME_MIX_LN,
+    LLM_TENSOR_TIME_MIX_OUTPUT,
+    LLM_TENSOR_CHANNEL_MIX_LERP_K,
+    LLM_TENSOR_CHANNEL_MIX_LERP_R,
+    LLM_TENSOR_CHANNEL_MIX_KEY,
+    LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,
+    LLM_TENSOR_CHANNEL_MIX_VALUE,
+    LLM_TENSOR_ATTN_Q_A,
+    LLM_TENSOR_ATTN_Q_B,
+    LLM_TENSOR_ATTN_KV_A_MQA,
+    LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_Q_A_NORM,
+    LLM_TENSOR_ATTN_KV_A_NORM,
+    LLM_TENSOR_ATTN_SUB_NORM,
+    LLM_TENSOR_FFN_SUB_NORM,
+    LLM_TENSOR_DEC_ATTN_NORM,
+    LLM_TENSOR_DEC_ATTN_Q,
+    LLM_TENSOR_DEC_ATTN_K,
+    LLM_TENSOR_DEC_ATTN_V,
+    LLM_TENSOR_DEC_ATTN_OUT,
+    LLM_TENSOR_DEC_ATTN_REL_B,
+    LLM_TENSOR_DEC_CROSS_ATTN_NORM,
+    LLM_TENSOR_DEC_CROSS_ATTN_Q,
+    LLM_TENSOR_DEC_CROSS_ATTN_K,
+    LLM_TENSOR_DEC_CROSS_ATTN_V,
+    LLM_TENSOR_DEC_CROSS_ATTN_OUT,
+    LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
+    LLM_TENSOR_DEC_FFN_NORM,
+    LLM_TENSOR_DEC_FFN_GATE,
+    LLM_TENSOR_DEC_FFN_DOWN,
+    LLM_TENSOR_DEC_FFN_UP,
+    LLM_TENSOR_DEC_OUTPUT_NORM,
+    LLM_TENSOR_ENC_ATTN_NORM,
+    LLM_TENSOR_ENC_ATTN_Q,
+    LLM_TENSOR_ENC_ATTN_K,
+    LLM_TENSOR_ENC_ATTN_V,
+    LLM_TENSOR_ENC_ATTN_OUT,
+    LLM_TENSOR_ENC_ATTN_REL_B,
+    LLM_TENSOR_ENC_FFN_NORM,
+    LLM_TENSOR_ENC_FFN_GATE,
+    LLM_TENSOR_ENC_FFN_DOWN,
+    LLM_TENSOR_ENC_FFN_UP,
+    LLM_TENSOR_ENC_OUTPUT_NORM,
+    LLM_TENSOR_CLS,
+    LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_CONV1D,
+    LLM_TENSOR_CONVNEXT_DW,
+    LLM_TENSOR_CONVNEXT_NORM,
+    LLM_TENSOR_CONVNEXT_PW1,
+    LLM_TENSOR_CONVNEXT_PW2,
+    LLM_TENSOR_CONVNEXT_GAMMA,
+    LLM_TENSOR_POS_NET_CONV1,
+    LLM_TENSOR_POS_NET_CONV2,
+    LLM_TENSOR_POS_NET_NORM,
+    LLM_TENSOR_POS_NET_NORM1,
+    LLM_TENSOR_POS_NET_NORM2,
+    LLM_TENSOR_POS_NET_ATTN_NORM,
+    LLM_TENSOR_POS_NET_ATTN_Q,
+    LLM_TENSOR_POS_NET_ATTN_K,
+    LLM_TENSOR_POS_NET_ATTN_V,
+    LLM_TENSOR_POS_NET_ATTN_OUT,
+};
+
+enum llm_tensor_layer {
+    LLM_TENSOR_LAYER_INPUT,
+    LLM_TENSOR_LAYER_REPEATING,
+    LLM_TENSOR_LAYER_OUTPUT,
+};
+
+struct LLM_KV {
+    LLM_KV(llm_arch arch, const char * suffix = nullptr);
+
+    llm_arch arch;
+    const char * suffix;
+
+    std::string operator()(llm_kv kv) const;
+};
+
+// helper to handle gguf constants
+// usage:
+//
+//   const auto tn = LLM_TN(LLM_ARCH_LLAMA);
+//
+//   std::string name = tn(LLM_TENSOR_OUTPUT);                     -> "output"
+//   std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias");         -> "token_embd.bias"
+//   std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3);     -> "blk.3.attn_norm.weight"
+//
+struct LLM_TN_IMPL {
+    const llm_arch arch;
+    const llm_tensor tensor;
+    const char * const suffix;
+    const int bid;
+    const int xid;
+
+    std::string str() const;
+
+    operator std::string() const {
+        return str();
+    }
+
+    friend bool operator==(const std::string & str, const LLM_TN_IMPL & tn) {
+        return str == tn.str();
+    }
+
+    friend bool operator!=(const std::string & str, const LLM_TN_IMPL & tn) {
+        return str != tn.str();
+    }
+};
+
+struct LLM_TN {
+    LLM_TN(llm_arch arch) : arch(arch) {}
+
+    llm_arch arch;
+
+    LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
+        return { arch, tensor, suffix, bid, xid };
+    }
+
+    LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
+        return { arch, tensor, nullptr, bid, xid };
+    }
+};
+
+
+struct llm_tensor_info {
+    llm_tensor_layer layer;
+    ggml_op op;
+};
+
+const char * llm_arch_name(llm_arch arch);
+
+llm_arch llm_arch_from_string(const std::string & name);
+
+const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);

examples/talk-llama/llama-batch.cpp

@@ -0,0 +1,368 @@
+#include "llama-batch.h"
+
+#include <cstring>
+#include <algorithm>
+
+llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
+    // clear empty sequences
+    // the previous ubatch is assumed to be gone,
+    // so nothing should refer to values in these sequences anymore.
+    for (size_t i = seq.size(); i-- > 0;) {
+        if (seq[i].length == 0) {
+            seq.pop_back();
+        } else {
+            break;
+        }
+    }
+    ubatch_token.resize(!has_embd ? n_ubatch : 0);
+    ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0);
+    ubatch_pos.resize(n_ubatch);
+    ubatch_n_seq_id.resize(n_ubatch);
+    ubatch_seq_id.resize(n_ubatch);
+    ubatch_output.resize(n_ubatch);
+    llama_ubatch ubatch = {
+        /*equal_seqs   =*/ true,
+        /*n_tokens     =*/ 0,
+        /*n_seq_tokens =*/ 0,
+        /*n_seqs       =*/ 0,
+        /*token        =*/ !has_embd ? ubatch_token.data() : nullptr,
+        /*embd         =*/ has_embd  ? ubatch_embd.data()  : nullptr,
+        /*pos          =*/ ubatch_pos.data(),
+        /*n_seq_id     =*/ ubatch_n_seq_id.data(),
+        /*seq_id       =*/ ubatch_seq_id.data(),
+        /*output       =*/ ubatch_output.data(),
+    };
+    return ubatch;
+}
+
+void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) {
+    GGML_ASSERT(batch != nullptr);
+    GGML_ASSERT(length <= seq.length);
+    // Can only add sequences of equal lengths to a batch,
+    // otherwise it isn't clear to which sequence a token belongs
+    GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs);
+    GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs);
+    // NOTE: loops are separated for cache-friendliness
+    if (batch->token) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]];
+            }
+        } else {
+            // simple split
+            ubatch.token = batch->token + seq.offset;
+        }
+    } else {
+        ubatch.token = nullptr;
+    }
+    if (batch->embd) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                memcpy(
+                        ubatch.embd + (n_embd * (ubatch.n_tokens + i)),
+                        batch->embd + (n_embd * ids[seq.offset + i]),
+                        n_embd * sizeof(float)
+                      );
+            }
+        } else {
+            // simple split
+            ubatch.embd = batch->embd + (n_embd * seq.offset);
+        }
+    } else {
+        ubatch.embd = nullptr;
+    }
+    if (ubatch.equal_seqs) {
+        for (size_t i = 0; i < length; ++i) {
+            ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
+        }
+    } else {
+        // simple split
+        ubatch.pos = batch->pos + seq.offset;
+    }
+    if (ubatch.equal_seqs) {
+        ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id;
+        if (seq.seq_id) {
+            ubatch.seq_id[ubatch.n_seqs] = seq.seq_id;
+        }
+    } else {
+        // simple split
+        if (batch->n_seq_id) {
+            ubatch.n_seq_id = batch->n_seq_id + seq.offset;
+        } else {
+            for (size_t i = 0; i < length; ++i) {
+                ubatch.n_seq_id[ubatch.n_seqs + i] = 1;
+            }
+        }
+        if (batch->seq_id) {
+            ubatch.seq_id = batch->seq_id + seq.offset;
+        }
+    }
+    if (logits_all) {
+        for (size_t i = 0; i < length; ++i) {
+            ubatch.output[ubatch.n_tokens + i] = 1;
+            out_ids.push_back(ids[seq.offset + i]);
+        }
+    } else if (batch->logits) {
+        if (ubatch.equal_seqs) {
+            for (size_t i = 0; i < length; ++i) {
+                size_t id = ids[seq.offset + i];
+                int8_t is_output = batch->logits[id];
+                ubatch.output[ubatch.n_tokens + i] = is_output;
+                if (is_output) { out_ids.push_back(id); }
+            }
+        } else {
+            // simple split
+            ubatch.output = batch->logits + seq.offset;
+            for (size_t i = 0; i < length; ++i) {
+                if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); }
+            }
+        }
+    } else {
+        // only get last output
+        for (size_t i = 0; i < length; ++i) {
+            size_t id = ids[seq.offset + i];
+            int8_t is_last = id == ids.size() - 1;
+            ubatch.output[ubatch.n_tokens + i] = is_last;
+            if (is_last) { out_ids.push_back(id); }
+        }
+    }
+    if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) {
+        ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1;
+    }
+    ubatch.n_tokens += length;
+    ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits
+    seq.offset += length;
+    seq.length -= length;
+    n_tokens -= length;
+    GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs);
+}
+
+llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    ubatch.equal_seqs = false;
+    if (!seq.empty()) {
+        llama_sbatch_seq & s = seq[0];
+        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
+        GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits
+        add_seq_to_ubatch(ubatch, s, length);
+    }
+    return ubatch;
+}
+
+llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    if (!seq.empty()) {
+        size_t length = 0;
+        size_t n_tokens_in_ubatch = 0;
+        GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits
+                                          // smallest first, because it's easier to split this way;
+                                          // starting from the end to pop in constant time.
+        for (size_t i = seq.size(); i-- > 0;) {
+            llama_sbatch_seq & s = seq[i];
+            GGML_ASSERT(s.length > 0);
+            if (length == 0) {
+                length = s.length < n_ubatch ? s.length : n_ubatch;
+            }
+            add_seq_to_ubatch(ubatch, s, length);
+            n_tokens_in_ubatch += length;
+            // shared prompts can't be mixed with any of their sequences,
+            // so it's safer to compute them in their own ubatch
+            if (s.n_seq_id > 1) { break; }
+            // stop when there isn't enough space for another sequence
+            if (length + n_tokens_in_ubatch > n_ubatch) { break; }
+        }
+    }
+    return ubatch;
+}
+
+llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
+    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
+    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
+    if (!seq.empty()) {
+        llama_sbatch_seq & s = seq[seq.size() - 1];
+        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
+        GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits
+        add_seq_to_ubatch(ubatch, s, length);
+    }
+    return ubatch;
+}
+
+void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+    GGML_ASSERT(batch.n_tokens >= 0);
+    this->batch = &batch;
+    this->n_embd = n_embd;
+    this->logits_all = logits_all;
+
+    n_tokens = batch.n_tokens;
+    ids.resize(n_tokens);
+    out_ids.clear();
+    // TODO: reserve out_ids and seq
+
+    for (size_t i = 0; i < n_tokens; ++i) {
+        ids[i] = i;
+    }
+    if (simple_split) {
+        seq.resize(1);
+        llama_sbatch_seq & s = seq[0];
+        s.n_seq_id = 0;
+        s.seq_id = nullptr;
+        s.offset = 0;
+        s.length = n_tokens;
+        return;
+    }
+    std::sort(ids.begin(), ids.end(),
+            [&batch](size_t a, size_t b) {
+                int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
+                int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1;
+                // sort by seq_id, then by pos
+                if (n_seq_a == n_seq_b) {
+                    if (batch.seq_id) {
+                        for (int32_t i = 0; i < n_seq_a; ++i) {
+                            llama_seq_id seq_id_a = batch.seq_id[a][i];
+                            llama_seq_id seq_id_b = batch.seq_id[b][i];
+                            // smaller seq_ids go first
+                            if (seq_id_a != seq_id_b) {
+                                return seq_id_a < seq_id_b;
+                            }
+                        }
+                    }
+                    // when all else is equal, sort by pos
+                    if (batch.pos) {
+                        return batch.pos[a] < batch.pos[b];
+                    }
+                    // no pos, sort by id
+                    return a < b;
+                }
+                // shared prompts go first
+                return n_seq_a > n_seq_b;
+            }
+    );
+    // init seq
+    llama_sbatch_seq * last_seq = nullptr;
+
+    for (size_t i = 0; i < n_tokens; ++i) {
+        const size_t bi = ids[i];
+        const int32_t n_seqs = batch.n_seq_id[bi];
+        llama_seq_id * seq_ids = batch.seq_id[bi];
+        if (last_seq != nullptr) {
+            bool same = n_seqs == last_seq->n_seq_id;
+            for (int32_t j = 0; same && j < n_seqs; ++j) {
+                if (seq_ids[j] != last_seq->seq_id[j]) {
+                    same = false;
+                }
+            }
+            if (same) {
+                last_seq->length += 1;
+                continue;
+            }
+        }
+        llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1};
+        seq.push_back(new_seq);
+        last_seq = &seq.back();
+    }
+    // keep shared prompts first at the end, then sort by length descending.
+    std::sort(seq.begin(), seq.end(),
+            [](llama_sbatch_seq & a, llama_sbatch_seq & b) {
+                if (a.n_seq_id == b.n_seq_id) {
+                    return a.length > b.length;
+                }
+                return a.n_seq_id < b.n_seq_id;
+            }
+            );
+}
+
+llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
+    batch = in_batch;
+    GGML_ASSERT(batch.n_tokens > 0);
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            pos[i] = i + p0;
+        }
+        batch.pos = pos.data();
+    }
+    if (!batch.n_seq_id) {
+        n_seq_id.resize(batch.n_tokens);
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            n_seq_id[i] = seq_id_0.size();
+        }
+        batch.n_seq_id = n_seq_id.data();
+    }
+    if (!batch.seq_id) {
+        seq_id.resize(batch.n_tokens + 1);
+        seq_id[batch.n_tokens] = NULL;
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            seq_id[i] = seq_id_0.data();
+        }
+        batch.seq_id = seq_id.data();
+    }
+    if (!batch.logits) {
+        logits.resize(batch.n_tokens);
+        logits[logits.size() - 1] = true;
+        batch.logits = logits.data();
+    }
+}
+
+//
+// interface implementation
+//
+
+struct llama_batch llama_batch_get_one(
+             llama_token * tokens,
+                 int32_t   n_tokens) {
+    return {
+        /*n_tokens       =*/ n_tokens,
+        /*tokens         =*/ tokens,
+        /*embd           =*/ nullptr,
+        /*pos            =*/ nullptr,
+        /*n_seq_id       =*/ nullptr,
+        /*seq_id         =*/ nullptr,
+        /*logits         =*/ nullptr,
+    };
+}
+
+struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
+    llama_batch batch = {
+        /*n_tokens       =*/ 0,
+        /*tokens         =*/ nullptr,
+        /*embd           =*/ nullptr,
+        /*pos            =*/ nullptr,
+        /*n_seq_id       =*/ nullptr,
+        /*seq_id         =*/ nullptr,
+        /*logits         =*/ nullptr,
+    };
+
+    if (embd) {
+        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+    } else {
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
+    }
+
+    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens_alloc);
+    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens_alloc);
+    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
+    for (int i = 0; i < n_tokens_alloc; ++i) {
+        batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
+    }
+    batch.seq_id[n_tokens_alloc] = nullptr;
+
+    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens_alloc);
+
+    return batch;
+}
+
+void llama_batch_free(struct llama_batch batch) {
+    if (batch.token)    free(batch.token);
+    if (batch.embd)     free(batch.embd);
+    if (batch.pos)      free(batch.pos);
+    if (batch.n_seq_id) free(batch.n_seq_id);
+    if (batch.seq_id) {
+        for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
+            free(batch.seq_id[i]);
+        }
+        free(batch.seq_id);
+    }
+    if (batch.logits)   free(batch.logits);
+}