whisper : alternative way to handle the external encoders

* CUDA: mul_mat_vec_q tiling, refactor mul mat logic

Co-authored-by: slaren <slarengh@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

.devops/cublas.Dockerfile

@@ -0,0 +1,28 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV WHISPER_CUBLAS=1
+
+RUN make
+
+ENTRYPOINT ["/app/main"]

.devops/main-cuda.Dockerfile

@@ -0,0 +1,38 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=12.3.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+WORKDIR /app
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV WHISPER_CUBLAS=1
+
+RUN apt-get update && \
+    apt-get install -y build-essential \
+    && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
+
+# Ref: https://stackoverflow.com/a/53464012
+ENV CUDA_MAIN_VERSION=12.3
+ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
+
+COPY .. .
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+WORKDIR /app
+
+RUN apt-get update && \
+  apt-get install -y curl ffmpeg \
+  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
+
+COPY --from=build /app /app
+ENTRYPOINT [ "bash", "-c" ]

.devops/main.Dockerfile

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

.github/workflows/build.yml

@@ -1,31 +1,42 @@
 name: CI
 on: [push, pull_request]
 
+env:
+  ubuntu_image: "ubuntu:22.04"
+
 jobs:
   ubuntu-latest:
     runs-on: ubuntu-latest
 
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install libsdl2-dev
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
 
-      - name: Build
+      - name: Build ${{ matrix.arch }}
         run: |
-          make
-          make stream
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential libsdl2-dev
+            make
+            make stream'
 
   macOS-latest:
     runs-on: macOS-latest
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         run: |
@@ -37,82 +48,107 @@ jobs:
           make
           make stream
 
+  freeBSD-latest:
+    runs-on: macos-12
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Build
+        uses: cross-platform-actions/action@v0.15.0
+        with:
+          operating_system: freebsd
+          version: '13.2'
+          run: |
+            sudo pkg update
+            sudo pkg install -y gmake sdl2
+            gmake
+            gmake stream
+
   ubuntu-latest-gcc:
     runs-on: ubuntu-latest
 
     strategy:
+      fail-fast: false
       matrix:
         build: [Debug, Release]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-          sudo apt-get install libsdl2-dev
-
-      - name: Configure
-        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+            make
+            ctest -L gh --output-on-failure'
 
   ubuntu-latest-clang:
     runs-on: ubuntu-latest
 
     strategy:
+      fail-fast: false
       matrix:
         build: [Debug, Release]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-          sudo apt-get install libsdl2-dev
-
-      - name: Configure
-        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y clang build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ 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
 
     strategy:
+      fail-fast: false
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        arch: [linux/amd64]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-
-      - name: Configure
-        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential cmake
+            cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
+            make
+            ctest -L gh --output-on-failure'
 
   windows:
     runs-on: windows-latest
@@ -130,11 +166,11 @@ jobs:
             s2arc: x64
             jnaPath: win32-x86-64
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
@@ -150,7 +186,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
         run: |
@@ -185,17 +221,20 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
+            clblast: OFF
           - arch: x64
-            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
+            clblast: ON
+            clver: 1.6.1
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
@@ -207,7 +246,7 @@ jobs:
           7z x blas.zip -oblas -y
           copy blas/include/cblas.h .
           copy blas/include/openblas_config.h .
-          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+          echo "OPENBLAS_PATH=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -216,13 +255,26 @@ jobs:
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
+      - name: Install OpenCL
+        if: matrix.clblast == 'ON'
+        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
+
+      - name: Fetch CLBlast and set CLBlast_DIR
+        if: matrix.clblast == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
+          7z x clblast.zip
+          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
+          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
+
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_SUPPORT_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
-          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_OPENBLAS=${{ matrix.blas }}
+          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
@@ -231,72 +283,87 @@ jobs:
 
       - name: Copy libopenblas.dll
         if: matrix.blas == 'ON'
-        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+        run: copy "$env:OPENBLAS_PATH/bin/libopenblas.dll" build/bin/${{ matrix.build }}
 
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
+      - name: Copy clblast.dll
+        if: matrix.clblast == 'ON'
+        run: copy "$env:CLBlast_DIR/../../clblast.dll" build/bin/${{ matrix.build }}
+
       - name: Upload binaries
         if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
         with:
-          name: whisper-blas-bin-${{ matrix.arch }}
+          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
-          
+
   windows-cublas:
     runs-on: windows-latest
-    
+
     strategy:
       matrix:
         build: [Release]
         arch: [x64]
         cublas: [ON]
         sdl2: [ON]
+        cuda-toolkit: [12.2.0, 11.8.0]
         include:
           - arch: x64
             s2arc: x64
           - sdl2: ON
-            s2ver: 2.26.0
-            
+            s2ver: 2.28.5
+
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
-        
+
       - name: Install CUDA Toolkit
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.10
-        
+        uses: Jimver/cuda-toolkit@v0.2.11
+        with:
+          cuda: '${{ matrix.cuda-toolkit }}'
+
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
         run: |
           C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
-        
+
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUBLAS=1
-      
-      - name: Build
+          -DWHISPER_CUBLAS=${{ matrix.cublas }}
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+
+      - name: Build ${{ matrix.cuda-toolkit }}
         run: |
           cd ./build
-          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
-          
+          cmake --build . --config ${{ matrix.build }}
+
+      - name: Copy CUDA DLLs
+        run: >
+          Copy-Item -PassThru
+          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
+          -Include cudart64_*,cublas64_*,cublasLt64_*
+          -Destination build/bin/${{ matrix.build }}
+
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
-          
+
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
         with:
-          name: whisper-cublas-bin-${{ matrix.arch }}
+          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   emscripten:
@@ -308,24 +375,16 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
-        run: |
-          wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
-          tar -xvf master.tar.gz
-          emsdk-master/emsdk update
-          emsdk-master/emsdk install latest
-          emsdk-master/emsdk activate latest
+      - name: Setup emsdk
+        uses: mymindstorm/setup-emsdk@v12
 
-      - name: Configure
-        run: echo "tmp"
+      - name: Verify
+        run: emcc -v
 
       - name: Build
         run: |
-          pushd emsdk-master
-          source ./emsdk_env.sh
-          popd
           emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           make
 
@@ -338,7 +397,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Configure
         run: |
@@ -347,7 +406,7 @@ jobs:
 
       - name: Build objc example
         run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
-      
+
       - name: Build swiftui example
         run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
 
@@ -356,27 +415,67 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
+        with:
+          path: whisper
+
+      - name: Clone
+        uses: actions/checkout@v3
+        with:
+          repository: ggerganov/ggml
+          path: ggml
 
       - name: Install Java
         uses: actions/setup-java@v3
         with:
           distribution: zulu
           java-version: 17
-      
+
       - name: Setup Android SDK
         uses: android-actions/setup-android@v2
 
       - name: Build
         run: |
-          cd examples/whisper.android
+          cd whisper/examples/whisper.android
           ./gradlew assembleRelease --no-daemon
 
+      - name: Build with external ggml
+        run: |
+          export PATH_TO_GGML=$PWD/ggml
+          cd whisper/examples/whisper.android
+          ./gradlew assembleRelease --no-daemon -PGGML_HOME=$PATH_TO_GGML
+
+  android_java:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: set up JDK 11
+        uses: actions/setup-java@v3
+        with:
+          java-version: '11'
+          distribution: 'temurin'
+          cache: gradle
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v2
+        with:
+          api-level: 30
+          build-tools-version: 30.0.3
+
+      - name: Build
+        run: |
+          cd examples/whisper.android.java
+          chmod +x ./gradlew 
+          ./gradlew assembleRelease
+
   java:
     needs: [ 'windows' ]
     runs-on: windows-latest
     steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v3
 
       - name: Install Java
         uses: actions/setup-java@v1
@@ -402,11 +501,27 @@ jobs:
           name: whispercpp.jar
           path: bindings/java/build/libs/whispercpp-*.jar
 
-#      - name: Publish package
-#        if: ${{ github.ref == 'refs/heads/master' }}
-#        uses: gradle/gradle-build-action@v2
-#        with:
-#          arguments: publish
-#        env:
-#          MAVEN_USERNAME: ${{ secrets.OSSRH_USERNAME }}
-#          MAVEN_PASSWORD: ${{ secrets.OSSRH_TOKEN }}
+      - name: Publish package
+        if: ${{ github.ref == 'refs/heads/master' }}
+        uses: gradle/gradle-build-action@v2.4.2
+        with:
+          arguments: publish
+          build-root-directory: bindings/java
+        env:
+          MAVEN_USERNAME: ${{ secrets.JIRA_USER }}
+          MAVEN_PASSWORD: ${{ secrets.JIRA_PASS }}
+          PGP_SECRET: ${{ secrets.GPG_PRIVATE_KEY }}
+          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+
+  quantize:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Test quantize
+        run: |
+          ./models/download-ggml-model.sh tiny.en
+          make quantize
+          ./quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0

.github/workflows/docker.yml

@@ -0,0 +1,57 @@
+name: Publish Docker image
+
+on:
+  pull_request:
+  push:
+    branches:
+      - master
+
+jobs:
+  push_to_registry:
+    name: Push Docker image to Docker Hub
+    if: github.event.pull_request.draft == false
+
+    runs-on: ubuntu-latest
+    env:
+      COMMIT_SHA: ${{ github.sha }}
+    strategy:
+      matrix:
+        config:
+          - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64,linux/arm64" }
+          - { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
+
+    steps:
+      - name: Check out the repo
+        uses: actions/checkout@v3
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Build and push Docker image (versioned)
+        if: github.event_name == 'push'
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          push: true
+          platforms: ${{ matrix.config.platforms }}
+          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          file: ${{ matrix.config.dockerfile }}
+
+      - name: Build and push Docker image (tagged)
+        uses: docker/build-push-action@v4
+        with:
+          context: .
+          push: ${{ github.event_name == 'push' }}
+          platforms: ${{ matrix.config.platforms }}
+          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
+          file: ${{ matrix.config.dockerfile }}

.gitignore

@@ -8,15 +8,22 @@
 .DS_Store
 
 build/
+build-coreml/
 build-em/
 build-debug/
 build-release/
+build-rwdi/
 build-static/
 build-cublas/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
+# SPM
+.build/
+.swiftpm
+*.metallib
+
 /main
 /stream
 /command
@@ -24,6 +31,8 @@ build-sanitize-thread/
 /talk-llama
 /bench
 /quantize
+/server
+/lsp
 
 arm_neon.h
 sync.sh
@@ -44,3 +53,9 @@ models/*.mlpackage
 bindings/java/.gradle/
 bindings/java/.idea/
 .idea/
+
+benchmark_results.csv
+cmake-build-debug/
+.cxx/
+.gradle/
+local.properties

CMakeLists.txt

@@ -1,6 +1,6 @@
-cmake_minimum_required (VERSION 3.0)
+cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.4.2)
+project(whisper.cpp VERSION 1.5.4)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -35,6 +35,12 @@ endif()
 
 # options
 
+if (APPLE)
+    set(WHISPER_METAL_DEFAULT ON)
+else()
+    set(WHISPER_METAL_DEFAULT OFF)
+endif()
+
 option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
 
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
@@ -58,13 +64,17 @@ option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
 
 if (APPLE)
     option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
+    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
+    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
     option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
     option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
+    option(WHISPER_METAL_EMBED_LIBRARY   "whisper: embed Metal library"          OFF)
 else()
     option(WHISPER_BLAS                  "whisper: use BLAS libraries"  OFF)
     option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor" Generic)
     option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"     OFF)
     option(WHISPER_CUBLAS                "whisper: support for cuBLAS"  OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS" OFF)
     option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
 endif()
 
@@ -108,7 +118,59 @@ if (APPLE)
             set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
         else()
-            message(WARNING "Accelerate framework not found")
+            message(FATAL_ERROR "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (WHISPER_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+        if (METAL_FRAMEWORK)
+            message(STATUS "Metal framework found")
+
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
+                ${FOUNDATION_LIBRARY}
+                ${METAL_FRAMEWORK}
+                ${METALKIT_FRAMEWORK}
+                )
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
+
+            if (WHISPER_METAL_NDEBUG)
+                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
+            endif()
+        else()
+            message(FATAL_ERROR "Metal framework not found")
+        endif()
+
+        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+        # copy ggml-metal.metal to bin directory
+        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+        if (WHISPER_METAL_EMBED_LIBRARY)
+            enable_language(ASM)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)
+
+            set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
+
+            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+            add_custom_command(
+                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
+                DEPENDS ${METALLIB_SOURCE}
+                COMMENT "Generate assembly for embedded Metal library"
+            )
+
+            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
         endif()
     endif()
 
@@ -121,7 +183,7 @@ if (APPLE)
 
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
         else()
-            message(WARNING "CoreML framework not found")
+            message(FATAL_ERROR "CoreML framework not found")
         endif()
 
         if (WHISPER_COREML_ALLOW_FALLBACK)
@@ -136,22 +198,34 @@ if (WHISPER_OPENBLAS)
 endif()
 
 if (WHISPER_BLAS)
-    set(BLA_STATIC 1)
-    set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-#    set(BLA_PREFER_PKGCONFIG 1)
-    set(BLA_SIZEOF_INTEGER 8)
-    find_package(BLAS)
+    if (WIN32)
+        if(DEFINED ENV{OPENBLAS_PATH})
+            set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories($ENV{OPENBLAS_PATH}/include)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else ()
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
+        endif ()
+    else ()
+        set(BLA_STATIC 1)
+        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
+        set(BLA_SIZEOF_INTEGER 8)
+        set(BLA_PREFER_PKGCONFIG 1)
+        find_package(BLAS)
 
-    if(BLAS_FOUND)
-        message(STATUS "BLAS compatible library found")
-        message(STATUS "Libraries ${BLAS_LIBRARIES}")
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-
-        include_directories(${BLAS_INCLUDE_DIRS})
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
-    else()
-        message(WARNING "BLAS library was not found")
-    endif()
+        if(BLAS_FOUND)
+            message(STATUS "BLAS compatible library found")
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories(${BLAS_INCLUDE_DIRS})
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else()
+            message(FATAL_ERROR "BLAS library was not found")
+        endif()
+    endif ()
 endif ()
 
 if (WHISPER_CUBLAS)
@@ -164,18 +238,55 @@ if (WHISPER_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
 
         if (WHISPER_STATIC)
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
         else()
             set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
     else()
-        message(WARNING "cuBLAS not found")
+        message(FATAL_ERROR "cuBLAS not found")
+    endif()
+endif()
+
+
+if (WHISPER_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (WHISPER_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
     endif()
 endif()
 
@@ -184,13 +295,13 @@ if (WHISPER_CLBLAST)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
         set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
     else()
-        message(WARNING "CLBlast not found")
+        message(FATAL_ERROR "CLBlast not found")
     endif()
 endif()
 
@@ -229,7 +340,8 @@ if (WHISPER_ALL_WARNINGS)
 endif()
 
 if (NOT MSVC)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
+    # TODO: temporary disabled until we figure out ggml-metal.m
+    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
     #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
 endif()
 
@@ -237,24 +349,29 @@ message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 
 if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
     message(STATUS "ARM detected")
+elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+    message(STATUS "PowerPC detected")
 else()
     message(STATUS "x86 detected")
     if (MSVC)
-            if(NOT WHISPER_NO_AVX2)
-                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
-                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-            else()
-                if(NOT WHISPER_NO_AVX)
-                    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
-                    set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
-                endif()
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
+        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
+        if(NOT WHISPER_NO_AVX2)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
+            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
+        else()
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
             endif()
+        endif()
     else()
         if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
         else()
             if(NOT WHISPER_NO_AVX)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
@@ -272,6 +389,53 @@ else()
     endif()
 endif()
 
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
 if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
@@ -330,8 +494,15 @@ set(TARGET whisper)
 add_library(${TARGET}
     ggml.h
     ggml.c
-    ${GGML_CUDA_SOURCES}
-    ${GGML_OPENCL_SOURCES}
+    ggml-alloc.h
+    ggml-alloc.c
+    ggml-backend.h
+    ggml-backend.c
+    ggml-quants.h
+    ggml-quants.c
+    ${GGML_SOURCES_METAL}
+    ${GGML_SOURCES_CUDA}
+    ${GGML_SOURCES_OPENCL}
     whisper.h
     whisper.cpp
     )
@@ -359,6 +530,7 @@ else()
 endif()
 
 if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_link_libraries(${TARGET} PUBLIC
         ${CMAKE_DL_LIBS}
         )
@@ -372,11 +544,23 @@ if (BUILD_SHARED_LIBS)
         WHISPER_BUILD
         GGML_BUILD
         )
+
+    if (WHISPER_METAL)
+        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
+        #       but for some reason it does not work here like it does in llama.cpp
+        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
 endif()
 
-if (GGML_CUDA_SOURCES)
+if (GGML_SOURCES_CUDA)
     message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES OFF)
+    # Only configure gmml CUDA architectures is not globally set
+    if (NOT DEFINED GGML_CUDA_ARCHITECTURES)
+        # Not overriden by user, so set defaults
+        set(GGML_CUDA_ARCHITECTURES 52 61 70)
+    endif()
+    message(STATUS "GGML Configuring CUDA architectures ${GGML_CUDA_ARCHITECTURES}")
+    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES ${GGML_CUDA_ARCHITECTURES})
     set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
 endif()
 
@@ -388,12 +572,15 @@ target_compile_definitions(${TARGET} PUBLIC
     ${WHISPER_EXTRA_FLAGS}
     )
 
-set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
+
+include(GNUInstallDirs)
 
 install(TARGETS ${TARGET}
-    LIBRARY DESTINATION lib
-    ARCHIVE DESTINATION lib/static
-    RUNTIME DESTINATION bin
+    LIBRARY  DESTINATION lib
+    ARCHIVE  DESTINATION lib/static
+    RUNTIME  DESTINATION bin
+    RESOURCE DESTINATION bin
     PUBLIC_HEADER DESTINATION include
     )
 

Makefile

@@ -1,4 +1,4 @@
-default: main bench quantize
+default: main bench quantize server
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -12,7 +12,13 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
+ifndef NVCC_VERSION
+	ifeq ($(call,$(shell which nvcc))$(.SHELLSTATUS),0)
+		NVCC_VERSION := $(shell nvcc --version | egrep -o "V[0-9]+.[0-9]+.[0-9]+" | cut -c2-)
+	endif
+endif
+
+CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)
 
 # Mac OS + Arm can report x86_64
@@ -36,101 +42,132 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
-# ref: https://github.com/ggerganov/whisper.cpp/issues/37
-ifneq ($(wildcard /usr/include/musl/*),)
-	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
-	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+ifdef MACOSX_DEPLOYMENT_TARGET
+	CFLAGS   += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+	CXXFLAGS += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+	LDFLAGS  += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+endif
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+CFLAGS   += -D_XOPEN_SOURCE=600
+CXXFLAGS += -D_XOPEN_SOURCE=600
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+ifeq ($(UNAME_S),OpenBSD)
+	CFLAGS   += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+	CXXFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+endif
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+ifeq ($(UNAME_S),Linux)
+	CFLAGS   += -D_GNU_SOURCE
+	CXXFLAGS += -D_GNU_SOURCE
 endif
 
 # RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
 # and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
 ifeq ($(UNAME_S),Darwin)
 	CFLAGS   += -D_DARWIN_C_SOURCE
 	CXXFLAGS += -D_DARWIN_C_SOURCE
 endif
+ifeq ($(UNAME_S),DragonFly)
+	CFLAGS   += -D__BSD_VISIBLE
+	CXXFLAGS += -D__BSD_VISIBLE
+endif
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+ifeq ($(UNAME_S),FreeBSD)
+	CFLAGS   += -D__BSD_VISIBLE
+	CXXFLAGS += -D__BSD_VISIBLE
+endif
+ifeq ($(UNAME_S),NetBSD)
+	CFLAGS   += -D_NETBSD_SOURCE
+	CXXFLAGS += -D_NETBSD_SOURCE
+endif
+ifeq ($(UNAME_S),OpenBSD)
+	CFLAGS   += -D_BSD_SOURCE
+	CXXFLAGS += -D_BSD_SOURCE
+endif
 
 # OS specific
 # TODO: support Windows
-ifeq ($(UNAME_S),Linux)
+ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$(UNAME_S))
 	CFLAGS   += -pthread
 	CXXFLAGS += -pthread
 endif
-ifeq ($(UNAME_S),Darwin)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
+
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
 endif
-ifeq ($(UNAME_S),FreeBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Haiku)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
 endif
 
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	ifeq ($(UNAME_S),Darwin)
-		CFLAGS += -mf16c
-		AVX1_M := $(shell sysctl machdep.cpu.features)
-		ifneq (,$(findstring FMA,$(AVX1_M)))
-			CFLAGS += -mfma
-		endif
-		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysctl machdep.cpu.leaf7_features)
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
+		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
 	else ifeq ($(UNAME_S),Linux)
-		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell grep "fma " /proc/cpuinfo)
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-
-			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
-		endif
-		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
-		ifneq (,$(findstring sse3,$(SSE3_M)))
-			CFLAGS += -msse3
-		endif
+		CPUINFO_CMD := cat /proc/cpuinfo
+	else ifneq (,$(filter MINGW32_NT% MINGW64_NT% MSYS_NT%,$(UNAME_S)))
+		CPUINFO_CMD := cat /proc/cpuinfo
+	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
+		CPUINFO_CMD := grep Features /var/run/dmesg.boot
 	else ifeq ($(UNAME_S),Haiku)
-		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell sysinfo -cpu | grep "FMA ")
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-
-			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
-		endif
-	else
-		CFLAGS += -mfma -mf16c -mavx -mavx2
+		CPUINFO_CMD := sysinfo -cpu
+	endif
+
+	ifdef CPUINFO_CMD
+		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
+		ifneq (,$(AVX_M))
+			CFLAGS   += -mavx
+			CXXFLAGS += -mavx
+		endif
+
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
+		ifneq (,$(AVX2_M))
+			CFLAGS   += -mavx2
+			CXXFLAGS += -mavx2
+		endif
+
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
+		endif
+
+		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
+		ifneq (,$(F16C_M))
+			CFLAGS   += -mf16c
+			CXXFLAGS += -mf16c
+		endif
+
+		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
+		ifneq (,$(SSE3_M))
+			CFLAGS   += -msse3
+			CXXFLAGS += -msse3
+		endif
+
+		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
+		ifneq (,$(SSSE3_M))
+			CFLAGS   += -mssse3
+			CXXFLAGS += -mssse3
+		endif
 	endif
-endif
-ifeq ($(UNAME_M),amd64)
-	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
 
 ifneq ($(filter ppc64%,$(UNAME_M)),)
@@ -161,30 +198,67 @@ ifdef WHISPER_COREML_ALLOW_FALLBACK
 endif
 endif
 
+ifndef WHISPER_NO_METAL
+	ifeq ($(UNAME_S),Darwin)
+		WHISPER_METAL := 1
+
+		CFLAGS   += -DGGML_USE_METAL
+		CXXFLAGS += -DGGML_USE_METAL
+		LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	endif
+endif
+
 ifdef WHISPER_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas
 endif
 
 ifdef WHISPER_CUBLAS
+	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
+		CUDA_ARCH_FLAG ?= native
+	else
+		CUDA_ARCH_FLAG ?= all
+	endif
+
 	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
+	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
 	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
-	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=any
+	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 
+ifdef WHISPER_HIPBLAS
+	ROCM_PATH   ?= /opt/rocm
+	HIPCC       ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	LDFLAGS     += -lhipblas -lamdhip64 -lrocblas
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	WHISPER_OBJ += ggml-cuda.o
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif
+
 ifdef WHISPER_CLBLAST
 	CFLAGS 		+= -DGGML_USE_CLBLAST
-	LDFLAGS	 	+= -lclblast -lOpenCL
+	CXXFLAGS 	+= -DGGML_USE_CLBLAST
+	LDFLAGS	 	+= -lclblast
+	ifeq ($(UNAME_S),Darwin)
+		LDFLAGS	 	+= -framework OpenCL
+	else
+		LDFLAGS	    += -lOpenCL
+	endif
 	WHISPER_OBJ	+= ggml-opencl.o
 
 ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
-	$(CC) $(CFLAGS) -c $< -o $@
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif
 
 ifdef WHISPER_GPROF
@@ -240,6 +314,17 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+WHISPER_OBJ += ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o
+
 whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
@@ -255,14 +340,39 @@ whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-imp
 WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
 endif
 
-libwhisper.a: ggml.o $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+ifdef WHISPER_METAL
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
 
-libwhisper.so: ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
+WHISPER_OBJ += ggml-metal.o
+
+ifdef WHISPER_METAL_EMBED_LIBRARY
+CFLAGS += -DGGML_METAL_EMBED_LIBRARY
+
+ggml-metal-embed.o: ggml-metal.metal
+	@echo "Embedding Metal library"
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
+
+WHISPER_OBJ += ggml-metal-embed.o
+endif
+endif
+
+libwhisper.a: $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a $(WHISPER_OBJ)
+
+libwhisper.so: $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk talk-llama bench quantize libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize server lsp libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -273,27 +383,33 @@ CC_SDL=`sdl2-config --cflags --libs`
 SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
+bench: examples/bench/bench.cpp $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
-	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
+server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS) $(LWINSOCK2)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
+
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples
@@ -309,6 +425,7 @@ samples:
 	@wget --quiet --show-progress -O samples/hp0.ogg https://upload.wikimedia.org/wikipedia/en/d/d4/En.henryfphillips.ogg
 	@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
@@ -318,6 +435,8 @@ samples:
 	@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
@@ -335,9 +454,10 @@ samples:
 .PHONY: medium.en
 .PHONY: medium
 .PHONY: large-v1
-.PHONY: large
+.PHONY: large-v2
+.PHONY: large-v3
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="
@@ -359,4 +479,4 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
 
 .PHONY: tests
 tests:
-	bash ./tests/run-tests.sh
+	bash ./tests/run-tests.sh $(word 2, $(MAKECMDGOALS))

Package.swift

@@ -0,0 +1,59 @@
+// 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"]),
+    ],
+    dependencies: [
+        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
+    ],
+    targets: [
+        .target(
+            name: "whisper",
+            dependencies: ["ggml"],
+            path: ".",
+            exclude: [
+               "bindings",
+               "cmake",
+               "coreml",
+               "examples",
+               "extra",
+               "models",
+               "samples",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
+            sources: [
+                "whisper.cpp",
+            ],
+            publicHeadersPath: "spm-headers",
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL")
+                // NOTE: NEW_LAPACK will required iOS version 16.4+
+                // We should consider add this in the future when we drop support for iOS 14
+                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+                // .define("ACCELERATE_NEW_LAPACK"),
+                // .define("ACCELERATE_LAPACK_ILP64")
+            ],
+            linkerSettings: [
+                .linkedFramework("Accelerate")
+            ]
+        )
+    ],
+    cxxLanguageStandard: .cxx11
+)

README.md

@@ -6,22 +6,21 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Beta: [v1.4.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.2) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [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 and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
-- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
-- Runs on the CPU
-- [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- Support for CPU-only inference
+- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
 - [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
-- [BLAS CPU support via OpenBLAS](https://github.com/ggerganov/whisper.cpp#blas-cpu-support-via-openblas)
+- [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
 Supported platforms:
@@ -34,11 +33,10 @@ Supported platforms:
 - [x] [WebAssembly](examples/whisper.wasm)
 - [x] Windows ([MSVC](https://github.com/ggerganov/whisper.cpp/blob/master/.github/workflows/build.yml#L117-L144) and [MinGW](https://github.com/ggerganov/whisper.cpp/issues/168)]
 - [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/discussions/166)
+- [x] [docker](https://github.com/ggerganov/whisper.cpp/pkgs/container/whisper.cpp)
 
-The entire implementation of the model is contained in 2 source files:
-
-- Tensor operations: [ggml.h](ggml.h) / [ggml.c](ggml.c)
-- Transformer inference: [whisper.h](whisper.h) / [whisper.cpp](whisper.cpp)
+The entire high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
+The rest of the code is part of the [`ggml`](https://github.com/ggerganov/ggml) machine learning library.
 
 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
 As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device: [whisper.objc](examples/whisper.objc)
@@ -49,6 +47,10 @@ You can also easily make your own offline voice assistant application: [command]
 
 https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
 
+On Apple Silicon, the inference runs fully on the GPU via Metal:
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/c82e8f86-60dc-49f2-b048-d2fdbd6b5225
+
 Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 
 ## Implementation details
@@ -59,22 +61,22 @@ Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 - Sample real-time audio transcription from the microphone is demonstrated in [stream.cpp](examples/stream)
 - Various other examples are available in the [examples](examples) folder
 
-The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD
-instrisics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
-the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
+The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
 
 ## Quick start
 
-First clone the repository.
+First clone the repository:
 
-Then, download one of the Whisper models converted in [ggml format](models). For example:
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+```
+
+Then, download one of the Whisper [models](models/README.md) converted in [`ggml` format](#ggml-format). For example:
 
 ```bash
 bash ./models/download-ggml-model.sh base.en
 ```
 
-If you wish to convert the Whisper models to ggml format yourself, instructions are in [models/README.md](models/README.md).
-
 Now build the [main](examples/main) example and transcribe an audio file like this:
 
 ```bash
@@ -89,7 +91,7 @@ make
 
 For a quick demo, simply run `make base.en`:
 
-```java
+```text
 $ make base.en
 
 cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
@@ -108,30 +110,39 @@ options:
   -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
   -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
   -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -sow,      --split-on-word     [false  ] split on word rather than on token
   -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -bs N,     --beam-size N       [5      ] beam size for beam search
   -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
   -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
   -tr,       --translate         [false  ] translate from source language to english
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
   -otxt,     --output-txt        [false  ] output result in a text file
   -ovtt,     --output-vtt        [false  ] output result in a vtt file
   -osrt,     --output-srt        [false  ] output result in a srt file
+  -olrc,     --output-lrc        [false  ] output result in a lrc file
   -owts,     --output-words      [false  ] output script for generating karaoke video
+  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
+  -ojf,      --output-json-full  [false  ] include more information in the JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
   -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -nt,       --no-timestamps     [false  ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
              --prompt PROMPT     [       ] initial prompt
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  -ls,       --log-score         [false  ] log best decoder scores of tokens
+  -ng,       --no-gpu            [false  ] disable GPU
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -196,7 +207,7 @@ For detailed usage instructions, run: `./main -h`
 Note that the [main](examples/main) example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.
 For example, you can use `ffmpeg` like this:
 
-```java
+```bash
 ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
 ```
 
@@ -222,18 +233,19 @@ make small
 make medium.en
 make medium
 make large-v1
-make large
+make large-v2
+make large-v3
 ```
 
 ## Memory usage
 
-| Model  | Disk   | Mem     | SHA                                        |
-| ---    | ---    | ---     | ---                                        |
-| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| Model  | Disk    | Mem     |
+| ------ | ------- | ------- |
+| tiny   | 75 MiB  | ~273 MB |
+| base   | 142 MiB | ~388 MB |
+| small  | 466 MiB | ~852 MB |
+| medium | 1.5 GiB | ~2.1 GB |
+| large  | 2.9 GiB | ~3.9 GB |
 
 ## Quantization
 
@@ -266,7 +278,7 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
   - To ensure `coremltools` operates correctly, please confirm that [Xcode](https://developer.apple.com/xcode/) is installed and execute `xcode-select --install` to install the command-line tools.
   - Python 3.10 is recommended.
-  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html)  for this step:
+  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for this step:
     - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
     - To activate the environment, use: `conda activate py310-whisper`
 
@@ -286,14 +298,14 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
   WHISPER_COREML=1 make -j
 
   # using CMake
-  cd build
-  cmake -DWHISPER_COREML=1 ..
+  cmake -B build -DWHISPER_COREML=1
+  cmake --build build -j --config Release
   ```
 
 - Run the examples as usual. For example:
 
-  ```bash
-  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```text
+  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -311,9 +323,94 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
 For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
 
-## NVIDIA GPU support via cuBLAS
+## OpenVINO support
 
-With NVIDIA cards the Encoder processing can to a large extent be offloaded to the GPU through cuBLAS.
+On platforms that support [OpenVINO](https://github.com/openvinotoolkit/openvino), the Encoder inference can be executed
+on OpenVINO-supported devices including x86 CPUs and Intel GPUs (integrated & discrete).
+
+This can result in significant speedup in encoder performance. Here are the instructions for generating the OpenVINO model and using it with `whisper.cpp`:
+
+- First, setup python virtual env. and install python dependencies. Python 3.10 is recommended.
+
+  Windows:
+
+  ```powershell
+  cd models
+  python -m venv openvino_conv_env
+  openvino_conv_env\Scripts\activate
+  python -m pip install --upgrade pip
+  pip install -r openvino-conversion-requirements.txt
+  ```
+
+  Linux and macOS:
+
+  ```bash
+  cd models
+  python3 -m venv openvino_conv_env
+  source openvino_conv_env/bin/activate
+  python -m pip install --upgrade pip
+  pip install -r openvino-conversion-requirements.txt
+  ```
+
+- Generate an OpenVINO encoder model. For example, to generate a `base.en` model, use:
+
+  ```
+  python convert-whisper-to-openvino.py --model base.en
+  ```
+
+  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as `ggml` models, as that
+  is the default location that the OpenVINO extension will search at runtime.
+
+- Build `whisper.cpp` with OpenVINO support:
+
+  Download OpenVINO package from [release page](https://github.com/openvinotoolkit/openvino/releases). The recommended version to use is [2023.0.0](https://github.com/openvinotoolkit/openvino/releases/tag/2023.0.0).
+
+  After downloading & extracting package onto your development system, set up required environment by sourcing setupvars script. For example:
+
+  Linux:
+
+  ```bash
+  source /path/to/l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64/setupvars.sh
+  ```
+
+  Windows (cmd):
+
+  ```powershell
+  C:\Path\To\w_openvino_toolkit_windows_2023.0.0.10926.b4452d56304_x86_64\setupvars.bat
+  ```
+
+  And then build the project using cmake:
+
+  ```bash
+  cmake -B build -DWHISPER_OPENVINO=1
+  cmake --build build -j --config Release
+  ```
+
+- Run the examples as usual. For example:
+
+  ```text
+  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+
+  ...
+
+  whisper_ctx_init_openvino_encoder: loading OpenVINO model from 'models/ggml-base.en-encoder-openvino.xml'
+  whisper_ctx_init_openvino_encoder: first run on a device may take a while ...
+  whisper_openvino_init: path_model = models/ggml-base.en-encoder-openvino.xml, device = GPU, cache_dir = models/ggml-base.en-encoder-openvino-cache
+  whisper_ctx_init_openvino_encoder: OpenVINO model loaded
+
+  system_info: n_threads = 4 / 8 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | COREML = 0 | OPENVINO = 1 |
+
+  ...
+  ```
+
+  The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
+  cached for the next run.
+
+For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
+
+## NVIDIA GPU support
+
+With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads
 
 Now build `whisper.cpp` with cuBLAS support:
@@ -338,14 +435,11 @@ make clean
 WHISPER_CLBLAST=1 make -j
 
 CMake:
-cd whisper.cpp ; mkdir build ; cd build
-cmake -DWHISPER_CLBLAST=ON  ..
-make clean
-make -j
-cp bin/* ../ 
+cd whisper.cpp
+cmake -B build -DWHISPER_CLBLAST=ON
+cmake --build build -j --config Release
 ```
 
-
 Run all the examples as usual.
 
 ## BLAS CPU support via OpenBLAS
@@ -360,6 +454,38 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```
 
+## Docker
+
+### Prerequisites
+
+- Docker must be installed and running on your system.
+- Create a folder to store big models & intermediate files (ex. /whisper/models)
+
+### Images
+
+We have two Docker images available for this project:
+
+1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
+
+### Usage
+
+```shell
+# download model and persist it in a local folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./models/download-ggml-model.sh base /models"
+# transcribe an audio file
+docker run -it --rm \
+  -v path/to/models:/models \
+  -v path/to/audios:/audios \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
+# transcribe an audio file in samples folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
+```
+
 ## Limitations
 
 - Inference only
@@ -372,7 +498,7 @@ in about half a minute on a MacBook M1 Pro, using `medium.en` model:
 <details>
   <summary>Expand to see the result</summary>
 
-```java
+```text
 $ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8
 
 whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
@@ -444,6 +570,7 @@ whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per
 whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
 whisper_print_timings:    total time = 32733.52 ms
 ```
+
 </details>
 
 ## Real-time audio input example
@@ -452,7 +579,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
-```java
+```bash
 make stream
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
@@ -464,7 +591,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:
 
-```java
+```bash
 ./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
 ```
 
@@ -474,8 +601,8 @@ to highlight words with high or low confidence:
 
 For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:
 
-```java
-./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
+```text
+$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -498,8 +625,8 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr
 
 The `--max-len` argument can be used to obtain word-level timestamps. Simply use `-ml 1`:
 
-```java
-./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
+```text
+$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -569,7 +696,7 @@ This requires to have `ffmpeg` installed.
 
 Here are a few *"typical"* examples:
 
-```java
+```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
 source ./samples/jfk.wav.wts
 ffplay ./samples/jfk.wav.mp4
@@ -579,7 +706,7 @@ https://user-images.githubusercontent.com/1991296/199337465-dbee4b5e-9aeb-48a3-b
 
 ---
 
-```java
+```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owts
 source ./samples/mm0.wav.wts
 ffplay ./samples/mm0.wav.mp4
@@ -589,7 +716,7 @@ https://user-images.githubusercontent.com/1991296/199337504-cc8fd233-0cb7-4920-9
 
 ---
 
-```java
+```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owts
 source ./samples/gb0.wav.wts
 ffplay ./samples/gb0.wav.mp4
@@ -603,7 +730,7 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
 
-```java
+```bash
 ./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```
@@ -620,7 +747,19 @@ took to execute it. The results are summarized in the following Github issue:
 
 [Benchmark results](https://github.com/ggerganov/whisper.cpp/issues/89)
 
-## ggml format
+Additionally a script to run whisper.cpp with different models and audio files is provided [bench.py](bench.py).
+
+You can run it with the following command, by default it will run against any standard model in the models folder.
+
+```bash
+python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+```
+
+It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
+
+It outputs a csv file with the results of the benchmarking.
+
+## `ggml` format
 
 The original models are converted to a custom binary format. This allows to pack everything needed into a single file:
 
@@ -635,49 +774,50 @@ or manually from here:
 - https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
-For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
-in [models](models).
+For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or [models/README.md](models/README.md).
 
 ## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
-- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
-- [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [x] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
+- [x] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
   - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
-- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
-- [X] Java:
+- [x] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [x] Java:
   - [GiviMAD/whisper-jni](https://github.com/GiviMAD/whisper-jni)
-- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
-- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+- [x] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
+- [x] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
   - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
-- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+- [x] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
   - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
-- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
   - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
   - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
-- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
-- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
+- [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
 There are various examples of using the library for different projects in the [examples](examples) folder.
 Some of the examples are even ported to run in the browser using WebAssembly. Check them out!
 
-| Example | Web | Description |
-| ---     | --- | ---         |
-| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
-| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
-| [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
-| [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
-| [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
-| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
-| [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
-| [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
-| [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
-| [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
-| [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
-| [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
-| [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
+| Example                                             | Web                                   | Description                                                                                                                     |
+| --------------------------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
+| [main](examples/main)                               | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
+| [bench](examples/bench)                             | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
+| [stream](examples/stream)                           | [stream.wasm](examples/stream.wasm)   | Real-time transcription of raw microphone capture                                                                               |
+| [command](examples/command)                         | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic                                                         |
+| [wchess](examples/wchess)                           | [wchess.wasm](examples/wchess)        | Voice-controlled chess                                                                                                          |
+| [talk](examples/talk)                               | [talk.wasm](examples/talk.wasm)       | Talk with a GPT-2 bot                                                                                                           |
+| [talk-llama](examples/talk-llama)                   |                                       | Talk with a LLaMA bot                                                                                                           |
+| [whisper.objc](examples/whisper.objc)               |                                       | iOS mobile application using whisper.cpp                                                                                        |
+| [whisper.swiftui](examples/whisper.swiftui)         |                                       | SwiftUI iOS / macOS application using whisper.cpp                                                                               |
+| [whisper.android](examples/whisper.android)         |                                       | Android mobile application using whisper.cpp                                                                                    |
+| [whisper.nvim](examples/whisper.nvim)               |                                       | Speech-to-text plugin for Neovim                                                                                                |
+| [generate-karaoke.sh](examples/generate-karaoke.sh) |                                       | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture                           |
+| [livestream.sh](examples/livestream.sh)             |                                       | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185)                                           |
+| [yt-wsp.sh](examples/yt-wsp.sh)                     |                                       | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
+| [server](examples/server)                           |                                       | HTTP transcription server with OAI-like API                                                                                     |
 
 ## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)
 

bindings/go/Makefile

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

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

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

bindings/go/params.go

@@ -19,6 +19,10 @@ func (p *Params) SetTranslate(v bool) {
 	p.translate = toBool(v)
 }
 
+func (p *Params) SetSplitOnWord(v bool) {
+	p.split_on_word = toBool(v)
+}
+
 func (p *Params) SetNoContext(v bool) {
 	p.no_context = toBool(v)
 }
@@ -114,6 +118,16 @@ func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)
 }
 
+// Set audio encoder context
+func (p *Params) SetAudioCtx(n int) {
+	p.audio_ctx = C.int(n)
+}
+
+// Set initial prompt
+func (p *Params) SetInitialPrompt(prompt string) {
+	p.initial_prompt = C.CString(prompt)
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS
 
@@ -137,6 +151,8 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" n_max_text_ctx=%d", p.n_max_text_ctx)
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
+	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
+	str += fmt.Sprintf(" initial_prompt=%s", C.GoString(p.initial_prompt))
 	if p.translate {
 		str += " translate"
 	}

bindings/go/pkg/whisper/context.go

@@ -81,6 +81,10 @@ func (context *context) SetSpeedup(v bool) {
 	context.params.SetSpeedup(v)
 }
 
+func (context *context) SetSplitOnWord(v bool) {
+	context.params.SetSplitOnWord(v)
+}
+
 // Set number of threads to use
 func (context *context) SetThreads(v uint) {
 	context.params.SetThreads(int(v))
@@ -121,6 +125,16 @@ func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
 }
 
+// Set audio encoder context
+func (context *context) SetAudioCtx(n uint) {
+	context.params.SetAudioCtx(int(n))
+}
+
+// Set initial prompt
+func (context *context) SetInitialPrompt(prompt string) {
+	context.params.SetInitialPrompt(prompt)
+}
+
 // ResetTimings resets the mode timings. Should be called before processing
 func (context *context) ResetTimings() {
 	context.model.ctx.Whisper_reset_timings()

bindings/go/pkg/whisper/interface.go

@@ -38,15 +38,18 @@ type Context interface {
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language
 
-	SetOffset(time.Duration)      // Set offset
-	SetDuration(time.Duration)    // Set duration
-	SetThreads(uint)              // Set number of threads to use
-	SetSpeedup(bool)              // Set speedup flag
-	SetTokenThreshold(float32)    // Set timestamp token probability threshold
-	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
-	SetMaxSegmentLength(uint)     // Set max segment length in characters
-	SetTokenTimestamps(bool)      // Set token timestamps flag
-	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)
+	SetOffset(time.Duration)        // Set offset
+	SetDuration(time.Duration)      // Set duration
+	SetThreads(uint)                // Set number of threads to use
+	SetSpeedup(bool)                // Set speedup flag
+	SetSplitOnWord(bool)            // Set split on word flag
+	SetTokenThreshold(float32)      // Set timestamp token probability threshold
+	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
+	SetMaxSegmentLength(uint)       // Set max segment length in characters
+	SetTokenTimestamps(bool)        // Set token timestamps flag
+	SetMaxTokensPerSegment(uint)    // Set max tokens per segment (0 = no limit)
+	SetAudioCtx(uint)               // Set audio encoder context
+	SetInitialPrompt(prompt string) // Set initial prompt
 
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the

bindings/go/whisper.go

@@ -83,7 +83,6 @@ const (
 	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
 	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
 	NumFFT     = C.WHISPER_N_FFT
-	NumMEL     = C.WHISPER_N_MEL
 	HopLength  = C.WHISPER_HOP_LENGTH
 	ChunkSize  = C.WHISPER_CHUNK_SIZE
 )
@@ -103,7 +102,7 @@ var (
 func Whisper_init(path string) *Context {
 	cPath := C.CString(path)
 	defer C.free(unsafe.Pointer(cPath))
-	if ctx := C.whisper_init_from_file(cPath); ctx != nil {
+	if ctx := C.whisper_init_from_file_with_params(cPath, C.whisper_context_default_params()); ctx != nil {
 		return (*Context)(ctx)
 	} else {
 		return nil

bindings/java/build.gradle

@@ -2,12 +2,14 @@ plugins {
     id 'java'
     id 'java-library'
     id 'maven-publish'
+    id 'signing'
 }
 
 archivesBaseName = 'whispercpp'
 group = 'io.github.ggerganov'
 version = '1.4.0'
 
+
 sourceCompatibility = 1.8
 targetCompatibility = 1.8
 
@@ -109,4 +111,23 @@ publishing {
             }
         }
     }
+
+    repositories {
+        maven {
+            def releasesRepoUrl = 'https://s01.oss.sonatype.org/service/local/staging/deploy/maven2/'
+            def snapshotsRepoUrl = 'https://s01.oss.sonatype.org/content/repositories/snapshots/'
+            url = version.endsWith('-SNAPSHOT') ? snapshotsRepoUrl : releasesRepoUrl
+            credentials {
+                username = System.getenv("MAVEN_USERNAME")
+                password = System.getenv("MAVEN_PASSWORD")
+            }
+        }
+    }
+}
+
+signing {
+    def signingKey = System.getenv("PGP_SECRET")
+    def signingPassword = System.getenv("PGP_PASSPHRASE")
+    useInMemoryPgpKeys(signingKey, signingPassword)
+    sign publishing.publications.mavenJava
 }

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

@@ -4,6 +4,7 @@ import com.sun.jna.Structure;
 import com.sun.jna.ptr.PointerByReference;
 import io.github.ggerganov.whispercpp.ggml.GgmlType;
 import io.github.ggerganov.whispercpp.WhisperModel;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 
 import java.util.List;
 
@@ -23,8 +24,9 @@ public class WhisperContext extends Structure {
     public PointerByReference vocab;
     public PointerByReference state;
 
-    /** populated by whisper_init_from_file() */
+    /** populated by whisper_init_from_file_with_params() */
     String path_model;
+    WhisperContextParams params;
 
 //    public static class ByReference extends WhisperContext implements Structure.ByReference {
 //    }

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

@@ -2,12 +2,16 @@ package io.github.ggerganov.whispercpp;
 
 import com.sun.jna.Native;
 import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.bean.WhisperSegment;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
 
 import java.io.File;
 import java.io.FileNotFoundException;
 import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
 
 /**
  * Before calling most methods, you must call `initContext(modelPath)` to initialise the `ctx` Pointer.
@@ -15,8 +19,9 @@ import java.io.IOException;
 public class WhisperCpp implements AutoCloseable {
     private WhisperCppJnaLibrary lib = WhisperCppJnaLibrary.instance;
     private Pointer ctx = null;
-    private Pointer greedyPointer = null;
-    private Pointer beamPointer = null;
+    private Pointer paramsPointer = null;
+    private Pointer greedyParamsPointer = null;
+    private Pointer beamParamsPointer = null;
 
     public File modelDir() {
         String modelDirPath = System.getenv("XDG_CACHE_HOME");
@@ -31,6 +36,18 @@ public class WhisperCpp implements AutoCloseable {
      * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
      */
     public void initContext(String modelPath) throws FileNotFoundException {
+        initContextImpl(modelPath, getContextDefaultParams());
+    }
+
+    /**
+     * @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 {
+        initContextImpl(modelPath, params);
+    }
+
+    private void initContextImpl(String modelPath, WhisperContextParams params) throws FileNotFoundException {
         if (ctx != null) {
             lib.whisper_free(ctx);
         }
@@ -43,13 +60,26 @@ public class WhisperCpp implements AutoCloseable {
             modelPath = new File(modelDir(), modelPath).getAbsolutePath();
         }
 
-        ctx = lib.whisper_init_from_file(modelPath);
+        ctx = lib.whisper_init_from_file_with_params(modelPath, params);
 
         if (ctx == null) {
             throw new FileNotFoundException(modelPath);
         }
     }
 
+    /**
+     * 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));`
+     */
+    public WhisperContextParams getContextDefaultParams() {
+        paramsPointer = lib.whisper_context_default_params_by_ref();
+        WhisperContextParams params = new WhisperContextParams(paramsPointer);
+        params.read();
+        return params;
+    }
+    
     /**
      * Provides default params which can be used with `whisper_full()` etc.
      * Because this function allocates memory for the params, the caller must call either:
@@ -63,15 +93,15 @@ public class WhisperCpp implements AutoCloseable {
 
         // whisper_full_default_params_by_ref allocates memory which we need to delete, so only create max 1 pointer for each strategy.
         if (strategy == WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY) {
-            if (greedyPointer == null) {
-                greedyPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            if (greedyParamsPointer == null) {
+                greedyParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
             }
-            pointer = greedyPointer;
+            pointer = greedyParamsPointer;
         } else {
-            if (beamPointer == null) {
-                beamPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            if (beamParamsPointer == null) {
+                beamParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
             }
-            pointer = beamPointer;
+            pointer = beamParamsPointer;
         }
 
         WhisperFullParams params = new WhisperFullParams(pointer);
@@ -93,13 +123,17 @@ public class WhisperCpp implements AutoCloseable {
     }
 
     private void freeParams() {
-        if (greedyPointer != null) {
-            Native.free(Pointer.nativeValue(greedyPointer));
-            greedyPointer = null;
+        if (paramsPointer != null) {
+            Native.free(Pointer.nativeValue(paramsPointer));
+            paramsPointer = null;
         }
-        if (beamPointer != null) {
-            Native.free(Pointer.nativeValue(beamPointer));
-            beamPointer = null;
+        if (greedyParamsPointer != null) {
+            Native.free(Pointer.nativeValue(greedyParamsPointer));
+            greedyParamsPointer = null;
+        }
+        if (beamParamsPointer != null) {
+            Native.free(Pointer.nativeValue(beamParamsPointer));
+            beamParamsPointer = null;
         }
     }
 
@@ -129,6 +163,28 @@ public class WhisperCpp implements AutoCloseable {
 
         return str.toString().trim();
     }
+    public List<WhisperSegment> fullTranscribeWithTime(WhisperFullParams whisperParams, float[] audioData) throws IOException {
+        if (ctx == null) {
+            throw new IllegalStateException("Model not initialised");
+        }
+
+        if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
+            throw new IOException("Failed to process audio");
+        }
+
+        int nSegments = lib.whisper_full_n_segments(ctx);
+        List<WhisperSegment> segments= new ArrayList<>(nSegments);
+
+
+        for (int i = 0; i < nSegments; i++) {
+            long t0 = lib.whisper_full_get_segment_t0(ctx, i);
+            String text = lib.whisper_full_get_segment_text(ctx, i);
+            long t1 = lib.whisper_full_get_segment_t1(ctx, i);
+            segments.add(new WhisperSegment(t0,t1,text));
+        }
+
+        return segments;
+    }
 
 //    public int getTextSegmentCount(Pointer ctx) {
 //        return lib.whisper_full_n_segments(ctx);

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

@@ -5,6 +5,7 @@ import com.sun.jna.Native;
 import com.sun.jna.Pointer;
 import io.github.ggerganov.whispercpp.model.WhisperModelLoader;
 import io.github.ggerganov.whispercpp.model.WhisperTokenData;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 
 public interface WhisperCppJnaLibrary extends Library {
@@ -13,12 +14,31 @@ public interface WhisperCppJnaLibrary extends Library {
     String whisper_print_system_info();
 
     /**
-     * Allocate (almost) all memory needed for the model by loading from a file.
+     * DEPRECATED. Allocate (almost) all memory needed for the model by loading from a file.
      *
      * @param path_model Path to the model file
      * @return Whisper context on success, null on failure
      */
     Pointer whisper_init_from_file(String path_model);
+    
+    /**
+     * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
+     * Because this function allocates memory for the params, the caller must call either:
+     * - call `whisper_free_context_params()`
+     * - `Native.free(Pointer.nativeValue(pointer));`
+     */
+    Pointer whisper_context_default_params_by_ref();
+
+    void whisper_free_context_params(Pointer params);
+
+    /**
+     * Allocate (almost) all memory needed for the model by loading from a file.
+     *
+     * @param path_model Path to the model file
+     * @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);
 
     /**
      * Allocate (almost) all memory needed for the model by loading from a buffer.

bindings/java/src/main/java/io/github/ggerganov/whispercpp/bean/WhisperSegment.java

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

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

@@ -0,0 +1,31 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.*;
+
+import java.util.Arrays;
+import java.util.List;
+
+/**
+ * Parameters for the whisper_init_from_file_with_params() function.
+ * If you change the order or add new parameters, make sure to update the default values in whisper.cpp:
+ * whisper_context_default_params()
+ */
+public class WhisperContextParams extends Structure {
+
+    public WhisperContextParams(Pointer p) {
+        super(p);
+    }
+
+    /** Use GPU for inference Number (default = true) */
+    public CBool use_gpu;
+
+    /** Use GPU for inference Number (default = true) */
+    public void useGpu(boolean enable) {
+        use_gpu = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("use_gpu");
+    }
+}

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

@@ -58,6 +58,9 @@ public class WhisperFullParams extends Structure {
         no_context = enable ? CBool.FALSE : CBool.TRUE;
     }
 
+    /** Generate timestamps or not? */
+    public CBool no_timestamps;
+
     /** Flag to force single segment output (useful for streaming). (default = false) */
     public CBool single_segment;
 
@@ -304,10 +307,16 @@ public class WhisperFullParams extends Structure {
         logits_filter_callback = CallbackReference.getFunctionPointer(callback);
     }
 
+    /** Grammar stuff */
+    public Pointer grammar_rules;
+    public long n_grammar_rules;
+    public long i_start_rule;
+    public float grammar_penalty;
+
     @Override
     protected List<String> getFieldOrder() {
         return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
-                "no_context", "single_segment",
+                "no_context", "single_segment", "no_timestamps",
                 "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
                 "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
                 "tdrz_enable", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
@@ -316,6 +325,7 @@ public class WhisperFullParams extends Structure {
                 "new_segment_callback", "new_segment_callback_user_data",
                 "progress_callback", "progress_callback_user_data",
                 "encoder_begin_callback", "encoder_begin_callback_user_data",
-                "logits_filter_callback", "logits_filter_callback_user_data");
+                "logits_filter_callback", "logits_filter_callback_user_data",
+                "grammar_rules", "n_grammar_rules", "i_start_rule", "grammar_penalty");
     }
 }

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

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

bindings/javascript/README.md

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

bindings/javascript/emscripten.cpp

@@ -20,7 +20,7 @@ struct whisper_context * g_context;
 EMSCRIPTEN_BINDINGS(whisper) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         if (g_context == nullptr) {
-            g_context = whisper_init_from_file(path_model.c_str());
+            g_context = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
             if (g_context != nullptr) {
                 return true;
             } else {

bindings/javascript/libwhisper.worker.js

@@ -1 +1 @@
-"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:function(f){(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f)},postMessage:function(msg){parentPort.postMessage(msg)},performance:global.performance||{now:function(){return Date.now()}}})}var initializedJS=false;var pendingNotifiedProxyingQueues=[];function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module).then(function(instance){Module=instance})}else if(e.data.cmd==="run"){Module["__performance_now_clock_drift"]=performance.now()-e.data.time;Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};
+"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:f=>(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f),postMessage:msg=>parentPort.postMessage(msg),performance:global.performance||{now:Date.now}})}var initializedJS=false;function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var module=Module["wasmModule"];Module["wasmModule"]=null;var instance=new WebAssembly.Instance(module,info);return receiveInstance(instance)};self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let messageQueue=[];self.onmessage=e=>messageQueue.push(e);self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"});for(let msg of messageQueue){handleMessage(msg)}self.onmessage=handleMessage};Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=(...args)=>{postMessage({cmd:"callHandler",handler:handler,args:args})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["__emscripten_thread_mailbox_await"](e.data.pthread_ptr);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){throw ex}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="checkMailbox"){if(initializedJS){Module["checkMailbox"]()}}else if(e.data.cmd){err(`worker.js received unknown command ${e.data.cmd}`);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}}self.onmessage=handleMessage;

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.4.2",
+  "version": "1.5.4",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/.gitignore

@@ -1,6 +1,8 @@
 Makefile
 ggml.c
 ggml.h
+ggml-alloc.c
+ggml-alloc.h
 whisper.bundle
 whisper.cpp
 whisper.h

bindings/ruby/ext/extconf.rb

@@ -3,6 +3,14 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
 
 

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

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

bindings/ruby/ext/ggml-backend.c

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

bindings/ruby/ext/ggml-backend.h

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

bindings/ruby/ext/ggml-impl.h

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

bindings/ruby/ext/ggml-quants.h

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

bindings/ruby/ext/ruby_whisper.cpp

@@ -87,7 +87,7 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   if (!rb_respond_to(whisper_model_file_path, rb_intern("to_s"))) {
     rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
   }
-  rw->context = whisper_init_from_file(StringValueCStr(whisper_model_file_path));
+  rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), whisper_context_default_params());
   if (rw->context == nullptr) {
     rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
   }

coreml/whisper-encoder-impl.h

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

coreml/whisper-encoder.h

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

coreml/whisper-encoder.mm

@@ -22,7 +22,13 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
 
-    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+    // select which device to run the Core ML model on
+    MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
+    // config.computeUnits = MLComputeUnitsCPUAndGPU;
+    //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
+    config.computeUnits = MLComputeUnitsAll;
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 
     if (data == NULL) {
         return NULL;
@@ -42,20 +48,24 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx) {
 
 void whisper_coreml_encode(
         const whisper_coreml_context * ctx,
+                             int64_t   n_ctx,
+                             int64_t   n_mel,
                                float * mel,
                                float * out) {
     MLMultiArray * inMultiArray = [
         [MLMultiArray alloc] initWithDataPointer: mel
-                                           shape: @[@1, @80, @3000]
+                                           shape: @[@1, @(n_mel), @(n_ctx)]
                                         dataType: MLMultiArrayDataTypeFloat32
-                                         strides: @[@(240000), @(3000), @1]
+                                         strides: @[@(n_ctx*n_mel), @(n_ctx), @1]
                                      deallocator: nil
                                            error: nil
     ];
 
-    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+    @autoreleasepool {
+        whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
 
-    memcpy(out, outCoreML.output.dataPointer, outCoreML.output.count * sizeof(float));
+        memcpy(out, outCoreML.output.dataPointer, outCoreML.output.count * sizeof(float));
+    }
 }
 
 #if __cplusplus

examples/CMakeLists.txt

@@ -14,6 +14,10 @@ if (WHISPER_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
+if (WHISPER_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
 # common
 
 set(TARGET common)
@@ -23,6 +27,7 @@ add_library(${TARGET} STATIC
     common.cpp
     common-ggml.h
     common-ggml.cpp
+    grammar-parser.cpp
     )
 
 include(DefaultTargetOptions)
@@ -64,9 +69,13 @@ elseif(CMAKE_JS_VERSION)
 else()
     add_subdirectory(main)
     add_subdirectory(stream)
+    add_subdirectory(server)
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(quantize)
     add_subdirectory(talk)
     add_subdirectory(talk-llama)
+    add_subdirectory(lsp)
 endif()
+
+add_subdirectory(wchess)

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

@@ -11,6 +11,7 @@ const whisperParamsMock = {
   language: "en",
   model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+  use_gpu: true,
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -36,6 +36,7 @@ struct whisper_params {
     bool print_colors   = false;
     bool print_progress = false;
     bool no_timestamps  = false;
+    bool use_gpu        = true;
 
     std::string language = "en";
     std::string prompt;
@@ -153,7 +154,9 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+    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");
@@ -315,10 +318,12 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   std::string language = whisper_params.Get("language").As<Napi::String>();
   std::string model = whisper_params.Get("model").As<Napi::String>();
   std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
 
   params.language = language;
   params.model = model;
   params.fname_inp.emplace_back(input);
+  params.use_gpu = use_gpu;
 
   Napi::Function callback = info[1].As<Napi::Function>();
   Worker* worker = new Worker(callback, params);

examples/addon.node/index.js

@@ -11,6 +11,7 @@ const whisperParams = {
   language: "en",
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
   fname_inp: "../../samples/jfk.wav",
+  use_gpu: true,
 };
 
 const arguments = process.argv.slice(2);

examples/bench.wasm/emscripten.cpp

@@ -23,7 +23,9 @@ void bench_main(size_t index) {
 
     fprintf(stderr, "%s: running benchmark with %d threads - please wait...\n", __func__, n_threads);
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return;
     }
@@ -57,7 +59,7 @@ EMSCRIPTEN_BINDINGS(bench) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     if (g_worker.joinable()) {
                         g_worker.join();

examples/bench/bench.cpp

@@ -1,6 +1,7 @@
 #include "whisper.h"
 
 #include <cstdio>
+#include <cstring>
 #include <string>
 #include <thread>
 
@@ -10,6 +11,8 @@ struct whisper_params {
     int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
 
     std::string model = "models/ggml-base.en.bin";
+
+    bool use_gpu = true;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -22,9 +25,10 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
-        else if (arg == "-m" || arg == "--model")   { params.model     = argv[++i]; }
-        else if (arg == "-w" || arg == "--what")    { params.what     = atoi(argv[++i]); }
+        else if (arg == "-t"  || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")    { params.what      = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")  { params.use_gpu   = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -44,16 +48,20 @@ 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, "                           %-7s  0 - whisper encoder\n",                         "");
+    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int whisper_bench_encoder(const whisper_params & params) {
+int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     {
         fprintf(stderr, "\n");
@@ -65,16 +73,65 @@ int whisper_bench_encoder(const whisper_params & params) {
         return 2;
     }
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return 3;
     }
-
+    // heat encoder
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
         return 4;
     }
 
+    whisper_token tokens[512];
+    memset(tokens, 0, sizeof(tokens));
+
+    // prompt heat
+    if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation heat
+    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    whisper_reset_timings(ctx);
+
+    // actual run
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // batched decoding
+    for (int i = 0; i < 64; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // prompt processing
+    for (int i = 0; i < 16; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
     whisper_print_timings(ctx);
     whisper_free(ctx);
 
@@ -103,7 +160,7 @@ int main(int argc, char ** argv) {
     int ret = -1;
 
     switch (params.what) {
-        case 0: ret = whisper_bench_encoder(params);                break;
+        case 0: ret = whisper_bench_full(params);                break;
         case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
         case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
         default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;

examples/command.wasm/emscripten.cpp

@@ -243,7 +243,7 @@ EMSCRIPTEN_BINDINGS(command) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/command/command.cpp

@@ -6,9 +6,10 @@
 // ref: https://github.com/ggerganov/whisper.cpp/issues/171
 //
 
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
+#include "grammar-parser.h"
 
 #include <sstream>
 #include <cassert>
@@ -21,6 +22,11 @@
 #include <vector>
 #include <map>
 
+bool file_exists(const std::string & fname) {
+    std::ifstream f(fname.c_str());
+    return f.good();
+}
+
 // command-line parameters
 struct whisper_params {
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -30,20 +36,27 @@ struct whisper_params {
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
 
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
+    float vad_thold  = 0.6f;
+    float freq_thold = 100.0f;
+
+    float grammar_penalty = 100.0f;
+
+    grammar_parser::parse_state grammar_parsed;
 
     bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
     bool no_timestamps = true;
+    bool use_gpu       = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
     std::string fname_out;
     std::string commands;
     std::string prompt;
+    std::string context;
+    std::string grammar;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -68,11 +81,15 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
         else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
         else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
+        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -101,21 +118,36 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
     fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
     fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
+    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
+std::string transcribe(
+                 whisper_context * ctx,
+            const whisper_params & params,
+        const std::vector<float> & pcmf32,
+               const std::string & grammar_rule,
+                           float & logprob_min,
+                           float & logprob_sum,
+                             int & n_tokens,
+                         int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
-    prob = 0.0f;
+    logprob_min = 0.0f;
+    logprob_sum = 0.0f;
+    n_tokens    = 0;
     t_ms = 0;
 
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    //whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH);
 
     wparams.print_progress   = false;
     wparams.print_special    = params.print_special;
@@ -123,19 +155,41 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     wparams.print_timestamps = !params.no_timestamps;
     wparams.translate        = params.translate;
     wparams.no_context       = true;
+    wparams.no_timestamps    = params.no_timestamps;
     wparams.single_segment   = true;
     wparams.max_tokens       = params.max_tokens;
     wparams.language         = params.language.c_str();
     wparams.n_threads        = params.n_threads;
 
-    wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
+    wparams.audio_ctx = params.audio_ctx;
+    wparams.speed_up  = params.speed_up;
+
+    wparams.temperature     = 0.4f;
+    wparams.temperature_inc = 1.0f;
+    wparams.greedy.best_of  = 5;
+
+    wparams.beam_search.beam_size = 5;
+
+    wparams.initial_prompt = params.context.data();
+
+    const auto & grammar_parsed = params.grammar_parsed;
+    auto grammar_rules = grammar_parsed.c_rules();
+
+    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
+        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
+            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, grammar_rule.c_str());
+        } else {
+            wparams.grammar_rules   = grammar_rules.data();
+            wparams.n_grammar_rules = grammar_rules.size();
+            wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
+            wparams.grammar_penalty = params.grammar_penalty;
+        }
+    }
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
     }
 
-    int prob_n = 0;
     std::string result;
 
     const int n_segments = whisper_full_n_segments(ctx);
@@ -144,19 +198,17 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
 
         result += text;
 
-        const int n_tokens = whisper_full_n_tokens(ctx, i);
-        for (int j = 0; j < n_tokens; ++j) {
+        const int n = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n; ++j) {
             const auto token = whisper_full_get_token_data(ctx, i, j);
 
-            prob += token.p;
-            ++prob_n;
+            if(token.plog > 0.0f) exit(0);
+            logprob_min = std::min(logprob_min, token.plog);
+            logprob_sum += token.plog;
+            ++n_tokens;
         }
     }
 
-    if (prob_n > 0) {
-        prob /= prob_n;
-    }
-
     const auto t_end = std::chrono::high_resolution_clock::now();
     t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
 
@@ -247,7 +299,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
         fprintf(stderr, " ]\n");
     }
 
-    std::string  k_prompt = "select one from the available words: ";
+    std::string k_prompt = "select one from the available words: ";
     for (int i = 0; i < (int) allowed_commands.size(); ++i) {
         if (i > 0) {
             k_prompt += ", ";
@@ -415,7 +467,9 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
     bool is_running = true;
     bool ask_prompt = true;
 
-    float prob = 0.0f;
+    float logprob_min = 0.0f;
+    float logprob_sum = 0.0f;
+    int   n_tokens    = 0;
 
     std::vector<float> pcmf32_cur;
 
@@ -453,7 +507,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
                 // detect the commands
                 audio.get(params.command_ms, pcmf32_cur);
 
-                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "", logprob_min, logprob_sum, n_tokens, t_ms));
 
                 const auto words = get_words(txt);
 
@@ -489,18 +543,27 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
 
 // general-purpose mode
 // freely transcribe the voice into text
-int process_general_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running  = true;
     bool have_prompt = false;
     bool ask_prompt  = true;
 
-    float prob0 = 0.0f;
-    float prob  = 0.0f;
+    float logprob_min0 = 0.0f;
+    float logprob_min  = 0.0f;
+
+    float logprob_sum0 = 0.0f;
+    float logprob_sum  = 0.0f;
+
+    int n_tokens0 = 0;
+    int n_tokens  = 0;
 
     std::vector<float> pcmf32_cur;
     std::vector<float> pcmf32_prompt;
 
-    const std::string k_prompt = "Ok Whisper, start listening for commands.";
+    std::string k_prompt = "Ok Whisper, start listening for commands.";
+    if (!params.prompt.empty()) {
+        k_prompt = params.prompt;
+    }
 
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: general-purpose mode\n", __func__);
@@ -533,9 +596,11 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                     // wait for activation phrase
                     audio.get(params.prompt_ms, pcmf32_cur);
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "prompt", logprob_min0, logprob_sum0, n_tokens0, t_ms));
 
-                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+                    const float p = 100.0f * std::exp(logprob_min0);
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms, p = %.2f%%)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms, p);
 
                     const float sim = similarity(txt, k_prompt);
 
@@ -556,19 +621,30 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                     // we have heard the activation phrase, now detect the commands
                     audio.get(params.command_ms, pcmf32_cur);
 
+                    //printf("len prompt:  %.4f\n", pcmf32_prompt.size() / (float) WHISPER_SAMPLE_RATE);
+                    //printf("len command: %.4f\n", pcmf32_cur.size() / (float) WHISPER_SAMPLE_RATE);
+
+                    // prepend 3 second of silence
+                    pcmf32_cur.insert(pcmf32_cur.begin(), 3.0f*WHISPER_SAMPLE_RATE, 0.0f);
+
                     // prepend the prompt audio
                     pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "root", logprob_min, logprob_sum, n_tokens, t_ms));
 
-                    prob = 100.0f*(prob - prob0);
+                    //const float p = 100.0f * std::exp((logprob - logprob0) / (n_tokens - n_tokens0));
+                    const float p = 100.0f * std::exp(logprob_min);
 
                     //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
 
                     // find the prompt in the text
                     float best_sim = 0.0f;
                     size_t best_len = 0;
-                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                    for (size_t n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        if (n >= txt.size()) {
+                            break;
+                        }
+
                         const auto prompt = txt.substr(0, n);
 
                         const float sim = similarity(prompt, k_prompt);
@@ -581,9 +657,16 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                         }
                     }
 
-                    const std::string command = ::trim(txt.substr(best_len));
+                    fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
+                    if (best_len == 0) {
+                        fprintf(stdout, "%s: WARNING: command not recognized, try again\n", __func__);
+                    } else {
+                        // cut the prompt from the decoded text
+                        const std::string command = ::trim(txt.substr(best_len));
+
+                        fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    }
 
-                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
                     fprintf(stdout, "\n");
                 }
 
@@ -610,7 +693,10 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     // print some info about the processing
     {
@@ -648,12 +734,36 @@ int main(int argc, char ** argv) {
 
     int  ret_val = 0;
 
-    if (!params.commands.empty()) {
-        ret_val = process_command_list(ctx, audio, params);
-    } else if (!params.prompt.empty()) {
-        ret_val = always_prompt_transcription(ctx, audio, params);
-    } else {
-        ret_val = process_general_transcription(ctx, audio, params);
+    if (!params.grammar.empty()) {
+        auto & grammar = params.grammar_parsed;
+        if (file_exists(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            ret_val = 1;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (ret_val == 0) {
+        if (!params.commands.empty()) {
+            ret_val = process_command_list(ctx, audio, params);
+        } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
+            ret_val = always_prompt_transcription(ctx, audio, params);
+        } else {
+            ret_val = process_general_transcription(ctx, audio, params);
+        }
     }
 
     audio.pause();

examples/common-ggml.cpp

@@ -9,6 +9,11 @@ static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
     {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
     {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
     {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+    {"q2_k", GGML_FTYPE_MOSTLY_Q2_K},
+    {"q3_k", GGML_FTYPE_MOSTLY_Q3_K},
+    {"q4_k", GGML_FTYPE_MOSTLY_Q4_K},
+    {"q5_k", GGML_FTYPE_MOSTLY_Q5_K},
+    {"q6_k", GGML_FTYPE_MOSTLY_Q6_K},
 };
 
 void ggml_print_ftypes(FILE * fp) {
@@ -48,15 +53,18 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
         case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
         case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
+        case GGML_FTYPE_MOSTLY_Q2_K: qtype = GGML_TYPE_Q2_K; break;
+        case GGML_FTYPE_MOSTLY_Q3_K: qtype = GGML_TYPE_Q3_K; break;
+        case GGML_FTYPE_MOSTLY_Q4_K: qtype = GGML_TYPE_Q4_K; break;
+        case GGML_FTYPE_MOSTLY_Q5_K: qtype = GGML_TYPE_Q5_K; break;
+        case GGML_FTYPE_MOSTLY_Q6_K: qtype = GGML_TYPE_Q6_K; break;
         case GGML_FTYPE_UNKNOWN:
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
-        case GGML_FTYPE_MOSTLY_Q2_K:
-        case GGML_FTYPE_MOSTLY_Q3_K:
-        case GGML_FTYPE_MOSTLY_Q4_K:
-        case GGML_FTYPE_MOSTLY_Q5_K:
-        case GGML_FTYPE_MOSTLY_Q6_K:
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:
+        case GGML_FTYPE_MOSTLY_IQ2_XS:
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -167,24 +175,17 @@ bool ggml_common_quantize_0(
 
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
-                    {
-                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q4_1:
-                    {
-                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q5_0:
-                    {
-                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q5_1:
-                    {
-                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], hist_cur.data(), nullptr);
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
@@ -192,12 +193,10 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_I16:
                 case GGML_TYPE_I32:
                 case GGML_TYPE_Q8_1:
-                case GGML_TYPE_Q2_K:
-                case GGML_TYPE_Q3_K:
-                case GGML_TYPE_Q4_K:
-                case GGML_TYPE_Q5_K:
-                case GGML_TYPE_Q6_K:
                 case GGML_TYPE_Q8_K:
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ2_XS:
+                case GGML_TYPE_IQ3_XXS:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/common-sdl.cpp

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

examples/common-sdl.h

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

examples/common.cpp

@@ -1,3 +1,5 @@
+#define _USE_MATH_DEFINES // for M_PI
+
 #include "common.h"
 
 // third-party utilities
@@ -13,53 +15,65 @@
 #include <codecvt>
 #include <sstream>
 
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+// Function to check if the next argument exists
+std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+    if (i + 1 < argc && argv[i + 1][0] != '-') {
+        return argv[++i];
+    } else {
+        fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
+        gpt_print_usage(argc, argv, params);
+        exit(0);
+    }
+}
+
 bool gpt_params_parse(int argc, char ** argv, gpt_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]);
+            params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-t" || arg == "--threads") {
-            params.n_threads = std::stoi(argv[++i]);
+            params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-p" || arg == "--prompt") {
-            params.prompt = argv[++i];
+            params.prompt = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-n" || arg == "--n_predict") {
-            params.n_predict = std::stoi(argv[++i]);
+            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-np" || arg == "--n_parallel") {
+            params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_k") {
-            params.top_k = std::max(1, std::stoi(argv[++i]));
+            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_p") {
-            params.top_p = std::stof(argv[++i]);
+            params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--temp") {
-            params.temp = std::stof(argv[++i]);
+            params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--repeat-last-n") {
-            params.repeat_last_n = std::stof(argv[++i]);
+            params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--repeat-penalty") {
-            params.repeat_penalty = std::stof(argv[++i]);
+            params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-b" || arg == "--batch_size") {
-            params.n_batch = std::stoi(argv[++i]);
+            params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-c" || arg == "--context") {
+            params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--ignore-eos") {
+            params.ignore_eos = true;
         } else if (arg == "-m" || arg == "--model") {
-            params.model = argv[++i];
+            params.model = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-i" || arg == "--interactive") {
             params.interactive = true;
         } else if (arg == "-ip" || arg == "--interactive-port") {
             params.interactive = true;
-            params.interactive_port = std::stoi(argv[++i]);
+            params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, params);
             exit(0);
         } else if (arg == "-f" || arg == "--file") {
-            if (++i > argc) {
-                fprintf(stderr, "Invalid file param");
-                break;
-            }
+            get_next_arg(i, argc, argv, arg, params);
             std::ifstream file(argv[i]);
             if (!file) {
                 fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@@ -70,7 +84,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 params.prompt.pop_back();
             }
         } else if (arg == "-tt" || arg == "--token_test") {
-            params.token_test = argv[++i];
+            params.token_test = get_next_arg(i, argc, argv, arg, params);
         }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -102,6 +116,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
     fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "\n");
@@ -598,6 +615,21 @@ 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
@@ -622,6 +654,12 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
 
         fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
     }
+    else if (is_wav_buffer(fname)) {
+        if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
+            return false;
+        }
+    }
     else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
         fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
         return false;
@@ -755,3 +793,46 @@ 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");
+}

examples/common.h

@@ -7,18 +7,25 @@
 #include <vector>
 #include <random>
 #include <thread>
+#include <ctime>
+#include <fstream>
 
 #define COMMON_SAMPLE_RATE 16000
 
 //
-// CLI argument parsing
+// GPT CLI argument parsing
 //
 
 struct gpt_params {
-    int32_t seed      = -1;  // RNG seed
-    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_predict = 200; // new tokens to predict
-    int32_t n_batch   = 8;   // batch size for prompt processing
+    int32_t seed         = -1;   // RNG seed
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict    = 200;  // new tokens to predict
+    int32_t n_parallel   = 1;    // number of parallel streams
+    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
+    int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
+
+    bool ignore_eos = false; // ignore EOS token when generating text
 
     // sampling parameters
     int32_t top_k          = 40;
@@ -128,7 +135,11 @@ 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(
@@ -137,6 +148,104 @@ bool read_wav(
         std::vector<std::vector<float>> & pcmf32s,
         bool stereo);
 
+// Write PCM data into WAV audio file
+class wav_writer {
+private:
+    std::ofstream file;
+    uint32_t dataSize = 0;
+    std::string wav_filename;
+
+    bool write_header(const uint32_t sample_rate,
+                      const uint16_t bits_per_sample,
+                      const uint16_t channels) {
+
+        file.write("RIFF", 4);
+        file.write("\0\0\0\0", 4);    // Placeholder for file size
+        file.write("WAVE", 4);
+        file.write("fmt ", 4);
+
+        const uint32_t sub_chunk_size = 16;
+        const uint16_t audio_format = 1;      // PCM format
+        const uint32_t byte_rate = sample_rate * channels * bits_per_sample / 8;
+        const uint16_t block_align = channels * bits_per_sample / 8;
+
+        file.write(reinterpret_cast<const char *>(&sub_chunk_size), 4);
+        file.write(reinterpret_cast<const char *>(&audio_format), 2);
+        file.write(reinterpret_cast<const char *>(&channels), 2);
+        file.write(reinterpret_cast<const char *>(&sample_rate), 4);
+        file.write(reinterpret_cast<const char *>(&byte_rate), 4);
+        file.write(reinterpret_cast<const char *>(&block_align), 2);
+        file.write(reinterpret_cast<const char *>(&bits_per_sample), 2);
+        file.write("data", 4);
+        file.write("\0\0\0\0", 4);    // Placeholder for data size
+
+        return true;
+    }
+
+    // It is assumed that PCM data is normalized to a range from -1 to 1
+    bool write_audio(const float * data, size_t length) {
+        for (size_t i = 0; i < length; ++i) {
+            const int16_t intSample = data[i] * 32767;
+            file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
+            dataSize += sizeof(int16_t);
+        }
+        if (file.is_open()) {
+            file.seekp(4, std::ios::beg);
+            uint32_t fileSize = 36 + dataSize;
+            file.write(reinterpret_cast<char *>(&fileSize), 4);
+            file.seekp(40, std::ios::beg);
+            file.write(reinterpret_cast<char *>(&dataSize), 4);
+            file.seekp(0, std::ios::end);
+        }
+        return true;
+    }
+
+    bool open_wav(const std::string & filename) {
+        if (filename != wav_filename) {
+            if (file.is_open()) {
+                file.close();
+            }
+        }
+        if (!file.is_open()) {
+            file.open(filename, std::ios::binary);
+            wav_filename = filename;
+            dataSize = 0;
+        }
+        return file.is_open();
+    }
+
+public:
+    bool open(const std::string & filename,
+              const    uint32_t   sample_rate,
+              const    uint16_t   bits_per_sample,
+              const    uint16_t   channels) {
+
+        if (open_wav(filename)) {
+            write_header(sample_rate, bits_per_sample, channels);
+        } else {
+            return false;
+        }
+
+        return true;
+    }
+
+    bool close() {
+        file.close();
+        return true;
+    }
+
+    bool write(const float * data, size_t length) {
+        return write_audio(data, length);
+    }
+
+    ~wav_writer() {
+        if (file.is_open()) {
+            file.close();
+        }
+    }
+};
+
+
 // Apply a high-pass frequency filter to PCM audio
 // Suppresses frequencies below cutoff Hz
 void high_pass_filter(
@@ -155,3 +264,20 @@ 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);

examples/grammar-parser.cpp

@@ -0,0 +1,423 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from whisper.cpp
+    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<whisper_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<whisper_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({WHISPER_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum whisper_gretype start_type = WHISPER_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = WHISPER_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum whisper_gretype type = last_sym_start < out_elements.size()
+                        ? WHISPER_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({WHISPER_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<whisper_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({WHISPER_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({WHISPER_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<whisper_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({WHISPER_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({WHISPER_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    bool is_char_element(whisper_grammar_element elem) {
+        switch (elem.type) {
+            case WHISPER_GRETYPE_CHAR:           return true;
+            case WHISPER_GRETYPE_CHAR_NOT:       return true;
+            case WHISPER_GRETYPE_CHAR_ALT:       return true;
+            case WHISPER_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
+                case WHISPER_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case WHISPER_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case WHISPER_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case WHISPER_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case WHISPER_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                case WHISPER_GRETYPE_ALT:
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                case WHISPER_GRETYPE_CHAR_NOT:
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<whisper_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != WHISPER_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with WHISPER_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            whisper_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case WHISPER_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case WHISPER_GRETYPE_CHAR_ALT:
+                    case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
+        std::vector<const whisper_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

examples/grammar-parser.h

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

examples/helpers.js

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

examples/livestream.sh

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

examples/lsp/CMakeLists.txt

@@ -0,0 +1,9 @@
+if (WHISPER_SDL2)
+    # stream
+    set(TARGET lsp)
+    add_executable(${TARGET} lsp.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/lsp/README.md

@@ -0,0 +1,104 @@
+# Language Server
+
+This example consists of a simple language server to expose both unguided
+and guided (command) transcriptions by sending json messages over stdout/stdin
+as well as a rather robust vim plugin that makes use of the language server.
+
+## Vim plugin quick start
+
+Compile the language server with
+
+```bash
+make lsp
+```
+Install the plugin itself by copying or symlinking whisper.vim into ~/.vim/autoload/
+
+In your vimrc, set the path of your whisper.cpp directory and optionally add some keybinds.
+
+```vim
+let g:whisper_dir = "~/whisper.cpp"
+" Start listening for commands when Ctrl - g is pressed in normal mode
+nnoremap <C-G> call whisper#requestCommands()<CR>
+" Start unguided transcription when Ctrl - g is pressed in insert mode
+inoremap <C-G> <Cmd>call whisper#doTranscription()<CR>
+```
+
+## Vim plugin usage
+
+The vim plugin was designed to closely follow the mnemonics of vim
+
+`s:spoken_dict` is used to translate keys to their spoken form.
+
+
+Keys corresponding to a string use that spoken value normally and when a motion is expected, but use the key itself when a character is expected.  
+Keys corresponding to a dict, like `i`, can have manual difinitions given to each possible commandset.
+
+0 is normal (insert), 1 is motion (inside), 2 is it's usage as a single key ([till] i), and 3 is it's usage in an area selection (s -> [around] sentence)
+
+Some punctuation items, like `-` are explicitly given pronunciations to prevent them from being picked as punctuation instead of an actual command word.
+
+Not all commands will tokenize to a single token and this can interfere with interpretation. "yank" as an example, takes multiple tokens and correspondingly, will give more accurate detection when only the first "ya" is used. While it could be changed to something else that is a single token (copy), value was placed on maintaining vim mnemonics.
+
+Commands that would normally move the editor into insert mode (insert, append, open, change) will begin unguided transcription.
+Unguided transcription will end when a speech segment ends in exit.
+Presence of punctuation can be designated by whether or not you add a pause between the previous speech segment and exit.
+Exiting only occurs if exit is the last word, so "Take the first exit on your right" would not cause transcription to end.
+
+After a command is evaluated, the plugin will continue listening for the next command.
+
+While in command mode, "Exit" will end listening.
+
+A best effort approach is taken to keep track of audio that is recorded while a previous chunk is still processing and immediately interpret it afterwards, but the current voice detection still needs a fairly sizable gap to determine when a command has been spoken.
+
+Log information is sent to a special `whisper_log` buffer and can be accessed with
+```vim
+:e whisper_log
+```
+
+## Vim plugin configuration
+
+`g:whisper_dir`  
+A full path to the whisper.cpp repo. It can be expanded in the definition like so:
+```vim
+let g:whisper_dir = expand("~/whisper.cpp/")
+```
+(The WHISPER_CPP_HOME environment variable is also checked for users of the existing whisper.nvim script)
+
+`g:whisper_lsp_path`  
+Can be used to manually set the path to the language server.
+If not defined, it will be inferred from the above whisper_dir
+
+`g:whisper_model_path`  
+A full path to the model to load. If not defined, it will default to ggml-base.en.bin
+
+`g:whisper_user_commands`  
+A dictionary of spoken commands that correspond to either strings or funcrefs.
+This can be used to create connections with other user plugins, for example
+```vim
+let g:whisper_user_commands = {"gen": "llama#doLlamaGen"}
+```
+will trigger the llama.cpp plugin to begin generation when "gen" is spoken
+
+## Language server methods
+
+`registerCommandset`  
+`params` is a list of strings that should be checked for with this commandset. The server prepends a space to these strings before tokenizing.  
+Responds with  
+`result.index` an integer index for the commandset registered, which should be included when initiating a guided transcription to select this commandset.
+Will return an error if any of the commands in the commandset have duplicate tokenizations
+
+`guided`  
+`params.commandset_index` An index returned by a corresponding commandset registration. If not set, the most recently registered commandset is used.
+`params.timestamp` A positive unsigned integer which designates a point in time which audio should begin processing from. If left blank, the start point of audio processing will be the moment the message is recieved. This should be left blank unless you have a timestamp from a previous response.  
+Responds with  
+`result.command_index` The numerical index (starting from 0) of the detected command in the selected commandset
+`result.command_text` A string containing the command as provided in the commandset
+`result.timestamp` A positive unsigned integer that designates the point in time which audio stopped being processed at. Pass this timestamp back in a subsequent message to mask the latency of transcription.
+
+`unguided`  
+`params.no_context` Sets the corresponding whisper `no_context` param. Defaults to true. Might provide more accurate results for consecutive unguided transcriptions if those after the first are set to false.
+`params.prompt` If provided, sets the initial prompt used during transcription.
+`params.timestamp` A positive unsigned integer which designates a point in time which audio should begin processing from. If left blank, the start point of audio processing will be the moment the message is recieved. This should be left blank unless you have a timestamp from a previous response.  
+Responds with  
+`result.transcription` A string containing the transcribed text.  N.B. This will almost always start with a space due to how text is tokenized.
+`result.timestamp` A positive unsigned integer that designates the point in time which audio stopped being processed at. Pass this timestamp back in a subsequent message to mask the latency of transcription.

examples/lsp/lsp.cpp

@@ -0,0 +1,463 @@
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "json.hpp"
+
+#include <iostream>
+#include <cassert>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+#include <deque>
+#include <set>
+
+using json = nlohmann::json;
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t prompt_ms  = 5000;
+    int32_t command_ms = 8000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool use_gpu       = true;
+
+    std::string language  = "en";
+    std::string model     = "models/ggml-base.en.bin";
+};
+struct command {
+    std::vector<whisper_token> tokens;
+    std::string plaintext;
+};
+struct commandset {
+    std::vector<struct command> commands;
+    std::vector<whisper_token> prompt_tokens;
+    // TODO: Store longest command?
+    // Multi-token commands should have probabilities of subsequent logits
+    // given that the prior logit is correct.
+    // In this case, all commands must be iterated.
+    // This however, is likely highly involved as different tokens
+    // almost certainly have different spoken lengths
+    // It would also have performance implications equivalent to a beam search
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms" || arg == "--command-ms")    { params.command_ms    = std::stoi(argv[++i]); }
+        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 == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,         --help           [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,       --threads N      [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -pms N,     --prompt-ms N    [%-7d] prompt duration in milliseconds\n",             params.prompt_ms);
+    fprintf(stderr, "  -cms N,     --command-ms N   [%-7d] command duration in milliseconds\n",            params.command_ms);
+    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, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "\n");
+}
+uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
+    using namespace std::chrono;
+    uint64_t time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+    uint64_t start_time = time_now;
+    if (jparams.contains("timestamp")) {
+        start_time = jparams.at("timestamp");
+    }
+    if(time_now - start_time < 500) {
+        //wait for a backlog of audio
+        std::this_thread::sleep_for(milliseconds(500 - (time_now - start_time)));
+        time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+    } else if (time_now - start_time > 1000) {
+        audio.get(time_now-start_time, pcmf32);
+        size_t max_offset = pcmf32.size() - WHISPER_SAMPLE_RATE;
+        for(size_t offset=0;offset < max_offset;offset+=WHISPER_SAMPLE_RATE/10) {
+            std::vector<float> audio_chunk(&pcmf32[offset], &pcmf32[offset+WHISPER_SAMPLE_RATE]);
+            if(::vad_simple(audio_chunk, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                pcmf32.resize(offset+WHISPER_SAMPLE_RATE);
+                if (offset*1000/WHISPER_SAMPLE_RATE+1000 > maxlength_ms) {
+                    //remove samples from the beginning
+                    pcmf32.erase(pcmf32.begin(),pcmf32.end()-(maxlength_ms*WHISPER_SAMPLE_RATE/1000));
+                    fprintf(stderr, "Shortened samples");
+                }
+                return start_time + offset*1000/WHISPER_SAMPLE_RATE+1000;
+            }
+        }
+    }
+    size_t window_duration = std::max((uint64_t)1000, time_now-start_time);
+    audio.get(window_duration, pcmf32);
+    while (!::vad_simple(pcmf32, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+        std::this_thread::sleep_for(milliseconds(100));
+        time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+        window_duration = std::max((uint64_t)1000,time_now-start_time);
+        audio.get(window_duration, pcmf32);
+    }
+    if (time_now - start_time > maxlength_ms) {
+        audio.get(maxlength_ms, pcmf32);
+    } else {
+        audio.get(time_now - start_time, pcmf32);
+    }
+
+    return time_now;
+}
+
+json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
+    std::vector<whisper_token> prompt_tokens;
+    std::vector<float> pcmf32;
+    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 10000U, pcmf32);
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    if (jparams.contains("prompt")) {
+        // unlikely to see much use. Under normal circumstances, no_context would be set to false
+        std::string prompt = jparams.at("prompt");
+        prompt_tokens.resize(1024);
+        int n = whisper_tokenize(ctx, prompt.c_str(), prompt_tokens.data(), 1024);
+        prompt_tokens.resize(n);
+
+        wparams.prompt_tokens    = prompt_tokens.data();
+        wparams.prompt_n_tokens  = prompt_tokens.size();
+    }
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = false;
+    wparams.translate        = params.translate;
+    wparams.no_context       = jparams.value("no_context", true);
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+    wparams.suppress_non_speech_tokens = true;
+    // run the transformer and a single decoding pass
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+        throw json{
+            {"code", -32803},
+            {"message", "ERROR: whisper_full() failed"}
+        };
+    }
+    std::string result = whisper_full_get_segment_text(ctx,0);
+    return json {
+        {"transcription", result},
+        {"timestamp", unprocessed_audio_timestamp}
+    };
+}
+
+// command-list mode
+// guide the transcription to match the most likely command from a provided list
+json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
+    struct commandset cs = commandset_list[jparams.value("commandset_index", commandset_list.size()-1)];
+    std::vector<float> pcmf32;
+    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 2000U, pcmf32);
+
+    fprintf(stderr, "%s: Speech detected! Processing ...\n", __func__);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = false;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = 1;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    // TODO: Do some time testing. Does an overly long prompt slow down processing?
+    // Set up command sets/precompute prompts
+    wparams.prompt_tokens    = cs.prompt_tokens.data();
+    wparams.prompt_n_tokens  = cs.prompt_tokens.size();
+    // TODO: properly expose as option
+    wparams.suppress_non_speech_tokens = true;
+
+    // run the transformer and a single decoding pass
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+        throw json{
+            {"code", -32803},
+            {"message", "ERROR: whisper_full() failed"}//TODO: format string (sprintf?)
+        };
+    }
+
+    // estimate command probability
+    // NOTE: not optimal
+    {
+        const auto * logits = whisper_get_logits(ctx);
+
+        std::vector<float> probs(whisper_n_vocab(ctx), 0.0f);
+
+        // compute probs from logits via softmax
+        {
+            float max = -1e9;
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                max = std::max(max, logits[i]);
+            }
+
+            float sum = 0.0f;
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                probs[i] = expf(logits[i] - max);
+                sum += probs[i];
+            }
+
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                probs[i] /= sum;
+            }
+        }
+
+        std::vector<std::pair<float, int>> probs_id;
+
+        // In my testing, the most verbose token is always the desired.
+        // TODO: Trim commandset struct once efficacy has been verified
+        for (int i = 0; i < (int) cs.commands.size(); ++i) {
+            probs_id.emplace_back(probs[cs.commands[i].tokens[0]], i);
+        }
+
+        // sort descending
+        {
+            using pair_type = decltype(probs_id)::value_type;
+            std::sort(probs_id.begin(), probs_id.end(), [](const pair_type & a, const pair_type & b) {
+                    return a.first > b.first;
+                    });
+        }
+        int id = probs_id[0].second;
+        return json{
+            {"command_index", id},
+                {"command_text", cs.commands[id].plaintext},
+                {"timestamp", unprocessed_audio_timestamp},
+        };
+    }
+}
+
+json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
+    // TODO: check for token collision
+    struct commandset cs;
+
+    std::string  k_prompt = " select one from the available words: ";
+    std::set<whisper_token> token_set;
+    whisper_token tokens[32];
+    for (std::string s : jparams) {
+        std::vector<whisper_token> token_vec;
+        // The existing command implementation uses a nested for loop to tokenize single characters
+        // I fail to see the purpose of this when ' a' has a wholly different pronunciation than the start of ' apple'
+        const int n = whisper_tokenize(ctx, (" " + s).c_str(), tokens, 32);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, s.c_str());
+            return 3;
+        }
+        token_vec.push_back(tokens[0]);
+        if (!token_set.insert(tokens[0]).second) {
+            fprintf(stderr, "%s: warning: %s is a duplicate of an existing token\n", __func__, s.c_str());
+            throw json{
+                {"code",-31000},
+                {"message", "Duplicate token in token set: " + s}
+            };
+        }
+        if (n > 1) {// empty string if n=0? Should never occur
+            fprintf(stderr, "%s: error: command is more than a single token: %s\n", __func__, s.c_str());
+        }
+        struct command command = {token_vec, s};
+        cs.commands.push_back(command);
+        k_prompt += s;
+    }
+    k_prompt = k_prompt.substr(0,k_prompt.length()-2) + ". Selected word:";
+    cs.prompt_tokens.resize(1024);
+    int n = whisper_tokenize(ctx, k_prompt.c_str(), cs.prompt_tokens.data(), 1024);
+    cs.prompt_tokens.resize(n);
+    // prepare response
+    int index = commandset_list.size();
+    commandset_list.push_back(cs);
+    return json{{"index",index}};
+}
+json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
+    // whisper_state has the pertinent offsets, but there also seem to be a large
+    // number of scratch buffers that would prevent rewinding context in a manner similar to llama
+    // I'll give this a another pass once everything else is implemented,
+    // but for now, it's unsupported
+    throw json {
+        {"code", -32601},
+            {"message", "Seeking is not yet supported."}
+    };
+}
+json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
+    // See: https://www.jsonrpc.org/specification
+    json id = body.at("id");
+    try {
+        std::string version = body.at("jsonrpc");
+        if (version != "2.0") {
+            // unsupported version
+            throw json{
+                {"code", -3260},
+                {"message", "invalid jsonrpc version"}
+            };
+        }
+        std::string method = body.at("method");
+        json jparams = json{{"dummy", "dummy"}};
+        if (body.contains("params"))
+            jparams = body.at("params");
+        json res;
+        // TODO: be consistent about argument order
+        fprintf(stderr, "Dispatching a job\n");
+        if (method == "unguided")                { res = unguided_transcription(ctx, audio, jparams, params); }
+        else if (method == "guided")             { res = guided_transcription(ctx, audio, params, jparams, commandset_list); }
+        else if (method == "seek")               { res = seek(ctx, audio, jparams); }
+        else if (method == "registerCommandset") { res = register_commandset(ctx, jparams, commandset_list); }
+        else if (method == "echo")               { res = jparams; }
+
+
+        return json{
+            {"jsonrpc", "2.0"},
+                {"result", res},
+                {"id", id}
+        };
+    } catch(json ex) {
+        return json {
+            {"jsonrpc", "2.0"},
+                {"error", ex},
+                {"id", id}
+        };
+    }
+}
+
+void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    std::deque<json> jobqueue;
+    std::vector<struct commandset> commandset_list;
+    while (true) {
+        // For eventual cancellation support, shouldn't block if job exists
+        if (std::cin.rdbuf()->in_avail() > 22 || jobqueue.size() == 0) {
+            int content_length;
+            if (scanf("Content-Length: %d", &content_length) != 1) {
+                fprintf(stderr, "Could not read input: %d", std::cin.peek());
+                return;
+            }
+            // scanf leaves the new lines intact
+            std::cin.ignore(2);
+            if (std::cin.peek() != 13) {
+                // Content-Type. jsonrpc necessitates utf8.
+                std::cin.ignore(200,10);
+            }
+            std::cin.ignore(2);
+            // A message is being sent and blocking is acceptable
+            std::string content(content_length,'\0');
+            std::cin.read(&content[0], content_length);
+            json job = json::parse(content);
+            // TODO: Some messages(cancellation) should skip queue here
+            if (job.is_array()) {
+                // response must also be batched. Will implement later
+                // for (subjob : job.begin())
+                // TODO: At the very least respond with an unsupported error.
+            } else {
+                jobqueue.push_back(job);
+            }
+        }
+        assert(jobqueue.size() > 0);
+        json job = jobqueue.front();
+        json resp = parse_job(job, ctx, audio, params, commandset_list);
+        if (resp != "unfinished") {
+            jobqueue.pop_front();
+            // send response
+            std::string data = resp.dump(-1, ' ', false, json::error_handler_t::replace);
+            fprintf(stdout, "Content-Length: %d\r\n\r\n%s\n", (int)data.length()+1, data.c_str());
+            std::cout.flush();
+
+        }
+    }
+}
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+    // TODO: Investigate why this is required. An extra second of startup latency is not great
+    // wait for 1 second to avoid any buffered noise
+    std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+    audio.clear();
+    // TODO: consider some sort of indicator to designate loading has finished?
+    // Potentially better for the client to just start with a non-blocking message (register commands)
+    process_loop(ctx, audio, params);
+
+    audio.pause();
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}

examples/lsp/whisper.vim

@@ -0,0 +1,362 @@
+if !exists("g:whisper_dir")
+    let g:whisper_dir = expand($WHISPER_CPP_HOME)
+    if g:whisper_dir == ""
+        echoerr "Please provide a path to the whisper.cpp repo in either the $WHISPER_CPP_HOME environment variable, or g:whisper_dir"
+    endif
+endif
+if !exists("g:whisper_lsp_path")
+    let g:whisper_lsp_path = g:whisper_dir .. "lsp"
+    if !filereadable(g:whisper_lsp_path)
+        echoerr "Was not able to locate a lsp executable at: " .. g:whisper_lsp_path
+        throw "Executable not found"
+    endif
+endif
+if !exists("g:whisper_model_path")
+    " TODO: allow custom paths relative to the repo dir
+    let g:whisper_model_path = g:whisper_dir .. "models/ggml-base.en.bin"
+    if !filereadable(g:whisper_model_path)
+        echoerr "Could not find model at: " .. g:whisper_model_path
+        throw "Model not found"
+    endif
+endif
+let s:output_buffer = bufnr("whisper_log", v:true)
+call setbufvar(s:output_buffer,"&buftype","nofile")
+let s:lsp_command = [g:whisper_lsp_path,"-m",g:whisper_model_path]
+" For faster execution. TODO: server load multiple models/run multiple servers?
+" let s:lsp_command = [g:whisper_lsp_path, "-m", g:whisper_dir .. "models/ggml-tiny.en.bin", "-ac", "128"]
+
+" requestCommands([params_dict])
+func whisper#requestCommands(...)
+    let l:req = {"method": "guided", "params": {"commandset_index": 0}}
+    if a:0 > 0
+        call extend(l:req.params, a:1)
+    endif
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:commandCallback", [l:req.params, 0])})
+endfunction
+
+" doTranscription([params_dict])
+func whisper#doTranscription(...)
+    let l:req = {"method": "unguided", "params": {}}
+    if a:0 > 0
+        call extend(l:req.params, a:1)
+    endif
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:transcriptionCallback", [function("s:insertText"),function("s:endTranscription")])})
+endfunction
+
+" For testing
+func whisper#uppertest(cha)
+    echo tr(a:cha, s:c_lowerkeys, s:c_upperkeys)
+endfunction
+
+
+" (upper, exit, count, motion, command, insert/append, save run) "base"
+" (upper, exit, count, motion, command, inside/around)           "motion/visual"
+" (upper, exit, count, motion, line,    inside/around)           "command already entered"
+" (upper, exit, key,                                 )           "from/till"
+
+" upper and lower keys is used to translate between cases with tr
+" Must be sunchronized
+let s:c_lowerkeys = "1234567890-=qwertyuiop[]\\asdfghjkl;'zxcvbnm,./\""
+let s:c_upperkeys = "!@#$%^&*()_+QWERTYUIOP{}|ASDFGHJKL:\"ZXCVBNM<>?'"
+let s:c_count = split("1234567890\"",'\zs')
+let s:c_command = split("ryuogpdxcv.iam", '\zs')
+let s:c_motion = split("wetf'hjklnb$^)",'\zs')
+" object words: Word, Sentence, Paragraph, [, (, <, Tag, {. ", '
+let s:c_area = split("wsp])>t}\"'",'\zs')
+"Special commands.
+let s:c_special_always = ["exit", "upper"]
+let s:c_special_normal = ["save", "run", "space"]
+
+" If not in dict, key is spoken word,
+" If key resolves to string, value is used for normal/motion, but key for chars
+" If key resolves to dict, {0: "normal",1: "motion",2:"single char",3: "area"}
+" Missing entries fall back as follows {0: "required", 1: 0, 2: "key", 3: 0}
+let s:spoken_dict = {"w": "word", "e": "end", "r": "replace", "t": {0: "till", 3: "tag"}, "y": "yank", "u": "undo", "i": {0: "insert", 1: "inside"}, "o": "open", "p": {0: "paste", 3: "paragraph"},  "a": {0: "append", 1: "around"}, "s": {0: "substitute", 3: "sentence"}, "d": "delete", "f": "from", "g": "go", "h": "left", "j": "down", "k": "up", "l": "right", "c": "change", "v": "visual", "b": "back", "n": "next", "m": "mark", ".": {0: "repeat", 2: "period"}, "]": {0: "bracket", 2: "bracket"}, "'": {0: "jump", 2: "apostrophe", 3:  "apostrophe"}, '"': {0: 'register', 2: "quotation", 3: "quotation"}, "-": {0: "minus", 2: "minus"}, "$": {0: "dollar", 2: "dollar"}, "^": {0: "carrot", 2: "carrot"}, ")": {0: "sentence", 2: "parenthesis",  3: "parenthesis"}, "}": {0: "paragraph", 2: "brace", 3: "brace"}, ">": {0: "indent", 2: "angle", 3: "angle"}}
+
+" Give this another pass. This seems overly hacky even if it's functional
+let s:sub_tran_msg = ""
+func s:subTranProg(msg)
+    if s:sub_tran_msg != ""
+        let s:sub_tran_msg = s:sub_tran_msg .. a:msg
+        if mode() !=? 'v'
+            exe "normal" "u" .. s:sub_tran_msg
+        endif
+    else
+        if s:command_backlog == ""
+            " this should not occur
+            call s:logCallback(0, "Warning: Encountered sub transcription without prior command")
+            let s:command_backlog = "a"
+        endif
+        if a:msg[0] == ' '
+            let s:sub_tran_msg = s:command_backlog .. a:msg[1:-1]
+        else
+            let s:sub_tran_msg = s:command_backlog  .. a:msg
+        endif
+        if mode() !=? 'v'
+            exe "normal" s:sub_tran_msg
+        endif
+    endif
+    call appendbufline(s:output_buffer, "$", s:sub_tran_msg ..  ":" .. string(a:msg ))
+endfunction
+
+func s:subTranFinish(params, timestamp)
+    let s:repeat_command = s:sub_tran_msg
+    " Visual selection is lot if used with streaming, so streaming of partial
+    " transcriptions is disabled in visual mode
+    if mode() ==? 'v'
+        exe "normal" s:sub_tran_msg
+    endif
+    let s:sub_tran_msg = ""
+    let s:command_backlog = ""
+    exe "normal a\<C-G>u"
+    let l:params = a:params
+    let l:params.timestamp = a:timestamp
+    if exists("l:params.commandset_index")
+        unlet l:params.commandset_index
+    endif
+    call whisper#requestCommands(a:params)
+endfunction
+
+func s:logCallback(channel, msg)
+    call appendbufline(s:output_buffer,"$",a:msg)
+endfunction
+
+
+func s:transcriptionCallback(progressCallback, finishedCallback, channel, msg)
+    let l:tr = a:msg.result.transcription
+
+    let l:ex_ind = match(tolower(l:tr),"exit", len(l:tr)-6)
+    " The worst case I've observed so far is " Exit.", which is 6 characters
+    if l:ex_ind != -1
+        call a:progressCallback(strpart(l:tr,0,l:ex_ind-1))
+        call a:finishedCallback(a:msg.result.timestamp)
+    else
+        call a:progressCallback(l:tr)
+        let req = {"method": "unguided", "params": {"timestamp": a:msg.result.timestamp, "no_context": v:true}}
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [a:progressCallback, a:finishedCallback])})
+    endif
+endfunc
+func s:insertText(msg)
+    exe "normal a" .. a:msg
+endfunction
+func s:endTranscription(timestamp)
+    call appendbufline(s:output_buffer, "$", "Ending unguided transcription")
+endfunction
+
+
+
+" If a command does not include a whole actionable step, attempting to execute
+" it discards the remainder of things. There is likely a simpler solution,
+" but it can be made functional now by storing a backbuffer until actionable
+let s:command_backlog = ""
+let s:repeat_command = ""
+let s:preceeding_upper = v:false
+func s:commandCallback(params, commandset_index, channel, msg)
+    let l:command_index = a:msg.result.command_index
+    let l:do_execute = v:false
+    let l:next_mode = a:commandset_index
+    let l:command = s:commandset_list[a:commandset_index][l:command_index]
+    call s:logCallback(0, string(a:msg) .. " " .. a:commandset_index .. " " .. l:command)
+    if l:command_index == 0
+        "exit
+        "if s:command_backlog == ""
+        call s:logCallback(0,"Stopping command mode")
+        echo "No longer listening"
+        let s:command_backlog = ""
+        return
+        "else
+        " Legacy code to clear an existing buffer with exit.
+        " Was found to be rarely desired and is better introduced as a
+        " standalone command (clear?)
+        "   call s:logCallback(0,"Clearing command_backlog" .. s:command_backlog)
+        "   let s:command_backlog = ""
+        "   let s:preceeding_upper = v:false
+        " endif
+    elseif l:command_index == 1
+        " upper
+        let s:preceeding_upper = !s:preceeding_upper
+    elseif l:command == "save"
+        " save and run can only happen in commandset 0,
+        exe "w"
+    elseif l:command == "run"
+        exe "make run"
+    elseif l:command == "space"
+        exe "normal i \<ESC>l"
+    elseif has_key(s:c_user, l:command)
+        let Userfunc = s:c_user[l:command]
+        if type(Userfunc) == v:t_string
+            let Userfunc = function(Userfunc)
+        endif
+        call Userfunc()
+    else
+        if s:preceeding_upper
+            " Upper should keep commandset
+            let s:preceeding_upper = v:false
+            let l:visual_command = tr(l:command, s:c_lowerkeys, s:c_upperkeys)
+        else
+            let l:visual_command = l:command
+        endif
+        echo s:command_backlog .. " - " .. l:visual_command
+        let s:command_backlog = s:command_backlog .. l:visual_command
+        if a:commandset_index == 2 || a:commandset_index == 3
+            " single key, either completes motion, replace, or register
+            " Should move to execute unless part of a register
+            " Change will be caught at execute
+            if s:command_backlog[-2:-2] !=# '"'
+                call s:logCallback(0,"not register")
+                let l:do_execute = v:true
+            end
+            let l:next_mode = 0
+            " commandset index only matters for a/i
+        elseif (l:command == "a" || l:command == "i") && a:commandset_index == 1
+            " inside/around. Is commandset 3
+            let l:next_mode = 3
+        elseif l:command ==# '"'
+            let l:next_mode = 2
+        elseif index(s:c_count, l:command) != -1
+            let l:next_mode = a:commandset_index
+        elseif index(s:c_motion, l:command) != -1
+            if l:command == 't' || l:command == 'f' || l:command == "'"
+                " prompt single key
+                let l:next_mode = 2
+            else
+                let l:do_execute = v:true
+                let l:next_mode = 0
+            endif
+        elseif index(s:c_command, l:command) != -1
+            if index(["y","g","d","c"], s:command_backlog[-1:-1]) != -1 && s:command_backlog[-1:-1] != s:command_backlog[-2:-2] && mode() !=? 'v'
+                " need motion or repeated command
+                " Potential for bad state here if disparaging command keys are
+                " entered (i.e. yd), but vim can handle checks for this at exe
+                " And checking for cases like y123d would complicate things
+                let l:next_mode = 1
+            elseif index(["i","a","c", "o", "s"], l:command) != -1 || s:command_backlog[-1:-1] ==# 'R'
+                "'Insert' mode, do general transcription
+                let l:req = {"method": "unguided", "params": a:params}
+                let l:req.params.timestamp = a:msg.result.timestamp
+                let l:req.params.no_context = v:true
+                let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [function("s:subTranProg"), function("s:subTranFinish", [a:params])])})
+                return
+            elseif l:command == 'r' || l:command == 'm'
+                let l:next_mode = 2
+            elseif l:command == '.'
+                let l:next_mode = 0
+                let l:do_execute = v:true
+                let s:command_backlog = s:command_backlog[0:-2] .. s:repeat_command
+            else
+                if l:command ==? 'v'
+                    let l:next_mode = 1
+                else
+                    let l:next_mode = 0
+                endif
+                let l:do_execute = v:true
+            endif
+        else
+            throw "Invalid command state: " .. l:command .. " " .. a:commandset_index .. " " .. s:command_backlog
+        endif
+    endif
+    if l:do_execute
+        if mode() ==?'v' && l:next_mode == 0
+            let l:next_mode = 1
+        elseif match(s:command_backlog, 'c') != -1
+            let l:req = {"method": "unguided", "params": a:params}
+            let l:req.params.timestamp = a:msg.result.timestamp
+            let l:req.params.no_context = v:true
+            let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [function("s:subTranProg"), function("s:subTranFinish", [a:params])])})
+            return
+        endif
+        exe "normal" s:command_backlog
+        if index(s:c_motion + ["u"],l:command) == -1
+            exe "normal a\<C-G>u"
+            let s:repeat_command = s:command_backlog
+            call s:logCallback(0, s:command_backlog)
+        endif
+        let s:command_backlog = ""
+    endif
+    let l:req = {"method": "guided", "params": a:params}
+    let l:req.params.timestamp = a:msg.result.timestamp
+    let l:req.params.commandset_index = l:next_mode
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:commandCallback",[a:params, l:next_mode])})
+endfunction
+
+func s:loadedCallback(channel, msg)
+    echo "Loading complete"
+    call s:logCallback(a:channel, a:msg)
+endfunction
+
+func s:registerCommandset(commandlist, is_final)
+    let req = {"method": "registerCommandset"}
+    let req.params = a:commandlist
+    call s:logCallback(0, join(a:commandlist))
+    call add(g:whisper_commandlist_spoken, a:commandlist)
+    if a:is_final
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": "s:loadedCallback"})
+    else
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": "s:logCallback"})
+    endif
+endfunction
+
+func s:registerAllCommands()
+    let l:normal = s:c_special_always + s:c_special_normal + s:c_count + s:c_command + s:c_motion + keys(s:c_user)
+    let l:visual = s:c_special_always + s:c_count + s:c_command + s:c_motion
+    " Currently the same as visual.
+    " let l:post_command = s:c_special_always + s:c_count + s:c_command + s:c_motion
+    let l:single_key = s:c_special_always + split(s:c_lowerkeys, '\zs')
+    let l:area = s:c_special_always + s:c_area
+
+    " Used only for compatibility with the testing script
+    let g:whisper_commandlist_spoken = []
+
+    let s:commandset_list = [l:normal, l:visual, l:single_key, l:area]
+    call s:registerCommandset(s:commandsetToSpoken(l:normal, 0), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:visual, 1), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:single_key, 2), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:area, 3), v:true)
+endfunction
+
+func s:commandsetToSpoken(commandset, spoken_index)
+    let l:spoken_list = []
+    for l:command in a:commandset
+        if has_key(s:spoken_dict, l:command)
+            let l:spoken_value = s:spoken_dict[l:command]
+            if type(l:spoken_value) == v:t_dict
+                if has_key(l:spoken_value, a:spoken_index)
+                    let l:spoken_value = l:spoken_value[a:spoken_index]
+                else
+                    if a:spoken_index == 2
+                        let l:spoken_value = l:command
+                    else
+                        let l:spoken_value = l:spoken_value[0]
+                    endif
+                endif
+            else
+                if a:spoken_index == 2
+                    let l:spoken_value = l:command
+                endif
+            endif
+        else
+            let l:spoken_value = l:command
+        endif
+        call add(l:spoken_list, l:spoken_value)
+    endfor
+    return l:spoken_list
+endfunction
+
+" TODO: Check lifetime. If the script is resourced, is the existing
+" s:lsp_job dropped and therefore killed?
+" This seems to not be the case and I've had to deal with zombie processes
+" that survive exiting vim, even though said behavior conflicts with my
+" understanding of the provided documentation
+let s:lsp_opts = {"in_mode": "lsp", "out_mode": "lsp", "err_mode": "nl", "err_io": "buffer", "err_buf": s:output_buffer}
+if !exists("g:lsp_job")
+    if exists("g:whisper_user_commands")
+        let s:c_user = g:whisper_user_commands
+    else
+        let s:c_user = {}
+    endif
+    let g:lsp_job = job_start(s:lsp_command, s:lsp_opts)
+    if job_status(g:lsp_job) == "fail"
+        echoerr "Failed to start whisper job"
+    endif
+    call s:registerAllCommands()
+endif

examples/main/README.md

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

examples/main/main.cpp

@@ -59,16 +59,19 @@ struct whisper_params {
     int32_t offset_t_ms  =  0;
     int32_t offset_n     =  0;
     int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
     int32_t max_context  = -1;
     int32_t max_len      =  0;
-    int32_t best_of      =  2;
-    int32_t beam_size    = -1;
+    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t audio_ctx   = 0;
 
     float word_thold    =  0.01f;
     float entropy_thold =  2.40f;
     float logprob_thold = -1.00f;
 
     bool speed_up        = false;
+    bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
     bool diarize         = false;
@@ -81,11 +84,15 @@ struct whisper_params {
     bool output_wts      = false;
     bool output_csv      = false;
     bool output_jsn      = false;
+    bool output_jsn_full = false;
     bool output_lrc      = false;
+    bool no_prints       = false;
     bool print_special   = false;
     bool print_colors    = false;
     bool print_progress  = false;
     bool no_timestamps   = false;
+    bool log_score       = false;
+    bool use_gpu         = true;
 
     std::string language  = "en";
     std::string prompt;
@@ -130,10 +137,12 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
         else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
         else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-context")   { params.audio_ctx       = std::stoi(argv[++i]); }
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
+        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
         else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
@@ -147,7 +156,9 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
         else if (arg == "-ocsv" || arg == "--output-csv")      { params.output_csv      = true; }
         else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
+        else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
         else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
         else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
         else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
         else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
@@ -158,6 +169,8 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
         else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
         else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
+        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -184,10 +197,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
     fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
     fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
     fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
@@ -200,7 +215,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
     fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
     fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
+    fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
+    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
     fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
@@ -211,6 +228,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
     fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
+    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
     fprintf(stderr, "\n");
 }
 
@@ -218,6 +237,7 @@ struct whisper_print_user_data {
     const whisper_params * params;
 
     const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
 };
 
 std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
@@ -252,6 +272,14 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
 
     return speaker;
 }
+void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
+    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
+    if (progress >= *progress_prev + progress_step) {
+        *progress_prev += progress_step;
+        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
+    }
+}
 
 void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
@@ -476,7 +504,31 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+    std::ofstream fout(fname);
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    // fprintf(stderr,"segments: %d\n",n_segments);
+    for (int i = 0; i < n_segments; ++i) {
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        // fprintf(stderr,"tokens: %d\n",n_tokens);
+        for (int j = 0; j < n_tokens; j++) {
+            auto token = whisper_full_get_token_text(ctx, i, j);
+            auto probability = whisper_full_get_token_p(ctx, i, j);
+            fout << token << '\t' << probability << std::endl;
+            // fprintf(stderr,"token: %s %f\n",token,probability);
+	    }
+    }
+    return true;
+}
+
+bool output_json(
+             struct whisper_context * ctx,
+                         const char * fname,
+               const whisper_params & params,
+    std::vector<std::vector<float>>   pcmf32s,
+                               bool   full) {
     std::ofstream fout(fname);
     int indent = 0;
 
@@ -493,7 +545,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
     auto end_arr = [&](bool end) {
         indent--;
         doindent();
-        fout << (end ? "]\n" : "},\n");
+        fout << (end ? "]\n" : "],\n");
     };
 
     auto start_obj = [&](const char *name) {
@@ -534,12 +586,29 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
         end_value(end);
     };
 
+    auto value_f = [&](const char *name, const float val, bool end) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
     auto value_b = [&](const char *name, const bool val, bool end) {
         start_value(name);
         fout << (val ? "true" : "false");
         end_value(end);
     };
 
+    auto times_o = [&](int64_t t0, int64_t t1, bool end) {
+        start_obj("timestamps");
+        value_s("from", to_timestamp(t0, true).c_str(), false);
+        value_s("to", to_timestamp(t1, true).c_str(), true);
+        end_obj(false);
+        start_obj("offsets");
+        value_i("from", t0 * 10, false);
+        value_i("to", t1 * 10, true);
+        end_obj(end);
+    };
+
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
         return false;
@@ -585,15 +654,26 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
                 const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
                 start_obj(nullptr);
-                    start_obj("timestamps");
-                        value_s("from", to_timestamp(t0, true).c_str(), false);
-                        value_s("to", to_timestamp(t1, true).c_str(), true);
-                    end_obj(false);
-                    start_obj("offsets");
-                        value_i("from", t0 * 10, false);
-                        value_i("to", t1 * 10, true);
-                    end_obj(false);
-                    value_s("text", text, !params.diarize && !params.tinydiarize);
+                    times_o(t0, t1, false);
+                    value_s("text", text, !params.diarize && !params.tinydiarize && !full);
+
+                    if (full) {
+                        start_arr("tokens");
+                        const int n = whisper_full_n_tokens(ctx, i);
+                        for (int j = 0; j < n; ++j) {
+                            auto token = whisper_full_get_token_data(ctx, i, j);
+                            start_obj(nullptr);
+                                value_s("text", whisper_token_to_str(ctx, token.id), false);
+                                if(token.t0 > -1 && token.t1 > -1) {
+                                    // If we have per-token timestamps, write them out
+                                    times_o(token.t0, token.t1, false);
+                                }
+                                value_i("id", token.id, false);
+                                value_f("p", token.p, true);
+                            end_obj(j == (n - 1));
+                        }
+                        end_arr(!params.diarize && !params.tinydiarize);
+                    }
 
                     if (params.diarize && pcmf32s.size() == 2) {
                         value_s("speaker", estimate_diarization_speaker(pcmf32s, t0, t1, true).c_str(), true);
@@ -778,6 +858,9 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
+
+void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -804,9 +887,16 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
+    if (params.no_prints) {
+        whisper_log_set(cb_log_disable, NULL);
+    }
+
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+
+    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");
@@ -828,29 +918,28 @@ int main(int argc, char ** argv) {
             continue;
         }
 
-        // print system information
-        {
+        if (!whisper_is_multilingual(ctx)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        if (params.detect_language) {
+            params.language = "auto";
+        }
+
+        if (!params.no_prints) {
+            // print system information
             fprintf(stderr, "\n");
             fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
                     params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
-        }
 
-        // print some info about the processing
-        {
+            // print some info about the processing
             fprintf(stderr, "\n");
-            if (!whisper_is_multilingual(ctx)) {
-                if (params.language != "en" || params.translate) {
-                    params.language = "en";
-                    params.translate = false;
-                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-                }
-            }
-            if (params.detect_language) {
-                params.language = "auto";
-            }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
-                    params.n_threads, params.n_processors,
+                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
                     params.tinydiarize ? "tdrz = 1, " : "",
@@ -877,12 +966,14 @@ int main(int argc, char ** argv) {
             wparams.offset_ms        = params.offset_t_ms;
             wparams.duration_ms      = params.duration_ms;
 
-            wparams.token_timestamps = params.output_wts || params.max_len > 0;
+            wparams.token_timestamps = params.output_wts || params.output_jsn_full || params.max_len > 0;
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
             wparams.split_on_word    = params.split_on_word;
+            wparams.audio_ctx        = params.audio_ctx;
 
             wparams.speed_up         = params.speed_up;
+            wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 
@@ -895,7 +986,9 @@ int main(int argc, char ** argv) {
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
-            whisper_print_user_data user_data = { &params, &pcmf32s };
+            wparams.no_timestamps    = params.no_timestamps;
+
+            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
@@ -903,8 +996,14 @@ int main(int argc, char ** argv) {
                 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
+            if (wparams.print_progress) {
+                wparams.progress_callback           = whisper_print_progress_callback;
+                wparams.progress_callback_user_data = &user_data;
+            }
+
+            // examples for abort mechanism
+            // in examples below, 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
@@ -916,6 +1015,17 @@ int main(int argc, char ** argv) {
                 wparams.encoder_begin_callback_user_data = &is_aborted;
             }
 
+            // the callback is called before every computation - if it returns true, the computation is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.abort_callback = [](void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return is_aborted;
+                };
+                wparams.abort_callback_user_data = &is_aborted;
+            }
+
             if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                 return 10;
@@ -959,7 +1069,7 @@ int main(int argc, char ** argv) {
             // output to JSON file
             if (params.output_jsn) {
                 const auto fname_jsn = fname_out + ".json";
-                output_json(ctx, fname_jsn.c_str(), params, pcmf32s);
+                output_json(ctx, fname_jsn.c_str(), params, pcmf32s, params.output_jsn_full);
             }
 
             // output to LRC file
@@ -967,6 +1077,12 @@ int main(int argc, char ** argv) {
                 const auto fname_lrc = fname_out + ".lrc";
                 output_lrc(ctx, fname_lrc.c_str(), params, pcmf32s);
             }
+
+            // output to score file
+            if (params.log_score) {
+                const auto fname_score = fname_out + ".score.txt";
+                output_score(ctx, fname_score.c_str(), params, pcmf32s);
+            }
         }
     }
 

examples/python/test_whisper_processor.py

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

examples/python/whisper_processor.py

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

examples/quantize/quantize.cpp

@@ -138,7 +138,7 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f
         //    return false;
         //}
 
-        char word[128];
+        char word[129];
 
         for (int i = 0; i < n_vocab; i++) {
             uint32_t len;

examples/server/CMakeLists.txt

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

examples/server/README.md

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

examples/stream.wasm/emscripten.cpp

@@ -132,7 +132,7 @@ EMSCRIPTEN_BINDINGS(stream) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/stream/README.md

@@ -4,7 +4,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The `stream` tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
-```java
+```bash
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
@@ -14,7 +14,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 
 Setting the `--step` argument to `0` enables the sliding window mode:
 
-```java
+```bash
  ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
@@ -39,6 +39,20 @@ brew install sdl2
 make stream
 ```
 
+Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
+as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
+`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
+
+```bash
+whisper.cpp/examples/stream$ make stream
+g++     stream.cpp   -o stream
+stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
+    6 | #include "common/sdl.h"
+      |          ^~~~~~~~~~~~~~
+compilation terminated.
+make: *** [<builtin>: stream] Error 1
+```
+
 ## Web version
 
 This tool can also run in the browser: [examples/stream.wasm](/examples/stream.wasm)

examples/stream/stream.cpp

@@ -2,9 +2,8 @@
 //
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
 //
-
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
 
 #include <cassert>
@@ -14,6 +13,7 @@
 #include <vector>
 #include <fstream>
 
+
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
 std::string to_timestamp(int64_t t) {
@@ -47,6 +47,9 @@ struct whisper_params {
     bool print_special = false;
     bool no_context    = true;
     bool no_timestamps = false;
+    bool tinydiarize   = false;
+    bool save_audio    = false; // save audio to wav file
+    bool use_gpu       = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -63,23 +66,27 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (                 arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
-        else if (                 arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
-        else if (                 arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
-        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 == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
-        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
-        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-t"    || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (                  arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
+        else if (                  arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
+        else if (                  arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
+        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 == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
+        else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-kc"   || arg == "--keep-context")  { params.no_context    = false; }
+        else if (arg == "-l"    || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"    || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")   { params.tinydiarize   = true; }
+        else if (arg == "-sa"   || arg == "--save-audio")    { params.save_audio    = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")        { params.use_gpu       = false; }
+
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -113,6 +120,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
     fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
+    fprintf(stderr, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
     fprintf(stderr, "\n");
 }
 
@@ -150,14 +160,16 @@ int main(int argc, char ** argv) {
     audio.resume();
 
     // whisper init
-
     if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
         fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
         whisper_print_usage(argc, argv, params);
         exit(0);
     }
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     std::vector<float> pcmf32    (n_samples_30s, 0.0f);
     std::vector<float> pcmf32_old;
@@ -208,14 +220,28 @@ int main(int argc, char ** argv) {
         }
     }
 
-    printf("[Start speaking]");
+    wav_writer wavWriter;
+    // save wav file
+    if (params.save_audio) {
+        // Get current date/time for filename
+        time_t now = time(0);
+        char buffer[80];
+        strftime(buffer, sizeof(buffer), "%Y%m%d%H%M%S", localtime(&now));
+        std::string filename = std::string(buffer) + ".wav";
+
+        wavWriter.open(filename, WHISPER_SAMPLE_RATE, 16, 1);
+    }
+    printf("[Start speaking]\n");
     fflush(stdout);
 
-          auto t_last  = std::chrono::high_resolution_clock::now();
+    auto t_last  = std::chrono::high_resolution_clock::now();
     const auto t_start = t_last;
 
     // main audio loop
     while (is_running) {
+        if (params.save_audio) {
+            wavWriter.write(pcmf32_new.data(), pcmf32_new.size());
+        }
         // handle Ctrl + C
         is_running = sdl_poll_events();
 
@@ -299,6 +325,8 @@ int main(int argc, char ** argv) {
             wparams.audio_ctx        = params.audio_ctx;
             wparams.speed_up         = params.speed_up;
 
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
+
             // disable temperature fallback
             //wparams.temperature_inc  = -1.0f;
             wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
@@ -344,10 +372,19 @@ int main(int argc, char ** argv) {
                         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-                        printf ("[%s --> %s]  %s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), text);
+                        std::string output = "[" + to_timestamp(t0) + " --> " + to_timestamp(t1) + "]  " + text;
+
+                        if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                            output += " [SPEAKER_TURN]";
+                        }
+
+                        output += "\n";
+
+                        printf("%s", output.c_str());
+                        fflush(stdout);
 
                         if (params.fname_out.length() > 0) {
-                            fout << "[" << to_timestamp(t0) << " --> " << to_timestamp(t1) << "]  " << text << std::endl;
+                            fout << output;
                         }
                     }
                 }
@@ -356,7 +393,7 @@ int main(int argc, char ** argv) {
                     fout << std::endl;
                 }
 
-                if (use_vad){
+                if (use_vad) {
                     printf("\n");
                     printf("### Transcription %d END\n", n_iter);
                 }

examples/talk-llama/CMakeLists.txt

@@ -1,16 +1,18 @@
 if (WHISPER_SDL2)
     # talk-llama
     set(TARGET talk-llama)
-    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
-    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
 
-    # TODO: this is temporary
-    #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+    if (WHISPER_CLBLAST)
+        set(CLBLAST_LIBNAME clblast)
+    endif ()
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CLBLAST_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    if(WIN32)
+        # It requires Windows 8.1 or later for PrefetchVirtualMemory
+        target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
+    endif()
 
     include(DefaultTargetOptions)
 endif ()

examples/talk-llama/README.md

@@ -2,6 +2,12 @@
 
 Talk with an LLaMA AI in your terminal
 
+*Latest perf as of 2 Nov 2023 using Whisper Medium + LLaMA v2 13B Q8_0 on M2 Ultra:*
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-60c6b391649e
+
+*Previous demo running on CPUs*
+
 [Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
 
 ## Building
@@ -19,7 +25,7 @@ brew install sdl2
 make talk-llama
 
 # Run it
-./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
 ```
 
 - The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
@@ -36,7 +42,7 @@ This feature is especially helpful for maintaining context in long conversations
 Example usage:
 
 ```bash
-./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
 ```
 
 ## TTS

examples/talk-llama/llama-util.h

@@ -1,474 +0,0 @@
-// Internal header to be included only by llama.cpp.
-// Contains wrappers around OS interfaces.
-
-#ifndef LLAMA_UTIL_H
-#define LLAMA_UTIL_H
-
-#include <cstdio>
-#include <cstdint>
-#include <cerrno>
-#include <cstring>
-#include <cstdarg>
-#include <cstdlib>
-#include <climits>
-
-#include <string>
-#include <vector>
-#include <stdexcept>
-
-#ifdef __has_include
-    #if __has_include(<unistd.h>)
-        #include <unistd.h>
-        #if defined(_POSIX_MAPPED_FILES)
-            #include <sys/mman.h>
-        #endif
-        #if defined(_POSIX_MEMLOCK_RANGE)
-            #include <sys/resource.h>
-        #endif
-    #endif
-#endif
-
-#if defined(_WIN32)
-    #define WIN32_LEAN_AND_MEAN
-    #ifndef NOMINMAX
-        #define NOMINMAX
-    #endif
-    #include <windows.h>
-    #include <io.h>
-    #include <stdio.h> // for _fseeki64
-#endif
-
-#define LLAMA_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            abort(); \
-        } \
-    } while (0)
-
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    LLAMA_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
-struct llama_file {
-    // use FILE * so we don't have to re-open the file to mmap
-    FILE * fp;
-    size_t size;
-
-    llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
-        if (fp == NULL) {
-            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
-        }
-        seek(0, SEEK_END);
-        size = tell();
-        seek(0, SEEK_SET);
-    }
-
-    size_t tell() const {
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
-        return (size_t) ret;
-    }
-
-    void seek(size_t offset, int whence) {
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        LLAMA_ASSERT(ret == 0); // same
-    }
-
-    void read_raw(void * ptr, size_t len) const {
-        if (len == 0) {
-            return;
-        }
-        errno = 0;
-        std::size_t ret = std::fread(ptr, len, 1, fp);
-        if (ferror(fp)) {
-            throw std::runtime_error(format("read error: %s", strerror(errno)));
-        }
-        if (ret != 1) {
-            throw std::runtime_error(std::string("unexpectedly reached end of file"));
-        }
-    }
-
-    std::uint32_t read_u32() {
-        std::uint32_t ret;
-        read_raw(&ret, sizeof(ret));
-        return ret;
-    }
-
-    std::string read_string(std::uint32_t len) {
-        std::vector<char> chars(len);
-        read_raw(chars.data(), len);
-        return std::string(chars.data(), len);
-    }
-
-    void write_raw(const void * ptr, size_t len) const {
-        if (len == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, len, 1, fp);
-        if (ret != 1) {
-            throw std::runtime_error(format("write error: %s", strerror(errno)));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
-    ~llama_file() {
-        if (fp) {
-            std::fclose(fp);
-        }
-    }
-};
-
-#if defined(_WIN32)
-static std::string llama_format_win_err(DWORD err) {
-    LPSTR buf;
-    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
-                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
-    if (!size) {
-        return "FormatMessageA failed";
-    }
-    std::string ret(buf, size);
-    LocalFree(buf);
-    return ret;
-}
-#endif
-
-struct llama_mmap {
-    void * addr;
-    size_t size;
-
-    llama_mmap(const llama_mmap &) = delete;
-
-#ifdef _POSIX_MAPPED_FILES
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
-        size = file->size;
-        int fd = fileno(file->fp);
-        int flags = MAP_SHARED;
-#ifdef __linux__
-        flags |= MAP_POPULATE;
-#endif
-        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        if (addr == MAP_FAILED) {
-            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
-        }
-
-        if (prefetch > 0) {
-            // Advise the kernel to preload the mapped memory
-            if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
-                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
-                        strerror(errno));
-            }
-        }
-    }
-
-    ~llama_mmap() {
-        munmap(addr, size);
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
-        size = file->size;
-
-        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
-
-        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
-        DWORD error = GetLastError();
-
-        if (hMapping == NULL) {
-            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
-        }
-
-        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
-        error = GetLastError();
-        CloseHandle(hMapping);
-
-        if (addr == NULL) {
-            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
-        }
-
-        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
-        if (prefetch) {
-            // Advise the kernel to preload the mapped memory
-            WIN32_MEMORY_RANGE_ENTRY range;
-            range.VirtualAddress = addr;
-            range.NumberOfBytes = (SIZE_T)size;
-            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-            }
-        }
-        #else
-        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
-        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
-    }
-
-    ~llama_mmap() {
-        if (!UnmapViewOfFile(addr)) {
-            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    llama_mmap(struct llama_file *, bool prefetch = true) {
-        (void)prefetch;
-        throw std::runtime_error(std::string("mmap not supported"));
-    }
-#endif
-};
-
-// Represents some region of memory being locked using mlock or VirtualLock;
-// will automatically unlock on destruction.
-struct llama_mlock {
-    void * addr = NULL;
-    size_t size = 0;
-    bool failed_already = false;
-
-    llama_mlock() {}
-    llama_mlock(const llama_mlock &) = delete;
-
-    ~llama_mlock() {
-        if (size) {
-            raw_unlock(addr, size);
-        }
-    }
-
-    void init(void * ptr) {
-        LLAMA_ASSERT(addr == NULL && size == 0);
-        addr = ptr;
-    }
-
-    void grow_to(size_t target_size) {
-        LLAMA_ASSERT(addr);
-        if (failed_already) {
-            return;
-        }
-        size_t granularity = lock_granularity();
-        target_size = (target_size + granularity - 1) & ~(granularity - 1);
-        if (target_size > size) {
-            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
-                size = target_size;
-            } else {
-                failed_already = true;
-            }
-        }
-    }
-
-#ifdef _POSIX_MEMLOCK_RANGE
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        return (size_t) sysconf(_SC_PAGESIZE);
-    }
-
-    #ifdef __APPLE__
-        #define MLOCK_SUGGESTION \
-            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
-    #else
-        #define MLOCK_SUGGESTION \
-            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
-    #endif
-
-    bool raw_lock(const void * addr, size_t size) {
-        if (!mlock(addr, size)) {
-            return true;
-        } else {
-            char* errmsg = std::strerror(errno);
-            bool suggest = (errno == ENOMEM);
-
-            // Check if the resource limit is fine after all
-            struct rlimit lock_limit;
-            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
-                suggest = false;
-            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
-                suggest = false;
-
-            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
-                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
-            return false;
-        }
-    }
-
-    #undef MLOCK_SUGGESTION
-
-    void raw_unlock(void * addr, size_t size) {
-        if (munlock(addr, size)) {
-            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
-        }
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        SYSTEM_INFO si;
-        GetSystemInfo(&si);
-        return (size_t) si.dwPageSize;
-    }
-
-    bool raw_lock(void * ptr, size_t len) {
-        for (int tries = 1; ; tries++) {
-            if (VirtualLock(ptr, len)) {
-                return true;
-            }
-            if (tries == 2) {
-                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
-                    len, size, llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-
-            // It failed but this was only the first try; increase the working
-            // set size and try again.
-            SIZE_T min_ws_size, max_ws_size;
-            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
-                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-            // Per MSDN: "The maximum number of pages that a process can lock
-            // is equal to the number of pages in its minimum working set minus
-            // a small overhead."
-            // Hopefully a megabyte is enough overhead:
-            size_t increment = len + 1048576;
-            // The minimum must be <= the maximum, so we need to increase both:
-            min_ws_size += increment;
-            max_ws_size += increment;
-            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
-                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-        }
-    }
-
-    void raw_unlock(void * ptr, size_t len) {
-        if (!VirtualUnlock(ptr, len)) {
-            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    size_t lock_granularity() {
-        return (size_t) 65536;
-    }
-
-    bool raw_lock(const void * addr, size_t len) {
-        fprintf(stderr, "warning: mlock not supported on this system\n");
-        return false;
-    }
-
-    void raw_unlock(const void * addr, size_t len) {}
-#endif
-};
-
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct llama_buffer {
-    uint8_t * addr = NULL;
-    size_t size = 0;
-
-    llama_buffer() = default;
-
-    void resize(size_t len) {
-        delete[] addr;
-        addr = new uint8_t[len];
-        size = len;
-    }
-
-    ~llama_buffer() {
-        delete[] addr;
-    }
-
-    // disable copy and move
-    llama_buffer(const llama_buffer&) = delete;
-    llama_buffer(llama_buffer&&) = delete;
-    llama_buffer& operator=(const llama_buffer&) = delete;
-    llama_buffer& operator=(llama_buffer&&) = delete;
-};
-
-#ifdef GGML_USE_CUBLAS
-#include "ggml-cuda.h"
-struct llama_ctx_buffer {
-    uint8_t * addr = NULL;
-    bool is_cuda;
-    size_t size = 0;
-
-    llama_ctx_buffer() = default;
-
-    void resize(size_t size) {
-        free();
-
-        addr = (uint8_t *) ggml_cuda_host_malloc(size);
-        if (addr) {
-            is_cuda = true;
-        }
-        else {
-            // fall back to pageable memory
-            addr = new uint8_t[size];
-            is_cuda = false;
-        }
-        this->size = size;
-    }
-
-    void free() {
-        if (addr) {
-            if (is_cuda) {
-                ggml_cuda_host_free(addr);
-            }
-            else {
-                delete[] addr;
-            }
-        }
-        addr = NULL;
-    }
-
-    ~llama_ctx_buffer() {
-        free();
-    }
-
-    // disable copy and move
-    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
-    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
-    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
-    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
-};
-#else
-typedef llama_buffer llama_ctx_buffer;
-#endif
-
-#endif

examples/talk-llama/llama.h

@@ -1,8 +1,12 @@
 #ifndef LLAMA_H
 #define LLAMA_H
 
+#include "ggml.h"
+#include "ggml-backend.h"
+
 #include <stddef.h>
 #include <stdint.h>
+#include <stdio.h>
 #include <stdbool.h>
 
 #ifdef LLAMA_SHARED
@@ -19,17 +23,23 @@
 #    define LLAMA_API
 #endif
 
-#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
-#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
-#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
-#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
-#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
+#ifdef __GNUC__
+#    define DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define DEPRECATED(func, hint) func
+#endif
 
-#define LLAMA_FILE_VERSION           3
-#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
-#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
-#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION        1
+#define LLAMA_DEFAULT_SEED 0xFFFFFFFF
+
+#define LLAMA_MAX_RNG_STATE (64*1024)
+
+#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
+#define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
+
+#define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
+#define LLAMA_SESSION_VERSION 4
 
 #ifdef __cplusplus
 extern "C" {
@@ -41,9 +51,72 @@ extern "C" {
     // TODO: show sample usage
     //
 
+    struct llama_model;
     struct llama_context;
 
-    typedef int llama_token;
+    typedef int32_t llama_pos;
+    typedef int32_t llama_token;
+    typedef int32_t llama_seq_id;
+
+    enum llama_vocab_type {
+        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
+    };
+
+    enum llama_token_type {
+        LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
+        LLAMA_TOKEN_TYPE_NORMAL       = 1,
+        LLAMA_TOKEN_TYPE_UNKNOWN      = 2,
+        LLAMA_TOKEN_TYPE_CONTROL      = 3,
+        LLAMA_TOKEN_TYPE_USER_DEFINED = 4,
+        LLAMA_TOKEN_TYPE_UNUSED       = 5,
+        LLAMA_TOKEN_TYPE_BYTE         = 6,
+    };
+
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32              = 0,
+        LLAMA_FTYPE_MOSTLY_F16           = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,  // tok_embeddings.weight and output.weight are F16
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5,  // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6,  // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K          = 10, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_S        = 11, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_M        = 12, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_L        = 13, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_S        = 14, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_K_M        = 15, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
+
+        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
+    };
+
+    enum llama_rope_scaling_type {
+        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_NONE        = 0,
+        LLAMA_ROPE_SCALING_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_YARN        = 2,
+        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
+    };
+
+    enum llama_split_mode {
+        LLAMA_SPLIT_NONE    = 0, // single GPU
+        LLAMA_SPLIT_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_ROW     = 2, // split rows across GPUs
+    };
 
     typedef struct llama_token_data {
         llama_token id; // token id
@@ -57,70 +130,258 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);
 
-    struct llama_context_params {
-        int n_ctx;        // text context
-        int n_gpu_layers; // number of layers to store in VRAM
-        int seed;         // RNG seed, -1 for random
+    // Input data for llama_decode
+    // A llama_batch object can contain input about one or many sequences
+    // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
+    //
+    // - token  : the token ids of the input (used when embd is NULL)
+    // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
+    // - pos    : the positions of the respective token in the sequence
+    // - seq_id : the sequence to which the respective token belongs
+    // - logits : if zero, the logits for the respective token will not be output
+    //
+    typedef struct llama_batch {
+        int32_t n_tokens;
 
-        bool f16_kv;     // use fp16 for KV cache
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        llama_token  *  token;
+        float        *  embd;
+        llama_pos    *  pos;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
+        int8_t       *  logits;
+
+        // NOTE: helpers for smooth API transition - can be deprecated in the future
+        //       for future-proof code, use the above fields instead and ignore everything below
+        //
+        // pos[i] = all_pos_0 + i*all_pos_1
+        //
+        llama_pos    all_pos_0;  // used if pos == NULL
+        llama_pos    all_pos_1;  // used if pos == NULL
+        llama_seq_id all_seq_id; // used if seq_id == NULL
+    } llama_batch;
+
+    enum llama_model_kv_override_type {
+        LLAMA_KV_OVERRIDE_INT,
+        LLAMA_KV_OVERRIDE_FLOAT,
+        LLAMA_KV_OVERRIDE_BOOL,
+    };
+
+    struct llama_model_kv_override {
+        char key[128];
+        enum llama_model_kv_override_type tag;
+        union {
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
+        };
+    };
+
+    struct llama_model_params {
+        int32_t n_gpu_layers; // number of layers to store in VRAM
+        enum llama_split_mode split_mode; // how to split the model across multiple GPUs
+
+        // main_gpu interpretation depends on split_mode:
+        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_LAYER: ignored
+        int32_t main_gpu;
+
+        // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
+        const float * tensor_split;
+
+        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+        // If the provided progress_callback returns true, model loading continues.
+        // If it returns false, model loading is immediately aborted.
+        llama_progress_callback progress_callback;
+
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+
+        // override key-value pairs of the model meta data
+        const struct llama_model_kv_override * kv_overrides;
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
-        bool embedding;  // embedding mode only
-
-        // called with a progress value between 0 and 1, pass NULL to disable
-        llama_progress_callback progress_callback;
-        // context pointer passed to the progress callback
-        void * progress_callback_user_data;
     };
 
-    // model file types
-    enum llama_ftype {
-        LLAMA_FTYPE_ALL_F32              = 0,
-        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
+    struct llama_context_params {
+        uint32_t seed;              // RNG seed, -1 for random
+        uint32_t n_ctx;             // text context, 0 = from model
+        uint32_t n_batch;           // prompt processing maximum batch size
+        uint32_t n_threads;         // number of threads to use for generation
+        uint32_t n_threads_batch;   // number of threads to use for batch processing
+        int32_t  rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
+
+        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        float    rope_freq_base;   // RoPE base frequency, 0 = from model
+        float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, negative = from model
+        float    yarn_attn_factor; // YaRN magnitude scaling factor
+        float    yarn_beta_fast;   // YaRN low correction dim
+        float    yarn_beta_slow;   // YaRN high correction dim
+        uint32_t yarn_orig_ctx;    // YaRN original context size
+
+        ggml_backend_sched_eval_callback cb_eval;
+        void * cb_eval_user_data;
+
+        enum ggml_type type_k; // data type for K cache
+        enum ggml_type type_v; // data type for V cache
+
+        // Keep the booleans together to avoid misalignment during copy-by-value.
+        bool mul_mat_q;   // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
+        bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
+        bool embedding;   // embedding mode only
+        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
     };
 
-    LLAMA_API struct llama_context_params llama_context_default_params();
+    // model quantization parameters
+    typedef struct llama_model_quantize_params {
+        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;      // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
+        void * imatrix;              // pointer to importance matrix data
+    } llama_model_quantize_params;
 
-    LLAMA_API bool llama_mmap_supported();
-    LLAMA_API bool llama_mlock_supported();
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
+    // performance timing information
+    struct llama_timings {
+        double t_start_ms;
+        double t_end_ms;
+        double t_load_ms;
+        double t_sample_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_sample;
+        int32_t n_p_eval;
+        int32_t n_eval;
+    };
+
+    // Helpers for getting default parameters
+    LLAMA_API struct llama_model_params llama_model_default_params(void);
+    LLAMA_API struct llama_context_params llama_context_default_params(void);
+    LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params(void);
 
-    // TODO: not great API - very likely to change
     // Initialize the llama + ggml backend
+    // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_init_backend();
+    LLAMA_API void llama_backend_init(bool numa);
 
-    LLAMA_API int64_t llama_time_us();
+    // Call once at the end of the program - currently only used for MPI
+    LLAMA_API void llama_backend_free(void);
 
-    // Various functions for loading a ggml llama model.
-    // Allocate (almost) all memory needed for the model.
-    // Return NULL on failure
-    LLAMA_API struct llama_context * llama_init_from_file(
+    LLAMA_API struct llama_model * llama_load_model_from_file(
                              const char * path_model,
+            struct llama_model_params     params);
+
+    LLAMA_API void llama_free_model(struct llama_model * model);
+
+    LLAMA_API struct llama_context * llama_new_context_with_model(
+                     struct llama_model * model,
             struct llama_context_params   params);
 
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
-    // TODO: not great API - very likely to change
+    LLAMA_API int64_t llama_time_us(void);
+
+    LLAMA_API size_t llama_max_devices(void);
+
+    LLAMA_API bool llama_supports_mmap       (void);
+    LLAMA_API bool llama_supports_mlock      (void);
+    LLAMA_API bool llama_supports_gpu_offload(void);
+
+    LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead");
+    LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead");
+
+    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
+
+    LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
+
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+
+    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
+
+    // Get the model's RoPE frequency scaling factor
+    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
+
+    // Functions to access the model's GGUF metadata scalar values
+    // - The functions return the length of the string on success, or -1 on failure
+    // - The output string is always null-terminated and cleared on failure
+    // - GGUF array values are not supported by these functions
+
+    // Get metadata value as a string by key name
+    LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size);
+
+    // Get the number of metadata key/value pairs
+    LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
+
+    // Get metadata key name by index
+    LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
+
+    // Get metadata value as a string by index
+    LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
+
+    // Get a string describing the model type
+    LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+
+    // Returns the total size of all the tensors in the model in bytes
+    LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
+
+    // Returns the total number of parameters in the model
+    LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
+
+    // Get a llama model tensor
+    LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
+
     // Returns 0 on success
-    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
-    LLAMA_API int llama_model_quantize(
+    LLAMA_API uint32_t llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-      enum llama_ftype   ftype,
-            int          nthread);
+            const llama_model_quantize_params * params);
 
     // Apply a LoRA adapter to a loaded model
     // path_base_model is the path to a higher quality model to use as a base for
@@ -128,17 +389,135 @@ extern "C" {
     // The model needs to be reloaded before applying a new adapter, otherwise the adapter
     // will be applied on top of the previous one
     // Returns 0 on success
-    LLAMA_API int llama_apply_lora_from_file(
+    LLAMA_API DEPRECATED(int32_t llama_apply_lora_from_file(
             struct llama_context * ctx,
                       const char * path_lora,
+                           float   scale,
                       const char * path_base_model,
-                             int   n_threads);
+                         int32_t   n_threads),
+            "use llama_model_apply_lora_from_file instead");
 
-    // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_model_apply_lora_from_file(
+            const struct llama_model * model,
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                         int32_t   n_threads);
 
-    // Sets the current rng seed.
-    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
+    //
+    // KV cache
+    //
+
+    // Information associated with an individual cell in the KV cache view.
+    struct llama_kv_cache_view_cell {
+        // The position for this cell. Takes KV cache shifts into account.
+        // May be negative if the cell is not populated.
+        llama_pos pos;
+    };
+
+    // An updateable view of the KV cache.
+    struct llama_kv_cache_view {
+        // Number of KV cache cells. This will be the same as the context size.
+        int32_t n_cells;
+
+        // Maximum number of sequences that can exist in a cell. It's not an error
+        // if there are more sequences in a cell than this value, however they will
+        // not be visible in the view cells_sequences.
+        int32_t n_max_seq;
+
+        // Number of tokens in the cache. For example, if there are two populated
+        // cells, the first with 1 sequence id in it and the second with 2 sequence
+        // ids then you'll have 3 tokens.
+        int32_t token_count;
+
+        // Number of populated cache cells.
+        int32_t used_cells;
+
+        // Maximum contiguous empty slots in the cache.
+        int32_t max_contiguous;
+
+        // Index to the start of the max_contiguous slot range. Can be negative
+        // when cache is full.
+        int32_t max_contiguous_idx;
+
+        // Information for an individual cell.
+        struct llama_kv_cache_view_cell * cells;
+
+        // The sequences for each cell. There will be n_max_seq items per cell.
+        llama_seq_id * cells_sequences;
+    };
+
+    // Create an empty KV cache view. (use only for debugging purposes)
+    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq);
+
+    // Free a KV cache view. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
+
+    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
+
+    // Returns the number of tokens in the KV cache (slow, use only for debug)
+    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
+    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
+
+    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
+    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
+
+    // Clear the KV cache
+    LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_shift(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta);
+
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
+    //
+    // State / sessions
+    //
 
     // Returns the maximum size in bytes of the state (rng, logits, embedding
     // and kv_cache) - will often be smaller after compacting tokens
@@ -147,85 +526,276 @@ extern "C" {
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
     // Returns the number of bytes copied
-    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
+    LLAMA_API size_t llama_copy_state_data(
+            struct llama_context * ctx,
+                         uint8_t * dst);
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(
+            struct llama_context * ctx,
+                         uint8_t * src);
 
     // Save/load session file
-    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
-    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    LLAMA_API bool llama_load_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+                     llama_token * tokens_out,
+                          size_t   n_token_capacity,
+                          size_t * n_token_count_out);
 
-    // Run the llama inference to obtain the logits and probabilities for the next token.
+    LLAMA_API bool llama_save_session_file(
+            struct llama_context * ctx,
+                      const char * path_session,
+               const llama_token * tokens,
+                          size_t   n_token_count);
+
+    //
+    // Decoding
+    //
+
+    // Run the llama inference to obtain the logits and probabilities for the next token(s).
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls
     // Returns 0 on success
-    LLAMA_API int llama_eval(
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval(
             struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
-
-    // Convert the provided text into tokens.
-    // The tokens pointer must be large enough to hold the resulting tokens.
-    // Returns the number of tokens on success, no more than n_max_tokens
-    // Returns a negative number on failure - the number of tokens that would have been returned
-    // TODO: not sure if correct
-    LLAMA_API int llama_tokenize(
-            struct llama_context * ctx,
-                      const char * text,
                      llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
+                         int32_t   n_tokens,
+                         int32_t   n_past),
+            "use llama_decode() instead");
 
-    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
+    // Same as llama_eval, but use float matrix input directly.
+    // DEPRECATED: use llama_decode() instead
+    LLAMA_API DEPRECATED(int llama_eval_embd(
+            struct llama_context * ctx,
+                           float * embd,
+                         int32_t   n_tokens,
+                         int32_t   n_past),
+            "use llama_decode() instead");
+
+    // Return batch for single sequence of tokens starting at pos_0
+    //
+    // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
+    //
+    LLAMA_API struct llama_batch llama_batch_get_one(
+                  llama_token * tokens,
+                      int32_t   n_tokens,
+                    llama_pos   pos_0,
+                 llama_seq_id   seq_id);
+
+    // Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
+    // Each token can be assigned up to n_seq_max sequence ids
+    // The batch has to be freed with llama_batch_free()
+    // If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
+    // Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
+    // The rest of the llama_batch members are allocated with size n_tokens
+    // All members are left uninitialized
+    LLAMA_API struct llama_batch llama_batch_init(
+            int32_t n_tokens,
+            int32_t embd,
+            int32_t n_seq_max);
+
+    // Frees a batch of tokens allocated with llama_batch_init()
+    LLAMA_API void llama_batch_free(struct llama_batch batch);
+
+    // Positive return values does not mean a fatal error, but rather a warning.
+    //   0 - success
+    //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
+    // < 0 - error
+    LLAMA_API int32_t llama_decode(
+            struct llama_context * ctx,
+              struct llama_batch   batch);
+
+    // Set the number of threads used for decoding
+    // n_threads is the number of threads used for generation (single token)
+    // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
+    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
     // Cols: n_vocab
     LLAMA_API float * llama_get_logits(struct llama_context * ctx);
 
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
+    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
+
     // Get the embeddings for the input
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
-    // Token Id -> String. Uses the vocabulary in the provided context
-    LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
+    //
+    // Vocab
+    //
+
+    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
+
+    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
+
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos();
-    LLAMA_API llama_token llama_token_eos();
-    LLAMA_API llama_token llama_token_nl();
+    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
+    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
+    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
 
+    // Returns -1 if unknown, 1 for true or 0 for false.
+    LLAMA_API int32_t         llama_add_bos_token(const struct llama_model * model);
+
+    // Returns -1 if unknown, 1 for true or 0 for false.
+    LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);
+
+    // codellama infill tokens
+    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
+    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
+    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
+    LLAMA_API llama_token llama_token_eot   (const struct llama_model * model); // End of infill middle
+
+    //
+    // Tokenization
+    //
+
+    /// @details Convert the provided text into tokens.
+    /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
+    /// @return Returns the number of tokens on success, no more than n_max_tokens
+    /// @return Returns a negative number on failure - the number of tokens that would have been returned
+    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
+    ///                Does not insert a leading space.
+    LLAMA_API int32_t llama_tokenize(
+        const struct llama_model * model,
+                      const char * text,
+                         int32_t   text_len,
+                     llama_token * tokens,
+                         int32_t   n_max_tokens,
+                            bool   add_bos,
+                            bool   special);
+
+    // Token Id -> Piece.
+    // Uses the vocabulary in the provided context.
+    // Does not write null terminator to the buffer.
+    // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
+    LLAMA_API int32_t llama_token_to_piece(
+              const struct llama_model * model,
+                           llama_token   token,
+                                  char * buf,
+                               int32_t   length);
+
+    //
+    // Grammar
+    //
+
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
+    LLAMA_API struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar);
+
+    //
     // Sampling functions
+    //
+
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
-    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
-
     /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
-    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+    LLAMA_API void llama_sample_repetition_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   penalty_last_n,
+                           float   penalty_repeat,
+                           float   penalty_freq,
+                           float   penalty_present);
+
+    /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
+    /// @param logits Logits extracted from the original generation context.
+    /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_apply_guidance(
+              struct llama_context * ctx,
+                             float * logits,
+                             float * logits_guidance,
+                             float   scale);
+
+    LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance(
+              struct llama_context * ctx,
+            llama_token_data_array * candidates,
+              struct llama_context * guidance_ctx,
+                             float   scale),
+              "use llama_sample_apply_guidance() instead");
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API void llama_sample_softmax(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
+    LLAMA_API void llama_sample_top_k(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                         int32_t   k,
+                          size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
-    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
+    LLAMA_API void llama_sample_top_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    LLAMA_API void llama_sample_min_p(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
 
     /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
-    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
+    LLAMA_API void llama_sample_tail_free(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   z,
+                          size_t   min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
-    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
-    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+    LLAMA_API void llama_sample_typical(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   p,
+                          size_t   min_keep);
+
+    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
+    LLAMA_API void llama_sample_entropy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates_p,
+                           float   min_temp,
+                           float   max_temp,
+                           float   exponent_val);
+
+    LLAMA_API void llama_sample_temp(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   temp);
+
+    LLAMA_API DEPRECATED(void llama_sample_temperature(
+                struct llama_context * ctx,
+              llama_token_data_array * candidates,
+                               float   temp),
+            "use llama_sample_temp instead");
+
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar);
 
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
@@ -233,28 +803,102 @@ extern "C" {
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                         int32_t   m,
+                           float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
     /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
     /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
-    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+                           float   tau,
+                           float   eta,
+                           float * mu);
 
     /// @details Selects the token with the highest probability.
-    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+    ///          Does not compute the token probabilities. Use llama_sample_softmax() instead.
+    LLAMA_API llama_token llama_sample_token_greedy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
 
     /// @details Randomly selects a token from the candidates based on their probabilities.
-    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+    LLAMA_API llama_token llama_sample_token(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates);
+
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(
+            struct llama_context * ctx,
+            struct llama_grammar * grammar,
+                     llama_token   token);
+
+    //
+    // Beam search
+    //
+
+    struct llama_beam_view {
+        const llama_token * tokens;
+
+        size_t n_tokens;
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
+    };
+
+    // Passed to beam_search_callback function.
+    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
+    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
+    // These pointers are valid only during the synchronous callback, so should not be saved.
+    struct llama_beams_state {
+        struct llama_beam_view * beam_views;
+
+        size_t n_beams;               // Number of elements in beam_views[].
+        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
+        bool   last_call;             // True iff this is the last callback invocation.
+    };
+
+    // Type of pointer to the beam_search_callback function.
+    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
+    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
+    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state);
+
+    /// @details Deterministically returns entire sentence constructed by a beam search.
+    /// @param ctx Pointer to the llama_context.
+    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
+    /// @param callback_data A pointer that is simply passed back to callback.
+    /// @param n_beams Number of beams to use.
+    /// @param n_past Number of tokens already evaluated.
+    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                int32_t   n_past,
+                                int32_t   n_predict);
 
     // Performance information
+    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
+
     LLAMA_API void llama_print_timings(struct llama_context * ctx);
     LLAMA_API void llama_reset_timings(struct llama_context * ctx);
 
     // Print system information
     LLAMA_API const char * llama_print_system_info(void);
 
+    // Set callback for all future logging events.
+    // If this is not called, or NULL is supplied, everything is output on stderr.
+    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
+
+    LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+
 #ifdef __cplusplus
 }
 #endif
@@ -264,10 +908,13 @@ extern "C" {
 
 #include <vector>
 #include <string>
+
 struct ggml_tensor;
 
-std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+);
 
-#endif
+#endif // LLAMA_API_INTERNAL
 
 #endif // LLAMA_H

examples/talk-llama/speak

@@ -9,6 +9,14 @@
 #
 #espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
 
+# piper
+#
+# https://github.com/rhasspy/piper
+#
+# Tested with Linux:
+#
+#echo "$2" | piper --model ~/en_US-lessac-medium.onnx --output-raw | aplay -q -r 22050 -f S16_LE -t raw -
+
 # for Mac
 say "$2"
 

examples/talk-llama/talk-llama.cpp

@@ -1,8 +1,8 @@
 // Talk with AI
 //
 
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
 #include "llama.h"
 
@@ -14,15 +14,37 @@
 #include <thread>
 #include <vector>
 #include <regex>
+#include <sstream>
 
 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
-    std::vector<llama_token> res(text.size() + (int)add_bos);
-    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
-    assert(n >= 0);
-    res.resize(n);
+    auto * model = llama_get_model(ctx);
 
-    return res;
+    // upper limit for the number of tokens
+    int n_tokens = text.length() + add_bos;
+    std::vector<llama_token> result(n_tokens);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+    return result;
+}
+
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+    std::vector<char> result(8, 0);
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+    if (n_tokens < 0) {
+        result.resize(-n_tokens);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+        GGML_ASSERT(check == -n_tokens);
+    } else {
+        result.resize(n_tokens);
+    }
+
+    return std::string(result.data(), result.size());
 }
 
 // command-line parameters
@@ -32,18 +54,23 @@ struct whisper_params {
     int32_t capture_id = -1;
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
+    int32_t n_gpu_layers = 999;
 
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
+    float vad_thold  = 0.6f;
+    float freq_thold = 100.0f;
 
-    bool speed_up      = false;
-    bool translate     = false;
-    bool print_special = false;
-    bool print_energy  = false;
-    bool no_timestamps = true;
+    bool speed_up       = false;
+    bool translate      = false;
+    bool print_special  = false;
+    bool print_energy   = false;
+    bool no_timestamps  = true;
     bool verbose_prompt = false;
+    bool use_gpu        = true;
 
     std::string person      = "Georgi";
+    std::string bot_name    = "LLaMA";
+    std::string wake_cmd    = "";
+    std::string heard_ok    = "";
     std::string language    = "en";
     std::string model_wsp   = "models/ggml-base.en.bin";
     std::string model_llama = "models/ggml-llama-7B.bin";
@@ -63,25 +90,30 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
-        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 == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
-        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
-        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
-        else if (arg == "--session")                        { params.path_session  = argv[++i];}
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
-        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
-        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
-        else if (arg == "--prompt-file")                    {
+        else if (arg == "-t"   || arg == "--threads")        { params.n_threads      = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")       { params.voice_ms       = std::stoi(argv[++i]); }
+        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 == "-ngl" || arg == "--n-gpu-layers")   { params.n_gpu_layers   = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")      { params.vad_thold      = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")     { params.freq_thold     = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")       { params.speed_up       = true; }
+        else if (arg == "-tr"  || arg == "--translate")      { params.translate      = true; }
+        else if (arg == "-ps"  || arg == "--print-special")  { params.print_special  = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")   { params.print_energy   = true; }
+        else if (arg == "-vp"  || arg == "--verbose-prompt") { params.verbose_prompt = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")         { params.use_gpu        = false; }
+        else if (arg == "-p"   || arg == "--person")         { params.person         = argv[++i]; }
+        else if (arg == "-bn"   || arg == "--bot-name")      { params.bot_name       = argv[++i]; }
+        else if (arg == "--session")                         { params.path_session   = argv[++i]; }
+        else if (arg == "-w"   || arg == "--wake-command")   { params.wake_cmd       = argv[++i]; }
+        else if (arg == "-ho"  || arg == "--heard-ok")       { params.heard_ok       = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")       { params.language       = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper")  { params.model_wsp      = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")    { params.model_llama    = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")          { params.speak          = argv[++i]; }
+        else if (arg == "--prompt-file")                     {
             std::ifstream file(argv[++i]);
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (params.prompt.back() == '\n') {
@@ -89,6 +121,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             }
         }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -104,27 +137,32 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "usage: %s [options]\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
-    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, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
-    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
-    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
-    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
-    fprintf(stderr, "  --session FNAME       file to cache model state in (may be large!) (default: none)\n");
-    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -h,       --help           [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N      [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N     [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    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, "  -ngl N,   --n-gpu-layers N [%-7d] number of layers to store in VRAM\n",           params.n_gpu_layers);
+    fprintf(stderr, "  -vth N,   --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -vp,      --verbose-prompt [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -p NAME,  --person NAME    [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
+    fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
+    fprintf(stderr, "  -ho TEXT, --heard-ok TEXT  [%-7s] said by TTS before generating reply\n",         params.heard_ok.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper  [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -ml FILE, --model-llama    [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  -s FILE,  --speak TEXT     [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  --prompt-file FNAME        [%-7s] file with custom prompt to start dialog\n",     "");
+    fprintf(stderr, "  --session FNAME                   file to cache model state in (may be large!) (default: none)\n");
+    fprintf(stderr, "  -f FNAME, --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
 
@@ -196,6 +234,18 @@ std::string transcribe(
     return result;
 }
 
+std::vector<std::string> get_words(const std::string &txt) {
+    std::vector<std::string> words;
+
+    std::istringstream iss(txt);
+    std::string word;
+    while (iss >> word) {
+        words.push_back(word);
+    }
+
+    return words;
+}
+
 const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";
 
 const std::string k_prompt_llama = R"(Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
@@ -223,7 +273,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (whisper_lang_id(params.language.c_str()) == -1) {
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
         fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
         whisper_print_usage(argc, argv, params);
         exit(0);
@@ -231,20 +281,32 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
 
     // llama init
 
-    llama_init_backend();
+    llama_backend_init(true);
 
-    auto lparams = llama_context_default_params();
+    auto lmparams = llama_model_default_params();
+    if (!params.use_gpu) {
+        lmparams.n_gpu_layers = 0;
+    } else {
+        lmparams.n_gpu_layers = params.n_gpu_layers;
+    }
+
+    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
+
+    llama_context_params lcparams = llama_context_default_params();
 
     // tune these to your liking
-    lparams.n_ctx      = 2048;
-    lparams.seed       = 1;
-    lparams.f16_kv     = true;
+    lcparams.n_ctx      = 2048;
+    lcparams.seed       = 1;
+    lcparams.n_threads  = params.n_threads;
 
-    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
 
     // print some info about the processing
     {
@@ -267,7 +329,6 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "\n");
     }
 
-
     // init audio
 
     audio_async audio(30*1000);
@@ -278,20 +339,17 @@ int main(int argc, char ** argv) {
 
     audio.resume();
 
-    int n_iter = 0;
-
     bool is_running  = true;
     bool force_speak = false;
 
     float prob0 = 0.0f;
 
     const std::string chat_symb = ":";
-    const std::string bot_name  = "LLaMA";
 
     std::vector<float> pcmf32_cur;
     std::vector<float> pcmf32_prompt;
 
-    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);
+    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", params.bot_name);
 
     // construct the initial prompt for LLaMA inference
     std::string prompt_llama = params.prompt.empty() ? k_prompt_llama : params.prompt;
@@ -300,7 +358,7 @@ int main(int argc, char ** argv) {
     prompt_llama.insert(0, 1, ' ');
 
     prompt_llama = ::replace(prompt_llama, "{0}", params.person);
-    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);
+    prompt_llama = ::replace(prompt_llama, "{1}", params.bot_name);
 
     {
         // get time string
@@ -343,7 +401,7 @@ int main(int argc, char ** argv) {
         if (fp != NULL) {
             std::fclose(fp);
 
-            session_tokens.resize(lparams.n_ctx);
+            session_tokens.resize(llama_n_ctx(ctx_llama));
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
@@ -365,7 +423,7 @@ int main(int argc, char ** argv) {
     printf("\n");
     printf("%s : initializing - please wait ...\n", __func__);
 
-    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0)) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return 1;
     }
@@ -402,6 +460,16 @@ int main(int argc, char ** argv) {
     bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < (embd_inp.size() * 3 / 4);
 
     printf("%s : done! start speaking in the microphone\n", __func__);
+
+    // show wake command if enabled
+    const std::string wake_cmd = params.wake_cmd;
+    const int wake_cmd_length = get_words(wake_cmd).size();
+    const bool use_wake_cmd = wake_cmd_length > 0;
+
+    if (use_wake_cmd) {
+        printf("%s : the wake-up command is: '%s%s%s'\n", __func__, "\033[1m", wake_cmd.c_str(), "\033[0m");
+    }
+
     printf("\n");
     printf("%s%s", params.person.c_str(), chat_symb.c_str());
     fflush(stdout);
@@ -447,10 +515,41 @@ int main(int argc, char ** argv) {
 
                 audio.get(params.voice_ms, pcmf32_cur);
 
-                std::string text_heard;
+                std::string all_heard;
 
                 if (!force_speak) {
-                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                    all_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                }
+
+                const auto words = get_words(all_heard);
+
+                std::string wake_cmd_heard;
+                std::string text_heard;
+
+                for (int i = 0; i < (int) words.size(); ++i) {
+                    if (i < wake_cmd_length) {
+                        wake_cmd_heard += words[i] + " ";
+                    } else {
+                        text_heard += words[i] + " ";
+                    }
+                }
+
+                // check if audio starts with the wake-up command if enabled
+                if (use_wake_cmd) {
+                    const float sim = similarity(wake_cmd_heard, wake_cmd);
+
+                    if ((sim < 0.7f) || (text_heard.empty())) {
+                        audio.clear();
+                        continue;
+                    }
+                }
+
+                // optionally give audio feedback that the current text is being processed
+                if (!params.heard_ok.empty()) {
+                    int ret = system((params.speak + " " + std::to_string(voice_id) + " '" + params.heard_ok + "'").c_str());
+                    if (ret != 0) {
+                        fprintf(stderr, "%s: failed to speak\n", __func__);
+                    }
                 }
 
                 // remove text between brackets using regex
@@ -487,7 +586,7 @@ int main(int argc, char ** argv) {
                 force_speak = false;
 
                 text_heard.insert(0, 1, ' ');
-                text_heard += "\n" + bot_name + chat_symb;
+                text_heard += "\n" + params.bot_name + chat_symb;
                 fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
                 fflush(stdout);
 
@@ -514,7 +613,7 @@ int main(int argc, char ** argv) {
                             //printf("\n---\n");
                             //printf("resetting: '");
                             //for (int i = 0; i < (int) embd.size(); i++) {
-                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                            //    printf("%s", llama_token_to_piece(ctx_llama, embd[i]));
                             //}
                             //printf("'\n");
                             //printf("\n---\n");
@@ -548,7 +647,7 @@ int main(int argc, char ** argv) {
                             n_session_consumed = session_tokens.size();
                         }
 
-                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past)) {
                             fprintf(stderr, "%s : failed to eval\n", __func__);
                             return 1;
                         }
@@ -580,9 +679,9 @@ int main(int argc, char ** argv) {
 
                         {
                             auto logits = llama_get_logits(ctx_llama);
-                            auto n_vocab = llama_n_vocab(ctx_llama);
+                            auto n_vocab = llama_n_vocab(model_llama);
 
-                            logits[llama_token_eos()] = 0;
+                            logits[llama_token_eos(model_llama)] = 0;
 
                             std::vector<llama_token_data> candidates;
                             candidates.reserve(n_vocab);
@@ -593,13 +692,13 @@ int main(int argc, char ** argv) {
                             llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
                             // apply repeat penalty
-                            const float nl_logit = logits[llama_token_nl()];
+                            const float nl_logit = logits[llama_token_nl(model_llama)];
 
-                            llama_sample_repetition_penalty(ctx_llama, &candidates_p,
+                            llama_sample_repetition_penalties(ctx_llama, &candidates_p,
                                     embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, repeat_penalty);
+                                    repeat_last_n, repeat_penalty, 0.0, 0.0f);
 
-                            logits[llama_token_nl()] = nl_logit;
+                            logits[llama_token_nl(model_llama)] = nl_logit;
 
                             if (temp <= 0) {
                                 // Greedy sampling
@@ -608,27 +707,28 @@ int main(int argc, char ** argv) {
                                 // Temperature sampling
                                 llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
                                 llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                                llama_sample_temperature(ctx_llama, &candidates_p, temp);
+                                llama_sample_temp (ctx_llama, &candidates_p, temp);
                                 id = llama_sample_token(ctx_llama, &candidates_p);
                             }
                         }
 
-                        if (id != llama_token_eos()) {
+                        if (id != llama_token_eos(model_llama)) {
                             // add it to the context
                             embd.push_back(id);
 
-                            text_to_speak += llama_token_to_str(ctx_llama, id);
+                            text_to_speak += llama_token_to_piece(ctx_llama, id);
 
-                            printf("%s", llama_token_to_str(ctx_llama, id));
+                            printf("%s", llama_token_to_piece(ctx_llama, id).c_str());
+                            fflush(stdout);
                         }
                     }
 
                     {
                         std::string last_output;
                         for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
-                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                            last_output += llama_token_to_piece(ctx_llama, embd_inp[i]);
                         }
-                        last_output += llama_token_to_str(ctx_llama, embd[0]);
+                        last_output += llama_token_to_piece(ctx_llama, embd[0]);
 
                         for (std::string & antiprompt : antiprompts) {
                             if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
@@ -648,12 +748,13 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                text_to_speak = ::replace(text_to_speak, "\"", "");
-                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+                text_to_speak = ::replace(text_to_speak, "'", "'\"'\"'");
+                int ret = system((params.speak + " " + std::to_string(voice_id) + " '" + text_to_speak + "'").c_str());
+                if (ret != 0) {
+                    fprintf(stderr, "%s: failed to speak\n", __func__);
+                }
 
                 audio.clear();
-
-                ++n_iter;
             }
         }
     }

examples/talk-llama/unicode.h

@@ -0,0 +1,463 @@
+#pragma once
+
+#include <cassert>
+#include <stdexcept>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static const std::vector<std::pair<uint32_t, uint32_t>> digit_ranges = {
+{0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F},
+{0xCE6, 0xCEF}, {0xD66, 0xD6F}, {0xDE6, 0xDEF}, {0xE50, 0xE59}, {0xED0, 0xED9}, {0xF20, 0xF29}, {0x1040, 0x1049}, {0x1090, 0x1099}, {0x1369, 0x1371}, {0x17E0, 0x17E9}, {0x1810, 0x1819}, {0x1946, 0x194F},
+{0x19D0, 0x19DA}, {0x1A80, 0x1A89}, {0x1A90, 0x1A99}, {0x1B50, 0x1B59}, {0x1BB0, 0x1BB9}, {0x1C40, 0x1C49}, {0x1C50, 0x1C59}, {0x2070, 0x2070}, {0x2074, 0x2079}, {0x2080, 0x2089}, {0x2460, 0x2468},
+{0x2474, 0x247C}, {0x2488, 0x2490}, {0x24EA, 0x24EA}, {0x24F5, 0x24FD}, {0x24FF, 0x24FF}, {0x2776, 0x277E}, {0x2780, 0x2788}, {0x278A, 0x2792}, {0xA620, 0xA629}, {0xA8D0, 0xA8D9}, {0xA900, 0xA909},
+{0xA9D0, 0xA9D9}, {0xA9F0, 0xA9F9}, {0xAA50, 0xAA59}, {0xABF0, 0xABF9}, {0xFF10, 0xFF19}, {0x104A0, 0x104A9}, {0x10A40, 0x10A43}, {0x10D30, 0x10D39}, {0x10E60, 0x10E68}, {0x11052, 0x1105A},
+{0x11066, 0x1106F}, {0x110F0, 0x110F9}, {0x11136, 0x1113F}, {0x111D0, 0x111D9}, {0x112F0, 0x112F9}, {0x11450, 0x11459}, {0x114D0, 0x114D9}, {0x11650, 0x11659}, {0x116C0, 0x116C9}, {0x11730, 0x11739},
+{0x118E0, 0x118E9}, {0x11950, 0x11959}, {0x11C50, 0x11C59}, {0x11D50, 0x11D59}, {0x11DA0, 0x11DA9}, {0x16A60, 0x16A69}, {0x16B50, 0x16B59}, {0x1D7CE, 0x1D7FF}, {0x1E140, 0x1E149}, {0x1E2F0, 0x1E2F9},
+{0x1E950, 0x1E959}, {0x1F100, 0x1F10A}, {0x1FBF0, 0x1FBF9},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> letter_ranges = {
+{0x41, 0x5A}, {0x61, 0x7A}, {0xAA, 0xAA}, {0xB5, 0xB5}, {0xBA, 0xBA}, {0xC0, 0xD6}, {0xD8, 0xF6}, {0xF8, 0x2C1}, {0x2C6, 0x2D1}, {0x2E0, 0x2E4}, {0x2EC, 0x2EC}, {0x2EE, 0x2EE}, {0x370, 0x374},
+{0x376, 0x377}, {0x37A, 0x37D}, {0x37F, 0x37F}, {0x386, 0x386}, {0x388, 0x38A}, {0x38C, 0x38C}, {0x38E, 0x3A1}, {0x3A3, 0x3F5}, {0x3F7, 0x481}, {0x48A, 0x52F}, {0x531, 0x556}, {0x559, 0x559},
+{0x560, 0x588}, {0x5D0, 0x5EA}, {0x5EF, 0x5F2}, {0x620, 0x64A}, {0x66E, 0x66F}, {0x671, 0x6D3}, {0x6D5, 0x6D5}, {0x6E5, 0x6E6}, {0x6EE, 0x6EF}, {0x6FA, 0x6FC}, {0x6FF, 0x6FF}, {0x710, 0x710},
+{0x712, 0x72F}, {0x74D, 0x7A5}, {0x7B1, 0x7B1}, {0x7CA, 0x7EA}, {0x7F4, 0x7F5}, {0x7FA, 0x7FA}, {0x800, 0x815}, {0x81A, 0x81A}, {0x824, 0x824}, {0x828, 0x828}, {0x840, 0x858}, {0x860, 0x86A},
+{0x8A0, 0x8B4}, {0x8B6, 0x8C7}, {0x904, 0x939}, {0x93D, 0x93D}, {0x950, 0x950}, {0x958, 0x961}, {0x971, 0x980}, {0x985, 0x98C}, {0x98F, 0x990}, {0x993, 0x9A8}, {0x9AA, 0x9B0}, {0x9B2, 0x9B2},
+{0x9B6, 0x9B9}, {0x9BD, 0x9BD}, {0x9CE, 0x9CE}, {0x9DC, 0x9DD}, {0x9DF, 0x9E1}, {0x9F0, 0x9F1}, {0x9FC, 0x9FC}, {0xA05, 0xA0A}, {0xA0F, 0xA10}, {0xA13, 0xA28}, {0xA2A, 0xA30}, {0xA32, 0xA33},
+{0xA35, 0xA36}, {0xA38, 0xA39}, {0xA59, 0xA5C}, {0xA5E, 0xA5E}, {0xA72, 0xA74}, {0xA85, 0xA8D}, {0xA8F, 0xA91}, {0xA93, 0xAA8}, {0xAAA, 0xAB0}, {0xAB2, 0xAB3}, {0xAB5, 0xAB9}, {0xABD, 0xABD},
+{0xAD0, 0xAD0}, {0xAE0, 0xAE1}, {0xAF9, 0xAF9}, {0xB05, 0xB0C}, {0xB0F, 0xB10}, {0xB13, 0xB28}, {0xB2A, 0xB30}, {0xB32, 0xB33}, {0xB35, 0xB39}, {0xB3D, 0xB3D}, {0xB5C, 0xB5D}, {0xB5F, 0xB61},
+{0xB71, 0xB71}, {0xB83, 0xB83}, {0xB85, 0xB8A}, {0xB8E, 0xB90}, {0xB92, 0xB95}, {0xB99, 0xB9A}, {0xB9C, 0xB9C}, {0xB9E, 0xB9F}, {0xBA3, 0xBA4}, {0xBA8, 0xBAA}, {0xBAE, 0xBB9}, {0xBD0, 0xBD0},
+{0xC05, 0xC0C}, {0xC0E, 0xC10}, {0xC12, 0xC28}, {0xC2A, 0xC39}, {0xC3D, 0xC3D}, {0xC58, 0xC5A}, {0xC60, 0xC61}, {0xC80, 0xC80}, {0xC85, 0xC8C}, {0xC8E, 0xC90}, {0xC92, 0xCA8}, {0xCAA, 0xCB3},
+{0xCB5, 0xCB9}, {0xCBD, 0xCBD}, {0xCDE, 0xCDE}, {0xCE0, 0xCE1}, {0xCF1, 0xCF2}, {0xD04, 0xD0C}, {0xD0E, 0xD10}, {0xD12, 0xD3A}, {0xD3D, 0xD3D}, {0xD4E, 0xD4E}, {0xD54, 0xD56}, {0xD5F, 0xD61},
+{0xD7A, 0xD7F}, {0xD85, 0xD96}, {0xD9A, 0xDB1}, {0xDB3, 0xDBB}, {0xDBD, 0xDBD}, {0xDC0, 0xDC6}, {0xE01, 0xE30}, {0xE32, 0xE33}, {0xE40, 0xE46}, {0xE81, 0xE82}, {0xE84, 0xE84}, {0xE86, 0xE8A},
+{0xE8C, 0xEA3}, {0xEA5, 0xEA5}, {0xEA7, 0xEB0}, {0xEB2, 0xEB3}, {0xEBD, 0xEBD}, {0xEC0, 0xEC4}, {0xEC6, 0xEC6}, {0xEDC, 0xEDF}, {0xF00, 0xF00}, {0xF40, 0xF47}, {0xF49, 0xF6C}, {0xF88, 0xF8C},
+{0x1000, 0x102A}, {0x103F, 0x103F}, {0x1050, 0x1055}, {0x105A, 0x105D}, {0x1061, 0x1061}, {0x1065, 0x1066}, {0x106E, 0x1070}, {0x1075, 0x1081}, {0x108E, 0x108E}, {0x10A0, 0x10C5}, {0x10C7, 0x10C7},
+{0x10CD, 0x10CD}, {0x10D0, 0x10FA}, {0x10FC, 0x1248}, {0x124A, 0x124D}, {0x1250, 0x1256}, {0x1258, 0x1258}, {0x125A, 0x125D}, {0x1260, 0x1288}, {0x128A, 0x128D}, {0x1290, 0x12B0}, {0x12B2, 0x12B5},
+{0x12B8, 0x12BE}, {0x12C0, 0x12C0}, {0x12C2, 0x12C5}, {0x12C8, 0x12D6}, {0x12D8, 0x1310}, {0x1312, 0x1315}, {0x1318, 0x135A}, {0x1380, 0x138F}, {0x13A0, 0x13F5}, {0x13F8, 0x13FD}, {0x1401, 0x166C},
+{0x166F, 0x167F}, {0x1681, 0x169A}, {0x16A0, 0x16EA}, {0x16F1, 0x16F8}, {0x1700, 0x170C}, {0x170E, 0x1711}, {0x1720, 0x1731}, {0x1740, 0x1751}, {0x1760, 0x176C}, {0x176E, 0x1770}, {0x1780, 0x17B3},
+{0x17D7, 0x17D7}, {0x17DC, 0x17DC}, {0x1820, 0x1878}, {0x1880, 0x1884}, {0x1887, 0x18A8}, {0x18AA, 0x18AA}, {0x18B0, 0x18F5}, {0x1900, 0x191E}, {0x1950, 0x196D}, {0x1970, 0x1974}, {0x1980, 0x19AB},
+{0x19B0, 0x19C9}, {0x1A00, 0x1A16}, {0x1A20, 0x1A54}, {0x1AA7, 0x1AA7}, {0x1B05, 0x1B33}, {0x1B45, 0x1B4B}, {0x1B83, 0x1BA0}, {0x1BAE, 0x1BAF}, {0x1BBA, 0x1BE5}, {0x1C00, 0x1C23}, {0x1C4D, 0x1C4F},
+{0x1C5A, 0x1C7D}, {0x1C80, 0x1C88}, {0x1C90, 0x1CBA}, {0x1CBD, 0x1CBF}, {0x1CE9, 0x1CEC}, {0x1CEE, 0x1CF3}, {0x1CF5, 0x1CF6}, {0x1CFA, 0x1CFA}, {0x1D00, 0x1DBF}, {0x1E00, 0x1F15}, {0x1F18, 0x1F1D},
+{0x1F20, 0x1F45}, {0x1F48, 0x1F4D}, {0x1F50, 0x1F57}, {0x1F59, 0x1F59}, {0x1F5B, 0x1F5B}, {0x1F5D, 0x1F5D}, {0x1F5F, 0x1F7D}, {0x1F80, 0x1FB4}, {0x1FB6, 0x1FBC}, {0x1FBE, 0x1FBE}, {0x1FC2, 0x1FC4},
+{0x1FC6, 0x1FCC}, {0x1FD0, 0x1FD3}, {0x1FD6, 0x1FDB}, {0x1FE0, 0x1FEC}, {0x1FF2, 0x1FF4}, {0x1FF6, 0x1FFC}, {0x2071, 0x2071}, {0x207F, 0x207F}, {0x2090, 0x209C}, {0x2102, 0x2102}, {0x2107, 0x2107},
+{0x210A, 0x2113}, {0x2115, 0x2115}, {0x2119, 0x211D}, {0x2124, 0x2124}, {0x2126, 0x2126}, {0x2128, 0x2128}, {0x212A, 0x212D}, {0x212F, 0x2139}, {0x213C, 0x213F}, {0x2145, 0x2149}, {0x214E, 0x214E},
+{0x2183, 0x2184}, {0x2C00, 0x2C2E}, {0x2C30, 0x2C5E}, {0x2C60, 0x2CE4}, {0x2CEB, 0x2CEE}, {0x2CF2, 0x2CF3}, {0x2D00, 0x2D25}, {0x2D27, 0x2D27}, {0x2D2D, 0x2D2D}, {0x2D30, 0x2D67}, {0x2D6F, 0x2D6F},
+{0x2D80, 0x2D96}, {0x2DA0, 0x2DA6}, {0x2DA8, 0x2DAE}, {0x2DB0, 0x2DB6}, {0x2DB8, 0x2DBE}, {0x2DC0, 0x2DC6}, {0x2DC8, 0x2DCE}, {0x2DD0, 0x2DD6}, {0x2DD8, 0x2DDE}, {0x2E2F, 0x2E2F}, {0x3005, 0x3006},
+{0x3031, 0x3035}, {0x303B, 0x303C}, {0x3041, 0x3096}, {0x309D, 0x309F}, {0x30A1, 0x30FA}, {0x30FC, 0x30FF}, {0x3105, 0x312F}, {0x3131, 0x318E}, {0x31A0, 0x31BF}, {0x31F0, 0x31FF}, {0x3400, 0x4DBF},
+{0x4E00, 0x9FFC}, {0xA000, 0xA48C}, {0xA4D0, 0xA4FD}, {0xA500, 0xA60C}, {0xA610, 0xA61F}, {0xA62A, 0xA62B}, {0xA640, 0xA66E}, {0xA67F, 0xA69D}, {0xA6A0, 0xA6E5}, {0xA717, 0xA71F}, {0xA722, 0xA788},
+{0xA78B, 0xA7BF}, {0xA7C2, 0xA7CA}, {0xA7F5, 0xA801}, {0xA803, 0xA805}, {0xA807, 0xA80A}, {0xA80C, 0xA822}, {0xA840, 0xA873}, {0xA882, 0xA8B3}, {0xA8F2, 0xA8F7}, {0xA8FB, 0xA8FB}, {0xA8FD, 0xA8FE},
+{0xA90A, 0xA925}, {0xA930, 0xA946}, {0xA960, 0xA97C}, {0xA984, 0xA9B2}, {0xA9CF, 0xA9CF}, {0xA9E0, 0xA9E4}, {0xA9E6, 0xA9EF}, {0xA9FA, 0xA9FE}, {0xAA00, 0xAA28}, {0xAA40, 0xAA42}, {0xAA44, 0xAA4B},
+{0xAA60, 0xAA76}, {0xAA7A, 0xAA7A}, {0xAA7E, 0xAAAF}, {0xAAB1, 0xAAB1}, {0xAAB5, 0xAAB6}, {0xAAB9, 0xAABD}, {0xAAC0, 0xAAC0}, {0xAAC2, 0xAAC2}, {0xAADB, 0xAADD}, {0xAAE0, 0xAAEA}, {0xAAF2, 0xAAF4},
+{0xAB01, 0xAB06}, {0xAB09, 0xAB0E}, {0xAB11, 0xAB16}, {0xAB20, 0xAB26}, {0xAB28, 0xAB2E}, {0xAB30, 0xAB5A}, {0xAB5C, 0xAB69}, {0xAB70, 0xABE2}, {0xAC00, 0xD7A3}, {0xD7B0, 0xD7C6}, {0xD7CB, 0xD7FB},
+{0xF900, 0xFA6D}, {0xFA70, 0xFAD9}, {0xFB00, 0xFB06}, {0xFB13, 0xFB17}, {0xFB1D, 0xFB1D}, {0xFB1F, 0xFB28}, {0xFB2A, 0xFB36}, {0xFB38, 0xFB3C}, {0xFB3E, 0xFB3E}, {0xFB40, 0xFB41}, {0xFB43, 0xFB44},
+{0xFB46, 0xFBB1}, {0xFBD3, 0xFD3D}, {0xFD50, 0xFD8F}, {0xFD92, 0xFDC7}, {0xFDF0, 0xFDFB}, {0xFE70, 0xFE74}, {0xFE76, 0xFEFC}, {0xFF21, 0xFF3A}, {0xFF41, 0xFF5A}, {0xFF66, 0xFFBE}, {0xFFC2, 0xFFC7},
+{0xFFCA, 0xFFCF}, {0xFFD2, 0xFFD7}, {0xFFDA, 0xFFDC}, {0x10000, 0x1000B}, {0x1000D, 0x10026}, {0x10028, 0x1003A}, {0x1003C, 0x1003D}, {0x1003F, 0x1004D}, {0x10050, 0x1005D}, {0x10080, 0x100FA},
+{0x10280, 0x1029C}, {0x102A0, 0x102D0}, {0x10300, 0x1031F}, {0x1032D, 0x10340}, {0x10342, 0x10349}, {0x10350, 0x10375}, {0x10380, 0x1039D}, {0x103A0, 0x103C3}, {0x103C8, 0x103CF}, {0x10400, 0x1049D},
+{0x104B0, 0x104D3}, {0x104D8, 0x104FB}, {0x10500, 0x10527}, {0x10530, 0x10563}, {0x10600, 0x10736}, {0x10740, 0x10755}, {0x10760, 0x10767}, {0x10800, 0x10805}, {0x10808, 0x10808}, {0x1080A, 0x10835},
+{0x10837, 0x10838}, {0x1083C, 0x1083C}, {0x1083F, 0x10855}, {0x10860, 0x10876}, {0x10880, 0x1089E}, {0x108E0, 0x108F2}, {0x108F4, 0x108F5}, {0x10900, 0x10915}, {0x10920, 0x10939}, {0x10980, 0x109B7},
+{0x109BE, 0x109BF}, {0x10A00, 0x10A00}, {0x10A10, 0x10A13}, {0x10A15, 0x10A17}, {0x10A19, 0x10A35}, {0x10A60, 0x10A7C}, {0x10A80, 0x10A9C}, {0x10AC0, 0x10AC7}, {0x10AC9, 0x10AE4}, {0x10B00, 0x10B35},
+{0x10B40, 0x10B55}, {0x10B60, 0x10B72}, {0x10B80, 0x10B91}, {0x10C00, 0x10C48}, {0x10C80, 0x10CB2}, {0x10CC0, 0x10CF2}, {0x10D00, 0x10D23}, {0x10E80, 0x10EA9}, {0x10EB0, 0x10EB1}, {0x10F00, 0x10F1C},
+{0x10F27, 0x10F27}, {0x10F30, 0x10F45}, {0x10FB0, 0x10FC4}, {0x10FE0, 0x10FF6}, {0x11003, 0x11037}, {0x11083, 0x110AF}, {0x110D0, 0x110E8}, {0x11103, 0x11126}, {0x11144, 0x11144}, {0x11147, 0x11147},
+{0x11150, 0x11172}, {0x11176, 0x11176}, {0x11183, 0x111B2}, {0x111C1, 0x111C4}, {0x111DA, 0x111DA}, {0x111DC, 0x111DC}, {0x11200, 0x11211}, {0x11213, 0x1122B}, {0x11280, 0x11286}, {0x11288, 0x11288},
+{0x1128A, 0x1128D}, {0x1128F, 0x1129D}, {0x1129F, 0x112A8}, {0x112B0, 0x112DE}, {0x11305, 0x1130C}, {0x1130F, 0x11310}, {0x11313, 0x11328}, {0x1132A, 0x11330}, {0x11332, 0x11333}, {0x11335, 0x11339},
+{0x1133D, 0x1133D}, {0x11350, 0x11350}, {0x1135D, 0x11361}, {0x11400, 0x11434}, {0x11447, 0x1144A}, {0x1145F, 0x11461}, {0x11480, 0x114AF}, {0x114C4, 0x114C5}, {0x114C7, 0x114C7}, {0x11580, 0x115AE},
+{0x115D8, 0x115DB}, {0x11600, 0x1162F}, {0x11644, 0x11644}, {0x11680, 0x116AA}, {0x116B8, 0x116B8}, {0x11700, 0x1171A}, {0x11800, 0x1182B}, {0x118A0, 0x118DF}, {0x118FF, 0x11906}, {0x11909, 0x11909},
+{0x1190C, 0x11913}, {0x11915, 0x11916}, {0x11918, 0x1192F}, {0x1193F, 0x1193F}, {0x11941, 0x11941}, {0x119A0, 0x119A7}, {0x119AA, 0x119D0}, {0x119E1, 0x119E1}, {0x119E3, 0x119E3}, {0x11A00, 0x11A00},
+{0x11A0B, 0x11A32}, {0x11A3A, 0x11A3A}, {0x11A50, 0x11A50}, {0x11A5C, 0x11A89}, {0x11A9D, 0x11A9D}, {0x11AC0, 0x11AF8}, {0x11C00, 0x11C08}, {0x11C0A, 0x11C2E}, {0x11C40, 0x11C40}, {0x11C72, 0x11C8F},
+{0x11D00, 0x11D06}, {0x11D08, 0x11D09}, {0x11D0B, 0x11D30}, {0x11D46, 0x11D46}, {0x11D60, 0x11D65}, {0x11D67, 0x11D68}, {0x11D6A, 0x11D89}, {0x11D98, 0x11D98}, {0x11EE0, 0x11EF2}, {0x11FB0, 0x11FB0},
+{0x12000, 0x12399}, {0x12480, 0x12543}, {0x13000, 0x1342E}, {0x14400, 0x14646}, {0x16800, 0x16A38}, {0x16A40, 0x16A5E}, {0x16AD0, 0x16AED}, {0x16B00, 0x16B2F}, {0x16B40, 0x16B43}, {0x16B63, 0x16B77},
+{0x16B7D, 0x16B8F}, {0x16E40, 0x16E7F}, {0x16F00, 0x16F4A}, {0x16F50, 0x16F50}, {0x16F93, 0x16F9F}, {0x16FE0, 0x16FE1}, {0x16FE3, 0x16FE3}, {0x17000, 0x187F7}, {0x18800, 0x18CD5}, {0x18D00, 0x18D08},
+{0x1B000, 0x1B11E}, {0x1B150, 0x1B152}, {0x1B164, 0x1B167}, {0x1B170, 0x1B2FB}, {0x1BC00, 0x1BC6A}, {0x1BC70, 0x1BC7C}, {0x1BC80, 0x1BC88}, {0x1BC90, 0x1BC99}, {0x1D400, 0x1D454}, {0x1D456, 0x1D49C},
+{0x1D49E, 0x1D49F}, {0x1D4A2, 0x1D4A2}, {0x1D4A5, 0x1D4A6}, {0x1D4A9, 0x1D4AC}, {0x1D4AE, 0x1D4B9}, {0x1D4BB, 0x1D4BB}, {0x1D4BD, 0x1D4C3}, {0x1D4C5, 0x1D505}, {0x1D507, 0x1D50A}, {0x1D50D, 0x1D514},
+{0x1D516, 0x1D51C}, {0x1D51E, 0x1D539}, {0x1D53B, 0x1D53E}, {0x1D540, 0x1D544}, {0x1D546, 0x1D546}, {0x1D54A, 0x1D550}, {0x1D552, 0x1D6A5}, {0x1D6A8, 0x1D6C0}, {0x1D6C2, 0x1D6DA}, {0x1D6DC, 0x1D6FA},
+{0x1D6FC, 0x1D714}, {0x1D716, 0x1D734}, {0x1D736, 0x1D74E}, {0x1D750, 0x1D76E}, {0x1D770, 0x1D788}, {0x1D78A, 0x1D7A8}, {0x1D7AA, 0x1D7C2}, {0x1D7C4, 0x1D7CB}, {0x1E100, 0x1E12C}, {0x1E137, 0x1E13D},
+{0x1E14E, 0x1E14E}, {0x1E2C0, 0x1E2EB}, {0x1E800, 0x1E8C4}, {0x1E900, 0x1E943}, {0x1E94B, 0x1E94B}, {0x1EE00, 0x1EE03}, {0x1EE05, 0x1EE1F}, {0x1EE21, 0x1EE22}, {0x1EE24, 0x1EE24}, {0x1EE27, 0x1EE27},
+{0x1EE29, 0x1EE32}, {0x1EE34, 0x1EE37}, {0x1EE39, 0x1EE39}, {0x1EE3B, 0x1EE3B}, {0x1EE42, 0x1EE42}, {0x1EE47, 0x1EE47}, {0x1EE49, 0x1EE49}, {0x1EE4B, 0x1EE4B}, {0x1EE4D, 0x1EE4F}, {0x1EE51, 0x1EE52},
+{0x1EE54, 0x1EE54}, {0x1EE57, 0x1EE57}, {0x1EE59, 0x1EE59}, {0x1EE5B, 0x1EE5B}, {0x1EE5D, 0x1EE5D}, {0x1EE5F, 0x1EE5F}, {0x1EE61, 0x1EE62}, {0x1EE64, 0x1EE64}, {0x1EE67, 0x1EE6A}, {0x1EE6C, 0x1EE72},
+{0x1EE74, 0x1EE77}, {0x1EE79, 0x1EE7C}, {0x1EE7E, 0x1EE7E}, {0x1EE80, 0x1EE89}, {0x1EE8B, 0x1EE9B}, {0x1EEA1, 0x1EEA3}, {0x1EEA5, 0x1EEA9}, {0x1EEAB, 0x1EEBB}, {0x20000, 0x2A6DD}, {0x2A700, 0x2B734},
+{0x2B740, 0x2B81D}, {0x2B820, 0x2CEA1}, {0x2CEB0, 0x2EBE0}, {0x2F800, 0x2FA1D}, {0x30000, 0x3134A},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> whitespace_ranges = {
+{0x9, 0xD}, {0x1C, 0x20}, {0x85, 0x85}, {0xA0, 0xA0}, {0x1680, 0x1680}, {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> accent_mark_ranges = {
+{0x300, 0x36F}, {0x483, 0x489}, {0x591, 0x5BD}, {0x5BF, 0x5BF}, {0x5C1, 0x5C2}, {0x5C4, 0x5C5}, {0x5C7, 0x5C7}, {0x610, 0x61A}, {0x64B, 0x65F}, {0x670, 0x670}, {0x6D6, 0x6DC}, {0x6DF, 0x6E4},
+{0x6E7, 0x6E8}, {0x6EA, 0x6ED}, {0x711, 0x711}, {0x730, 0x74A}, {0x7A6, 0x7B0}, {0x7EB, 0x7F3}, {0x7FD, 0x7FD}, {0x816, 0x819}, {0x81B, 0x823}, {0x825, 0x827}, {0x829, 0x82D}, {0x859, 0x85B},
+{0x8D3, 0x8E1}, {0x8E3, 0x903}, {0x93A, 0x93C}, {0x93E, 0x94F}, {0x951, 0x957}, {0x962, 0x963}, {0x981, 0x983}, {0x9BC, 0x9BC}, {0x9BE, 0x9C4}, {0x9C7, 0x9C8}, {0x9CB, 0x9CD}, {0x9D7, 0x9D7},
+{0x9E2, 0x9E3}, {0x9FE, 0x9FE}, {0xA01, 0xA03}, {0xA3C, 0xA3C}, {0xA3E, 0xA42}, {0xA47, 0xA48}, {0xA4B, 0xA4D}, {0xA51, 0xA51}, {0xA70, 0xA71}, {0xA75, 0xA75}, {0xA81, 0xA83}, {0xABC, 0xABC},
+{0xABE, 0xAC5}, {0xAC7, 0xAC9}, {0xACB, 0xACD}, {0xAE2, 0xAE3}, {0xAFA, 0xAFF}, {0xB01, 0xB03}, {0xB3C, 0xB3C}, {0xB3E, 0xB44}, {0xB47, 0xB48}, {0xB4B, 0xB4D}, {0xB55, 0xB57}, {0xB62, 0xB63},
+{0xB82, 0xB82}, {0xBBE, 0xBC2}, {0xBC6, 0xBC8}, {0xBCA, 0xBCD}, {0xBD7, 0xBD7}, {0xC00, 0xC04}, {0xC3E, 0xC44}, {0xC46, 0xC48}, {0xC4A, 0xC4D}, {0xC55, 0xC56}, {0xC62, 0xC63}, {0xC81, 0xC83},
+{0xCBC, 0xCBC}, {0xCBE, 0xCC4}, {0xCC6, 0xCC8}, {0xCCA, 0xCCD}, {0xCD5, 0xCD6}, {0xCE2, 0xCE3}, {0xD00, 0xD03}, {0xD3B, 0xD3C}, {0xD3E, 0xD44}, {0xD46, 0xD48}, {0xD4A, 0xD4D}, {0xD57, 0xD57},
+{0xD62, 0xD63}, {0xD81, 0xD83}, {0xDCA, 0xDCA}, {0xDCF, 0xDD4}, {0xDD6, 0xDD6}, {0xDD8, 0xDDF}, {0xDF2, 0xDF3}, {0xE31, 0xE31}, {0xE34, 0xE3A}, {0xE47, 0xE4E}, {0xEB1, 0xEB1}, {0xEB4, 0xEBC},
+{0xEC8, 0xECD}, {0xF18, 0xF19}, {0xF35, 0xF35}, {0xF37, 0xF37}, {0xF39, 0xF39}, {0xF3E, 0xF3F}, {0xF71, 0xF84}, {0xF86, 0xF87}, {0xF8D, 0xF97}, {0xF99, 0xFBC}, {0xFC6, 0xFC6}, {0x102B, 0x103E},
+{0x1056, 0x1059}, {0x105E, 0x1060}, {0x1062, 0x1064}, {0x1067, 0x106D}, {0x1071, 0x1074}, {0x1082, 0x108D}, {0x108F, 0x108F}, {0x109A, 0x109D}, {0x135D, 0x135F}, {0x1712, 0x1714}, {0x1732, 0x1734},
+{0x1752, 0x1753}, {0x1772, 0x1773}, {0x17B4, 0x17D3}, {0x17DD, 0x17DD}, {0x180B, 0x180D}, {0x1885, 0x1886}, {0x18A9, 0x18A9}, {0x1920, 0x192B}, {0x1930, 0x193B}, {0x1A17, 0x1A1B}, {0x1A55, 0x1A5E},
+{0x1A60, 0x1A7C}, {0x1A7F, 0x1A7F}, {0x1AB0, 0x1AC0}, {0x1B00, 0x1B04}, {0x1B34, 0x1B44}, {0x1B6B, 0x1B73}, {0x1B80, 0x1B82}, {0x1BA1, 0x1BAD}, {0x1BE6, 0x1BF3}, {0x1C24, 0x1C37}, {0x1CD0, 0x1CD2},
+{0x1CD4, 0x1CE8}, {0x1CED, 0x1CED}, {0x1CF4, 0x1CF4}, {0x1CF7, 0x1CF9}, {0x1DC0, 0x1DF9}, {0x1DFB, 0x1DFF}, {0x20D0, 0x20F0}, {0x2CEF, 0x2CF1}, {0x2D7F, 0x2D7F}, {0x2DE0, 0x2DFF}, {0x302A, 0x302F},
+{0x3099, 0x309A}, {0xA66F, 0xA672}, {0xA674, 0xA67D}, {0xA69E, 0xA69F}, {0xA6F0, 0xA6F1}, {0xA802, 0xA802}, {0xA806, 0xA806}, {0xA80B, 0xA80B}, {0xA823, 0xA827}, {0xA82C, 0xA82C}, {0xA880, 0xA881},
+{0xA8B4, 0xA8C5}, {0xA8E0, 0xA8F1}, {0xA8FF, 0xA8FF}, {0xA926, 0xA92D}, {0xA947, 0xA953}, {0xA980, 0xA983}, {0xA9B3, 0xA9C0}, {0xA9E5, 0xA9E5}, {0xAA29, 0xAA36}, {0xAA43, 0xAA43}, {0xAA4C, 0xAA4D},
+{0xAA7B, 0xAA7D}, {0xAAB0, 0xAAB0}, {0xAAB2, 0xAAB4}, {0xAAB7, 0xAAB8}, {0xAABE, 0xAABF}, {0xAAC1, 0xAAC1}, {0xAAEB, 0xAAEF}, {0xAAF5, 0xAAF6}, {0xABE3, 0xABEA}, {0xABEC, 0xABED}, {0xFB1E, 0xFB1E},
+{0xFE00, 0xFE0F}, {0xFE20, 0xFE2F}, {0x101FD, 0x101FD}, {0x102E0, 0x102E0}, {0x10376, 0x1037A}, {0x10A01, 0x10A03}, {0x10A05, 0x10A06}, {0x10A0C, 0x10A0F}, {0x10A38, 0x10A3A}, {0x10A3F, 0x10A3F},
+{0x10AE5, 0x10AE6}, {0x10D24, 0x10D27}, {0x10EAB, 0x10EAC}, {0x10F46, 0x10F50}, {0x11000, 0x11002}, {0x11038, 0x11046}, {0x1107F, 0x11082}, {0x110B0, 0x110BA}, {0x11100, 0x11102}, {0x11127, 0x11134},
+{0x11145, 0x11146}, {0x11173, 0x11173}, {0x11180, 0x11182}, {0x111B3, 0x111C0}, {0x111C9, 0x111CC}, {0x111CE, 0x111CF}, {0x1122C, 0x11237}, {0x1123E, 0x1123E}, {0x112DF, 0x112EA}, {0x11300, 0x11303},
+{0x1133B, 0x1133C}, {0x1133E, 0x11344}, {0x11347, 0x11348}, {0x1134B, 0x1134D}, {0x11357, 0x11357}, {0x11362, 0x11363}, {0x11366, 0x1136C}, {0x11370, 0x11374}, {0x11435, 0x11446}, {0x1145E, 0x1145E},
+{0x114B0, 0x114C3}, {0x115AF, 0x115B5}, {0x115B8, 0x115C0}, {0x115DC, 0x115DD}, {0x11630, 0x11640}, {0x116AB, 0x116B7}, {0x1171D, 0x1172B}, {0x1182C, 0x1183A}, {0x11930, 0x11935}, {0x11937, 0x11938},
+{0x1193B, 0x1193E}, {0x11940, 0x11940}, {0x11942, 0x11943}, {0x119D1, 0x119D7}, {0x119DA, 0x119E0}, {0x119E4, 0x119E4}, {0x11A01, 0x11A0A}, {0x11A33, 0x11A39}, {0x11A3B, 0x11A3E}, {0x11A47, 0x11A47},
+{0x11A51, 0x11A5B}, {0x11A8A, 0x11A99}, {0x11C2F, 0x11C36}, {0x11C38, 0x11C3F}, {0x11C92, 0x11CA7}, {0x11CA9, 0x11CB6}, {0x11D31, 0x11D36}, {0x11D3A, 0x11D3A}, {0x11D3C, 0x11D3D}, {0x11D3F, 0x11D45},
+{0x11D47, 0x11D47}, {0x11D8A, 0x11D8E}, {0x11D90, 0x11D91}, {0x11D93, 0x11D97}, {0x11EF3, 0x11EF6}, {0x16AF0, 0x16AF4}, {0x16B30, 0x16B36}, {0x16F4F, 0x16F4F}, {0x16F51, 0x16F87}, {0x16F8F, 0x16F92},
+{0x16FE4, 0x16FE4}, {0x16FF0, 0x16FF1}, {0x1BC9D, 0x1BC9E}, {0x1D165, 0x1D169}, {0x1D16D, 0x1D172}, {0x1D17B, 0x1D182}, {0x1D185, 0x1D18B}, {0x1D1AA, 0x1D1AD}, {0x1D242, 0x1D244}, {0x1DA00, 0x1DA36},
+{0x1DA3B, 0x1DA6C}, {0x1DA75, 0x1DA75}, {0x1DA84, 0x1DA84}, {0x1DA9B, 0x1DA9F}, {0x1DAA1, 0x1DAAF}, {0x1E000, 0x1E006}, {0x1E008, 0x1E018}, {0x1E01B, 0x1E021}, {0x1E023, 0x1E024}, {0x1E026, 0x1E02A},
+{0x1E130, 0x1E136}, {0x1E2EC, 0x1E2EF}, {0x1E8D0, 0x1E8D6}, {0x1E944, 0x1E94A}, {0xE0100, 0xE01EF},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> punctuation_ranges = {
+{0x21, 0x23}, {0x25, 0x2A}, {0x2C, 0x2F}, {0x3A, 0x3B}, {0x3F, 0x40}, {0x5B, 0x5D}, {0x5F, 0x5F}, {0x7B, 0x7B}, {0x7D, 0x7D}, {0xA1, 0xA1}, {0xA7, 0xA7}, {0xAB, 0xAB}, {0xB6, 0xB7}, {0xBB, 0xBB},
+{0xBF, 0xBF}, {0x37E, 0x37E}, {0x387, 0x387}, {0x55A, 0x55F}, {0x589, 0x58A}, {0x5BE, 0x5BE}, {0x5C0, 0x5C0}, {0x5C3, 0x5C3}, {0x5C6, 0x5C6}, {0x5F3, 0x5F4}, {0x609, 0x60A}, {0x60C, 0x60D},
+{0x61B, 0x61B}, {0x61E, 0x61F}, {0x66A, 0x66D}, {0x6D4, 0x6D4}, {0x700, 0x70D}, {0x7F7, 0x7F9}, {0x830, 0x83E}, {0x85E, 0x85E}, {0x964, 0x965}, {0x970, 0x970}, {0x9FD, 0x9FD}, {0xA76, 0xA76},
+{0xAF0, 0xAF0}, {0xC77, 0xC77}, {0xC84, 0xC84}, {0xDF4, 0xDF4}, {0xE4F, 0xE4F}, {0xE5A, 0xE5B}, {0xF04, 0xF12}, {0xF14, 0xF14}, {0xF3A, 0xF3D}, {0xF85, 0xF85}, {0xFD0, 0xFD4}, {0xFD9, 0xFDA},
+{0x104A, 0x104F}, {0x10FB, 0x10FB}, {0x1360, 0x1368}, {0x1400, 0x1400}, {0x166E, 0x166E}, {0x169B, 0x169C}, {0x16EB, 0x16ED}, {0x1735, 0x1736}, {0x17D4, 0x17D6}, {0x17D8, 0x17DA}, {0x1800, 0x180A},
+{0x1944, 0x1945}, {0x1A1E, 0x1A1F}, {0x1AA0, 0x1AA6}, {0x1AA8, 0x1AAD}, {0x1B5A, 0x1B60}, {0x1BFC, 0x1BFF}, {0x1C3B, 0x1C3F}, {0x1C7E, 0x1C7F}, {0x1CC0, 0x1CC7}, {0x1CD3, 0x1CD3}, {0x2010, 0x2027},
+{0x2030, 0x2043}, {0x2045, 0x2051}, {0x2053, 0x205E}, {0x207D, 0x207E}, {0x208D, 0x208E}, {0x2308, 0x230B}, {0x2329, 0x232A}, {0x2768, 0x2775}, {0x27C5, 0x27C6}, {0x27E6, 0x27EF}, {0x2983, 0x2998},
+{0x29D8, 0x29DB}, {0x29FC, 0x29FD}, {0x2CF9, 0x2CFC}, {0x2CFE, 0x2CFF}, {0x2D70, 0x2D70}, {0x2E00, 0x2E2E}, {0x2E30, 0x2E4F}, {0x2E52, 0x2E52}, {0x3001, 0x3003}, {0x3008, 0x3011}, {0x3014, 0x301F},
+{0x3030, 0x3030}, {0x303D, 0x303D}, {0x30A0, 0x30A0}, {0x30FB, 0x30FB}, {0xA4FE, 0xA4FF}, {0xA60D, 0xA60F}, {0xA673, 0xA673}, {0xA67E, 0xA67E}, {0xA6F2, 0xA6F7}, {0xA874, 0xA877}, {0xA8CE, 0xA8CF},
+{0xA8F8, 0xA8FA}, {0xA8FC, 0xA8FC}, {0xA92E, 0xA92F}, {0xA95F, 0xA95F}, {0xA9C1, 0xA9CD}, {0xA9DE, 0xA9DF}, {0xAA5C, 0xAA5F}, {0xAADE, 0xAADF}, {0xAAF0, 0xAAF1}, {0xABEB, 0xABEB}, {0xFD3E, 0xFD3F},
+{0xFE10, 0xFE19}, {0xFE30, 0xFE52}, {0xFE54, 0xFE61}, {0xFE63, 0xFE63}, {0xFE68, 0xFE68}, {0xFE6A, 0xFE6B}, {0xFF01, 0xFF03}, {0xFF05, 0xFF0A}, {0xFF0C, 0xFF0F}, {0xFF1A, 0xFF1B}, {0xFF1F, 0xFF20},
+{0xFF3B, 0xFF3D}, {0xFF3F, 0xFF3F}, {0xFF5B, 0xFF5B}, {0xFF5D, 0xFF5D}, {0xFF5F, 0xFF65}, {0x10100, 0x10102}, {0x1039F, 0x1039F}, {0x103D0, 0x103D0}, {0x1056F, 0x1056F}, {0x10857, 0x10857},
+{0x1091F, 0x1091F}, {0x1093F, 0x1093F}, {0x10A50, 0x10A58}, {0x10A7F, 0x10A7F}, {0x10AF0, 0x10AF6}, {0x10B39, 0x10B3F}, {0x10B99, 0x10B9C}, {0x10EAD, 0x10EAD}, {0x10F55, 0x10F59}, {0x11047, 0x1104D},
+{0x110BB, 0x110BC}, {0x110BE, 0x110C1}, {0x11140, 0x11143}, {0x11174, 0x11175}, {0x111C5, 0x111C8}, {0x111CD, 0x111CD}, {0x111DB, 0x111DB}, {0x111DD, 0x111DF}, {0x11238, 0x1123D}, {0x112A9, 0x112A9},
+{0x1144B, 0x1144F}, {0x1145A, 0x1145B}, {0x1145D, 0x1145D}, {0x114C6, 0x114C6}, {0x115C1, 0x115D7}, {0x11641, 0x11643}, {0x11660, 0x1166C}, {0x1173C, 0x1173E}, {0x1183B, 0x1183B}, {0x11944, 0x11946},
+{0x119E2, 0x119E2}, {0x11A3F, 0x11A46}, {0x11A9A, 0x11A9C}, {0x11A9E, 0x11AA2}, {0x11C41, 0x11C45}, {0x11C70, 0x11C71}, {0x11EF7, 0x11EF8}, {0x11FFF, 0x11FFF}, {0x12470, 0x12474}, {0x16A6E, 0x16A6F},
+{0x16AF5, 0x16AF5}, {0x16B37, 0x16B3B}, {0x16B44, 0x16B44}, {0x16E97, 0x16E9A}, {0x16FE2, 0x16FE2}, {0x1BC9F, 0x1BC9F}, {0x1DA87, 0x1DA8B}, {0x1E95E, 0x1E95F},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> symbol_ranges = {
+{0x24, 0x24}, {0x2B, 0x2B}, {0x3C, 0x3E}, {0x5E, 0x5E}, {0x60, 0x60}, {0x7C, 0x7C}, {0x7E, 0x7E}, {0xA2, 0xA6}, {0xA8, 0xA9}, {0xAC, 0xAC}, {0xAE, 0xB1}, {0xB4, 0xB4}, {0xB8, 0xB8}, {0xD7, 0xD7},
+{0xF7, 0xF7}, {0x2C2, 0x2C5}, {0x2D2, 0x2DF}, {0x2E5, 0x2EB}, {0x2ED, 0x2ED}, {0x2EF, 0x2FF}, {0x375, 0x375}, {0x384, 0x385}, {0x3F6, 0x3F6}, {0x482, 0x482}, {0x58D, 0x58F}, {0x606, 0x608},
+{0x60B, 0x60B}, {0x60E, 0x60F}, {0x6DE, 0x6DE}, {0x6E9, 0x6E9}, {0x6FD, 0x6FE}, {0x7F6, 0x7F6}, {0x7FE, 0x7FF}, {0x9F2, 0x9F3}, {0x9FA, 0x9FB}, {0xAF1, 0xAF1}, {0xB70, 0xB70}, {0xBF3, 0xBFA},
+{0xC7F, 0xC7F}, {0xD4F, 0xD4F}, {0xD79, 0xD79}, {0xE3F, 0xE3F}, {0xF01, 0xF03}, {0xF13, 0xF13}, {0xF15, 0xF17}, {0xF1A, 0xF1F}, {0xF34, 0xF34}, {0xF36, 0xF36}, {0xF38, 0xF38}, {0xFBE, 0xFC5},
+{0xFC7, 0xFCC}, {0xFCE, 0xFCF}, {0xFD5, 0xFD8}, {0x109E, 0x109F}, {0x1390, 0x1399}, {0x166D, 0x166D}, {0x17DB, 0x17DB}, {0x1940, 0x1940}, {0x19DE, 0x19FF}, {0x1B61, 0x1B6A}, {0x1B74, 0x1B7C},
+{0x1FBD, 0x1FBD}, {0x1FBF, 0x1FC1}, {0x1FCD, 0x1FCF}, {0x1FDD, 0x1FDF}, {0x1FED, 0x1FEF}, {0x1FFD, 0x1FFE}, {0x2044, 0x2044}, {0x2052, 0x2052}, {0x207A, 0x207C}, {0x208A, 0x208C}, {0x20A0, 0x20BF},
+{0x2100, 0x2101}, {0x2103, 0x2106}, {0x2108, 0x2109}, {0x2114, 0x2114}, {0x2116, 0x2118}, {0x211E, 0x2123}, {0x2125, 0x2125}, {0x2127, 0x2127}, {0x2129, 0x2129}, {0x212E, 0x212E}, {0x213A, 0x213B},
+{0x2140, 0x2144}, {0x214A, 0x214D}, {0x214F, 0x214F}, {0x218A, 0x218B}, {0x2190, 0x2307}, {0x230C, 0x2328}, {0x232B, 0x2426}, {0x2440, 0x244A}, {0x249C, 0x24E9}, {0x2500, 0x2767}, {0x2794, 0x27C4},
+{0x27C7, 0x27E5}, {0x27F0, 0x2982}, {0x2999, 0x29D7}, {0x29DC, 0x29FB}, {0x29FE, 0x2B73}, {0x2B76, 0x2B95}, {0x2B97, 0x2BFF}, {0x2CE5, 0x2CEA}, {0x2E50, 0x2E51}, {0x2E80, 0x2E99}, {0x2E9B, 0x2EF3},
+{0x2F00, 0x2FD5}, {0x2FF0, 0x2FFB}, {0x3004, 0x3004}, {0x3012, 0x3013}, {0x3020, 0x3020}, {0x3036, 0x3037}, {0x303E, 0x303F}, {0x309B, 0x309C}, {0x3190, 0x3191}, {0x3196, 0x319F}, {0x31C0, 0x31E3},
+{0x3200, 0x321E}, {0x322A, 0x3247}, {0x3250, 0x3250}, {0x3260, 0x327F}, {0x328A, 0x32B0}, {0x32C0, 0x33FF}, {0x4DC0, 0x4DFF}, {0xA490, 0xA4C6}, {0xA700, 0xA716}, {0xA720, 0xA721}, {0xA789, 0xA78A},
+{0xA828, 0xA82B}, {0xA836, 0xA839}, {0xAA77, 0xAA79}, {0xAB5B, 0xAB5B}, {0xAB6A, 0xAB6B}, {0xFB29, 0xFB29}, {0xFBB2, 0xFBC1}, {0xFDFC, 0xFDFD}, {0xFE62, 0xFE62}, {0xFE64, 0xFE66}, {0xFE69, 0xFE69},
+{0xFF04, 0xFF04}, {0xFF0B, 0xFF0B}, {0xFF1C, 0xFF1E}, {0xFF3E, 0xFF3E}, {0xFF40, 0xFF40}, {0xFF5C, 0xFF5C}, {0xFF5E, 0xFF5E}, {0xFFE0, 0xFFE6}, {0xFFE8, 0xFFEE}, {0xFFFC, 0xFFFD}, {0x10137, 0x1013F},
+{0x10179, 0x10189}, {0x1018C, 0x1018E}, {0x10190, 0x1019C}, {0x101A0, 0x101A0}, {0x101D0, 0x101FC}, {0x10877, 0x10878}, {0x10AC8, 0x10AC8}, {0x1173F, 0x1173F}, {0x11FD5, 0x11FF1}, {0x16B3C, 0x16B3F},
+{0x16B45, 0x16B45}, {0x1BC9C, 0x1BC9C}, {0x1D000, 0x1D0F5}, {0x1D100, 0x1D126}, {0x1D129, 0x1D164}, {0x1D16A, 0x1D16C}, {0x1D183, 0x1D184}, {0x1D18C, 0x1D1A9}, {0x1D1AE, 0x1D1E8}, {0x1D200, 0x1D241},
+{0x1D245, 0x1D245}, {0x1D300, 0x1D356}, {0x1D6C1, 0x1D6C1}, {0x1D6DB, 0x1D6DB}, {0x1D6FB, 0x1D6FB}, {0x1D715, 0x1D715}, {0x1D735, 0x1D735}, {0x1D74F, 0x1D74F}, {0x1D76F, 0x1D76F}, {0x1D789, 0x1D789},
+{0x1D7A9, 0x1D7A9}, {0x1D7C3, 0x1D7C3}, {0x1D800, 0x1D9FF}, {0x1DA37, 0x1DA3A}, {0x1DA6D, 0x1DA74}, {0x1DA76, 0x1DA83}, {0x1DA85, 0x1DA86}, {0x1E14F, 0x1E14F}, {0x1E2FF, 0x1E2FF}, {0x1ECAC, 0x1ECAC},
+{0x1ECB0, 0x1ECB0}, {0x1ED2E, 0x1ED2E}, {0x1EEF0, 0x1EEF1}, {0x1F000, 0x1F02B}, {0x1F030, 0x1F093}, {0x1F0A0, 0x1F0AE}, {0x1F0B1, 0x1F0BF}, {0x1F0C1, 0x1F0CF}, {0x1F0D1, 0x1F0F5}, {0x1F10D, 0x1F1AD},
+{0x1F1E6, 0x1F202}, {0x1F210, 0x1F23B}, {0x1F240, 0x1F248}, {0x1F250, 0x1F251}, {0x1F260, 0x1F265}, {0x1F300, 0x1F6D7}, {0x1F6E0, 0x1F6EC}, {0x1F6F0, 0x1F6FC}, {0x1F700, 0x1F773}, {0x1F780, 0x1F7D8},
+{0x1F7E0, 0x1F7EB}, {0x1F800, 0x1F80B}, {0x1F810, 0x1F847}, {0x1F850, 0x1F859}, {0x1F860, 0x1F887}, {0x1F890, 0x1F8AD}, {0x1F8B0, 0x1F8B1}, {0x1F900, 0x1F978}, {0x1F97A, 0x1F9CB}, {0x1F9CD, 0x1FA53},
+{0x1FA60, 0x1FA6D}, {0x1FA70, 0x1FA74}, {0x1FA78, 0x1FA7A}, {0x1FA80, 0x1FA86}, {0x1FA90, 0x1FAA8}, {0x1FAB0, 0x1FAB6}, {0x1FAC0, 0x1FAC2}, {0x1FAD0, 0x1FAD6}, {0x1FB00, 0x1FB92}, {0x1FB94, 0x1FBCA},
+};
+
+static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
+{0x0, 0x8}, {0xE, 0x1B}, {0x7F, 0x84}, {0x86, 0x9F}, {0xAD, 0xAD}, {0x378, 0x379}, {0x380, 0x383}, {0x38B, 0x38B}, {0x38D, 0x38D}, {0x3A2, 0x3A2}, {0x530, 0x530}, {0x557, 0x558}, {0x58B, 0x58C},
+{0x590, 0x590}, {0x5C8, 0x5CF}, {0x5EB, 0x5EE}, {0x5F5, 0x605}, {0x61C, 0x61D}, {0x6DD, 0x6DD}, {0x70E, 0x70F}, {0x74B, 0x74C}, {0x7B2, 0x7BF}, {0x7FB, 0x7FC}, {0x82E, 0x82F}, {0x83F, 0x83F},
+{0x85C, 0x85D}, {0x85F, 0x85F}, {0x86B, 0x89F}, {0x8B5, 0x8B5}, {0x8C8, 0x8D2}, {0x8E2, 0x8E2}, {0x984, 0x984}, {0x98D, 0x98E}, {0x991, 0x992}, {0x9A9, 0x9A9}, {0x9B1, 0x9B1}, {0x9B3, 0x9B5},
+{0x9BA, 0x9BB}, {0x9C5, 0x9C6}, {0x9C9, 0x9CA}, {0x9CF, 0x9D6}, {0x9D8, 0x9DB}, {0x9DE, 0x9DE}, {0x9E4, 0x9E5}, {0x9FF, 0xA00}, {0xA04, 0xA04}, {0xA0B, 0xA0E}, {0xA11, 0xA12}, {0xA29, 0xA29},
+{0xA31, 0xA31}, {0xA34, 0xA34}, {0xA37, 0xA37}, {0xA3A, 0xA3B}, {0xA3D, 0xA3D}, {0xA43, 0xA46}, {0xA49, 0xA4A}, {0xA4E, 0xA50}, {0xA52, 0xA58}, {0xA5D, 0xA5D}, {0xA5F, 0xA65}, {0xA77, 0xA80},
+{0xA84, 0xA84}, {0xA8E, 0xA8E}, {0xA92, 0xA92}, {0xAA9, 0xAA9}, {0xAB1, 0xAB1}, {0xAB4, 0xAB4}, {0xABA, 0xABB}, {0xAC6, 0xAC6}, {0xACA, 0xACA}, {0xACE, 0xACF}, {0xAD1, 0xADF}, {0xAE4, 0xAE5},
+{0xAF2, 0xAF8}, {0xB00, 0xB00}, {0xB04, 0xB04}, {0xB0D, 0xB0E}, {0xB11, 0xB12}, {0xB29, 0xB29}, {0xB31, 0xB31}, {0xB34, 0xB34}, {0xB3A, 0xB3B}, {0xB45, 0xB46}, {0xB49, 0xB4A}, {0xB4E, 0xB54},
+{0xB58, 0xB5B}, {0xB5E, 0xB5E}, {0xB64, 0xB65}, {0xB78, 0xB81}, {0xB84, 0xB84}, {0xB8B, 0xB8D}, {0xB91, 0xB91}, {0xB96, 0xB98}, {0xB9B, 0xB9B}, {0xB9D, 0xB9D}, {0xBA0, 0xBA2}, {0xBA5, 0xBA7},
+{0xBAB, 0xBAD}, {0xBBA, 0xBBD}, {0xBC3, 0xBC5}, {0xBC9, 0xBC9}, {0xBCE, 0xBCF}, {0xBD1, 0xBD6}, {0xBD8, 0xBE5}, {0xBFB, 0xBFF}, {0xC0D, 0xC0D}, {0xC11, 0xC11}, {0xC29, 0xC29}, {0xC3A, 0xC3C},
+{0xC45, 0xC45}, {0xC49, 0xC49}, {0xC4E, 0xC54}, {0xC57, 0xC57}, {0xC5B, 0xC5F}, {0xC64, 0xC65}, {0xC70, 0xC76}, {0xC8D, 0xC8D}, {0xC91, 0xC91}, {0xCA9, 0xCA9}, {0xCB4, 0xCB4}, {0xCBA, 0xCBB},
+{0xCC5, 0xCC5}, {0xCC9, 0xCC9}, {0xCCE, 0xCD4}, {0xCD7, 0xCDD}, {0xCDF, 0xCDF}, {0xCE4, 0xCE5}, {0xCF0, 0xCF0}, {0xCF3, 0xCFF}, {0xD0D, 0xD0D}, {0xD11, 0xD11}, {0xD45, 0xD45}, {0xD49, 0xD49},
+{0xD50, 0xD53}, {0xD64, 0xD65}, {0xD80, 0xD80}, {0xD84, 0xD84}, {0xD97, 0xD99}, {0xDB2, 0xDB2}, {0xDBC, 0xDBC}, {0xDBE, 0xDBF}, {0xDC7, 0xDC9}, {0xDCB, 0xDCE}, {0xDD5, 0xDD5}, {0xDD7, 0xDD7},
+{0xDE0, 0xDE5}, {0xDF0, 0xDF1}, {0xDF5, 0xE00}, {0xE3B, 0xE3E}, {0xE5C, 0xE80}, {0xE83, 0xE83}, {0xE85, 0xE85}, {0xE8B, 0xE8B}, {0xEA4, 0xEA4}, {0xEA6, 0xEA6}, {0xEBE, 0xEBF}, {0xEC5, 0xEC5},
+{0xEC7, 0xEC7}, {0xECE, 0xECF}, {0xEDA, 0xEDB}, {0xEE0, 0xEFF}, {0xF48, 0xF48}, {0xF6D, 0xF70}, {0xF98, 0xF98}, {0xFBD, 0xFBD}, {0xFCD, 0xFCD}, {0xFDB, 0xFFF}, {0x10C6, 0x10C6}, {0x10C8, 0x10CC},
+{0x10CE, 0x10CF}, {0x1249, 0x1249}, {0x124E, 0x124F}, {0x1257, 0x1257}, {0x1259, 0x1259}, {0x125E, 0x125F}, {0x1289, 0x1289}, {0x128E, 0x128F}, {0x12B1, 0x12B1}, {0x12B6, 0x12B7}, {0x12BF, 0x12BF},
+{0x12C1, 0x12C1}, {0x12C6, 0x12C7}, {0x12D7, 0x12D7}, {0x1311, 0x1311}, {0x1316, 0x1317}, {0x135B, 0x135C}, {0x137D, 0x137F}, {0x139A, 0x139F}, {0x13F6, 0x13F7}, {0x13FE, 0x13FF}, {0x169D, 0x169F},
+{0x16F9, 0x16FF}, {0x170D, 0x170D}, {0x1715, 0x171F}, {0x1737, 0x173F}, {0x1754, 0x175F}, {0x176D, 0x176D}, {0x1771, 0x1771}, {0x1774, 0x177F}, {0x17DE, 0x17DF}, {0x17EA, 0x17EF}, {0x17FA, 0x17FF},
+{0x180E, 0x180F}, {0x181A, 0x181F}, {0x1879, 0x187F}, {0x18AB, 0x18AF}, {0x18F6, 0x18FF}, {0x191F, 0x191F}, {0x192C, 0x192F}, {0x193C, 0x193F}, {0x1941, 0x1943}, {0x196E, 0x196F}, {0x1975, 0x197F},
+{0x19AC, 0x19AF}, {0x19CA, 0x19CF}, {0x19DB, 0x19DD}, {0x1A1C, 0x1A1D}, {0x1A5F, 0x1A5F}, {0x1A7D, 0x1A7E}, {0x1A8A, 0x1A8F}, {0x1A9A, 0x1A9F}, {0x1AAE, 0x1AAF}, {0x1AC1, 0x1AFF}, {0x1B4C, 0x1B4F},
+{0x1B7D, 0x1B7F}, {0x1BF4, 0x1BFB}, {0x1C38, 0x1C3A}, {0x1C4A, 0x1C4C}, {0x1C89, 0x1C8F}, {0x1CBB, 0x1CBC}, {0x1CC8, 0x1CCF}, {0x1CFB, 0x1CFF}, {0x1DFA, 0x1DFA}, {0x1F16, 0x1F17}, {0x1F1E, 0x1F1F},
+{0x1F46, 0x1F47}, {0x1F4E, 0x1F4F}, {0x1F58, 0x1F58}, {0x1F5A, 0x1F5A}, {0x1F5C, 0x1F5C}, {0x1F5E, 0x1F5E}, {0x1F7E, 0x1F7F}, {0x1FB5, 0x1FB5}, {0x1FC5, 0x1FC5}, {0x1FD4, 0x1FD5}, {0x1FDC, 0x1FDC},
+{0x1FF0, 0x1FF1}, {0x1FF5, 0x1FF5}, {0x1FFF, 0x1FFF}, {0x200B, 0x200F}, {0x202A, 0x202E}, {0x2060, 0x206F}, {0x2072, 0x2073}, {0x208F, 0x208F}, {0x209D, 0x209F}, {0x20C0, 0x20CF}, {0x20F1, 0x20FF},
+{0x218C, 0x218F}, {0x2427, 0x243F}, {0x244B, 0x245F}, {0x2B74, 0x2B75}, {0x2B96, 0x2B96}, {0x2C2F, 0x2C2F}, {0x2C5F, 0x2C5F}, {0x2CF4, 0x2CF8}, {0x2D26, 0x2D26}, {0x2D28, 0x2D2C}, {0x2D2E, 0x2D2F},
+{0x2D68, 0x2D6E}, {0x2D71, 0x2D7E}, {0x2D97, 0x2D9F}, {0x2DA7, 0x2DA7}, {0x2DAF, 0x2DAF}, {0x2DB7, 0x2DB7}, {0x2DBF, 0x2DBF}, {0x2DC7, 0x2DC7}, {0x2DCF, 0x2DCF}, {0x2DD7, 0x2DD7}, {0x2DDF, 0x2DDF},
+{0x2E53, 0x2E7F}, {0x2E9A, 0x2E9A}, {0x2EF4, 0x2EFF}, {0x2FD6, 0x2FEF}, {0x2FFC, 0x2FFF}, {0x3040, 0x3040}, {0x3097, 0x3098}, {0x3100, 0x3104}, {0x3130, 0x3130}, {0x318F, 0x318F}, {0x31E4, 0x31EF},
+{0x321F, 0x321F}, {0x9FFD, 0x9FFF}, {0xA48D, 0xA48F}, {0xA4C7, 0xA4CF}, {0xA62C, 0xA63F}, {0xA6F8, 0xA6FF}, {0xA7C0, 0xA7C1}, {0xA7CB, 0xA7F4}, {0xA82D, 0xA82F}, {0xA83A, 0xA83F}, {0xA878, 0xA87F},
+{0xA8C6, 0xA8CD}, {0xA8DA, 0xA8DF}, {0xA954, 0xA95E}, {0xA97D, 0xA97F}, {0xA9CE, 0xA9CE}, {0xA9DA, 0xA9DD}, {0xA9FF, 0xA9FF}, {0xAA37, 0xAA3F}, {0xAA4E, 0xAA4F}, {0xAA5A, 0xAA5B}, {0xAAC3, 0xAADA},
+{0xAAF7, 0xAB00}, {0xAB07, 0xAB08}, {0xAB0F, 0xAB10}, {0xAB17, 0xAB1F}, {0xAB27, 0xAB27}, {0xAB2F, 0xAB2F}, {0xAB6C, 0xAB6F}, {0xABEE, 0xABEF}, {0xABFA, 0xABFF}, {0xD7A4, 0xD7AF}, {0xD7C7, 0xD7CA},
+{0xD7FC, 0xF8FF}, {0xFA6E, 0xFA6F}, {0xFADA, 0xFAFF}, {0xFB07, 0xFB12}, {0xFB18, 0xFB1C}, {0xFB37, 0xFB37}, {0xFB3D, 0xFB3D}, {0xFB3F, 0xFB3F}, {0xFB42, 0xFB42}, {0xFB45, 0xFB45}, {0xFBC2, 0xFBD2},
+{0xFD40, 0xFD4F}, {0xFD90, 0xFD91}, {0xFDC8, 0xFDEF}, {0xFDFE, 0xFDFF}, {0xFE1A, 0xFE1F}, {0xFE53, 0xFE53}, {0xFE67, 0xFE67}, {0xFE6C, 0xFE6F}, {0xFE75, 0xFE75}, {0xFEFD, 0xFF00}, {0xFFBF, 0xFFC1},
+{0xFFC8, 0xFFC9}, {0xFFD0, 0xFFD1}, {0xFFD8, 0xFFD9}, {0xFFDD, 0xFFDF}, {0xFFE7, 0xFFE7}, {0xFFEF, 0xFFFB}, {0xFFFE, 0xFFFF}, {0x1000C, 0x1000C}, {0x10027, 0x10027}, {0x1003B, 0x1003B},
+{0x1003E, 0x1003E}, {0x1004E, 0x1004F}, {0x1005E, 0x1007F}, {0x100FB, 0x100FF}, {0x10103, 0x10106}, {0x10134, 0x10136}, {0x1018F, 0x1018F}, {0x1019D, 0x1019F}, {0x101A1, 0x101CF}, {0x101FE, 0x1027F},
+{0x1029D, 0x1029F}, {0x102D1, 0x102DF}, {0x102FC, 0x102FF}, {0x10324, 0x1032C}, {0x1034B, 0x1034F}, {0x1037B, 0x1037F}, {0x1039E, 0x1039E}, {0x103C4, 0x103C7}, {0x103D6, 0x103FF}, {0x1049E, 0x1049F},
+{0x104AA, 0x104AF}, {0x104D4, 0x104D7}, {0x104FC, 0x104FF}, {0x10528, 0x1052F}, {0x10564, 0x1056E}, {0x10570, 0x105FF}, {0x10737, 0x1073F}, {0x10756, 0x1075F}, {0x10768, 0x107FF}, {0x10806, 0x10807},
+{0x10809, 0x10809}, {0x10836, 0x10836}, {0x10839, 0x1083B}, {0x1083D, 0x1083E}, {0x10856, 0x10856}, {0x1089F, 0x108A6}, {0x108B0, 0x108DF}, {0x108F3, 0x108F3}, {0x108F6, 0x108FA}, {0x1091C, 0x1091E},
+{0x1093A, 0x1093E}, {0x10940, 0x1097F}, {0x109B8, 0x109BB}, {0x109D0, 0x109D1}, {0x10A04, 0x10A04}, {0x10A07, 0x10A0B}, {0x10A14, 0x10A14}, {0x10A18, 0x10A18}, {0x10A36, 0x10A37}, {0x10A3B, 0x10A3E},
+{0x10A49, 0x10A4F}, {0x10A59, 0x10A5F}, {0x10AA0, 0x10ABF}, {0x10AE7, 0x10AEA}, {0x10AF7, 0x10AFF}, {0x10B36, 0x10B38}, {0x10B56, 0x10B57}, {0x10B73, 0x10B77}, {0x10B92, 0x10B98}, {0x10B9D, 0x10BA8},
+{0x10BB0, 0x10BFF}, {0x10C49, 0x10C7F}, {0x10CB3, 0x10CBF}, {0x10CF3, 0x10CF9}, {0x10D28, 0x10D2F}, {0x10D3A, 0x10E5F}, {0x10E7F, 0x10E7F}, {0x10EAA, 0x10EAA}, {0x10EAE, 0x10EAF}, {0x10EB2, 0x10EFF},
+{0x10F28, 0x10F2F}, {0x10F5A, 0x10FAF}, {0x10FCC, 0x10FDF}, {0x10FF7, 0x10FFF}, {0x1104E, 0x11051}, {0x11070, 0x1107E}, {0x110BD, 0x110BD}, {0x110C2, 0x110CF}, {0x110E9, 0x110EF}, {0x110FA, 0x110FF},
+{0x11135, 0x11135}, {0x11148, 0x1114F}, {0x11177, 0x1117F}, {0x111E0, 0x111E0}, {0x111F5, 0x111FF}, {0x11212, 0x11212}, {0x1123F, 0x1127F}, {0x11287, 0x11287}, {0x11289, 0x11289}, {0x1128E, 0x1128E},
+{0x1129E, 0x1129E}, {0x112AA, 0x112AF}, {0x112EB, 0x112EF}, {0x112FA, 0x112FF}, {0x11304, 0x11304}, {0x1130D, 0x1130E}, {0x11311, 0x11312}, {0x11329, 0x11329}, {0x11331, 0x11331}, {0x11334, 0x11334},
+{0x1133A, 0x1133A}, {0x11345, 0x11346}, {0x11349, 0x1134A}, {0x1134E, 0x1134F}, {0x11351, 0x11356}, {0x11358, 0x1135C}, {0x11364, 0x11365}, {0x1136D, 0x1136F}, {0x11375, 0x113FF}, {0x1145C, 0x1145C},
+{0x11462, 0x1147F}, {0x114C8, 0x114CF}, {0x114DA, 0x1157F}, {0x115B6, 0x115B7}, {0x115DE, 0x115FF}, {0x11645, 0x1164F}, {0x1165A, 0x1165F}, {0x1166D, 0x1167F}, {0x116B9, 0x116BF}, {0x116CA, 0x116FF},
+{0x1171B, 0x1171C}, {0x1172C, 0x1172F}, {0x11740, 0x117FF}, {0x1183C, 0x1189F}, {0x118F3, 0x118FE}, {0x11907, 0x11908}, {0x1190A, 0x1190B}, {0x11914, 0x11914}, {0x11917, 0x11917}, {0x11936, 0x11936},
+{0x11939, 0x1193A}, {0x11947, 0x1194F}, {0x1195A, 0x1199F}, {0x119A8, 0x119A9}, {0x119D8, 0x119D9}, {0x119E5, 0x119FF}, {0x11A48, 0x11A4F}, {0x11AA3, 0x11ABF}, {0x11AF9, 0x11BFF}, {0x11C09, 0x11C09},
+{0x11C37, 0x11C37}, {0x11C46, 0x11C4F}, {0x11C6D, 0x11C6F}, {0x11C90, 0x11C91}, {0x11CA8, 0x11CA8}, {0x11CB7, 0x11CFF}, {0x11D07, 0x11D07}, {0x11D0A, 0x11D0A}, {0x11D37, 0x11D39}, {0x11D3B, 0x11D3B},
+{0x11D3E, 0x11D3E}, {0x11D48, 0x11D4F}, {0x11D5A, 0x11D5F}, {0x11D66, 0x11D66}, {0x11D69, 0x11D69}, {0x11D8F, 0x11D8F}, {0x11D92, 0x11D92}, {0x11D99, 0x11D9F}, {0x11DAA, 0x11EDF}, {0x11EF9, 0x11FAF},
+{0x11FB1, 0x11FBF}, {0x11FF2, 0x11FFE}, {0x1239A, 0x123FF}, {0x1246F, 0x1246F}, {0x12475, 0x1247F}, {0x12544, 0x12FFF}, {0x1342F, 0x143FF}, {0x14647, 0x167FF}, {0x16A39, 0x16A3F}, {0x16A5F, 0x16A5F},
+{0x16A6A, 0x16A6D}, {0x16A70, 0x16ACF}, {0x16AEE, 0x16AEF}, {0x16AF6, 0x16AFF}, {0x16B46, 0x16B4F}, {0x16B5A, 0x16B5A}, {0x16B62, 0x16B62}, {0x16B78, 0x16B7C}, {0x16B90, 0x16E3F}, {0x16E9B, 0x16EFF},
+{0x16F4B, 0x16F4E}, {0x16F88, 0x16F8E}, {0x16FA0, 0x16FDF}, {0x16FE5, 0x16FEF}, {0x16FF2, 0x16FFF}, {0x187F8, 0x187FF}, {0x18CD6, 0x18CFF}, {0x18D09, 0x1AFFF}, {0x1B11F, 0x1B14F}, {0x1B153, 0x1B163},
+{0x1B168, 0x1B16F}, {0x1B2FC, 0x1BBFF}, {0x1BC6B, 0x1BC6F}, {0x1BC7D, 0x1BC7F}, {0x1BC89, 0x1BC8F}, {0x1BC9A, 0x1BC9B}, {0x1BCA0, 0x1CFFF}, {0x1D0F6, 0x1D0FF}, {0x1D127, 0x1D128}, {0x1D173, 0x1D17A},
+{0x1D1E9, 0x1D1FF}, {0x1D246, 0x1D2DF}, {0x1D2F4, 0x1D2FF}, {0x1D357, 0x1D35F}, {0x1D379, 0x1D3FF}, {0x1D455, 0x1D455}, {0x1D49D, 0x1D49D}, {0x1D4A0, 0x1D4A1}, {0x1D4A3, 0x1D4A4}, {0x1D4A7, 0x1D4A8},
+{0x1D4AD, 0x1D4AD}, {0x1D4BA, 0x1D4BA}, {0x1D4BC, 0x1D4BC}, {0x1D4C4, 0x1D4C4}, {0x1D506, 0x1D506}, {0x1D50B, 0x1D50C}, {0x1D515, 0x1D515}, {0x1D51D, 0x1D51D}, {0x1D53A, 0x1D53A}, {0x1D53F, 0x1D53F},
+{0x1D545, 0x1D545}, {0x1D547, 0x1D549}, {0x1D551, 0x1D551}, {0x1D6A6, 0x1D6A7}, {0x1D7CC, 0x1D7CD}, {0x1DA8C, 0x1DA9A}, {0x1DAA0, 0x1DAA0}, {0x1DAB0, 0x1DFFF}, {0x1E007, 0x1E007}, {0x1E019, 0x1E01A},
+{0x1E022, 0x1E022}, {0x1E025, 0x1E025}, {0x1E02B, 0x1E0FF}, {0x1E12D, 0x1E12F}, {0x1E13E, 0x1E13F}, {0x1E14A, 0x1E14D}, {0x1E150, 0x1E2BF}, {0x1E2FA, 0x1E2FE}, {0x1E300, 0x1E7FF}, {0x1E8C5, 0x1E8C6},
+{0x1E8D7, 0x1E8FF}, {0x1E94C, 0x1E94F}, {0x1E95A, 0x1E95D}, {0x1E960, 0x1EC70}, {0x1ECB5, 0x1ED00}, {0x1ED3E, 0x1EDFF}, {0x1EE04, 0x1EE04}, {0x1EE20, 0x1EE20}, {0x1EE23, 0x1EE23}, {0x1EE25, 0x1EE26},
+{0x1EE28, 0x1EE28}, {0x1EE33, 0x1EE33}, {0x1EE38, 0x1EE38}, {0x1EE3A, 0x1EE3A}, {0x1EE3C, 0x1EE41}, {0x1EE43, 0x1EE46}, {0x1EE48, 0x1EE48}, {0x1EE4A, 0x1EE4A}, {0x1EE4C, 0x1EE4C}, {0x1EE50, 0x1EE50},
+{0x1EE53, 0x1EE53}, {0x1EE55, 0x1EE56}, {0x1EE58, 0x1EE58}, {0x1EE5A, 0x1EE5A}, {0x1EE5C, 0x1EE5C}, {0x1EE5E, 0x1EE5E}, {0x1EE60, 0x1EE60}, {0x1EE63, 0x1EE63}, {0x1EE65, 0x1EE66}, {0x1EE6B, 0x1EE6B},
+{0x1EE73, 0x1EE73}, {0x1EE78, 0x1EE78}, {0x1EE7D, 0x1EE7D}, {0x1EE7F, 0x1EE7F}, {0x1EE8A, 0x1EE8A}, {0x1EE9C, 0x1EEA0}, {0x1EEA4, 0x1EEA4}, {0x1EEAA, 0x1EEAA}, {0x1EEBC, 0x1EEEF}, {0x1EEF2, 0x1EFFF},
+{0x1F02C, 0x1F02F}, {0x1F094, 0x1F09F}, {0x1F0AF, 0x1F0B0}, {0x1F0C0, 0x1F0C0}, {0x1F0D0, 0x1F0D0}, {0x1F0F6, 0x1F0FF}, {0x1F1AE, 0x1F1E5}, {0x1F203, 0x1F20F}, {0x1F23C, 0x1F23F}, {0x1F249, 0x1F24F},
+{0x1F252, 0x1F25F}, {0x1F266, 0x1F2FF}, {0x1F6D8, 0x1F6DF}, {0x1F6ED, 0x1F6EF}, {0x1F6FD, 0x1F6FF}, {0x1F774, 0x1F77F}, {0x1F7D9, 0x1F7DF}, {0x1F7EC, 0x1F7FF}, {0x1F80C, 0x1F80F}, {0x1F848, 0x1F84F},
+{0x1F85A, 0x1F85F}, {0x1F888, 0x1F88F}, {0x1F8AE, 0x1F8AF}, {0x1F8B2, 0x1F8FF}, {0x1F979, 0x1F979}, {0x1F9CC, 0x1F9CC}, {0x1FA54, 0x1FA5F}, {0x1FA6E, 0x1FA6F}, {0x1FA75, 0x1FA77}, {0x1FA7B, 0x1FA7F},
+{0x1FA87, 0x1FA8F}, {0x1FAA9, 0x1FAAF}, {0x1FAB7, 0x1FABF}, {0x1FAC3, 0x1FACF}, {0x1FAD7, 0x1FAFF}, {0x1FB93, 0x1FB93}, {0x1FBCB, 0x1FBEF}, {0x1FBFA, 0x1FFFF}, {0x2A6DE, 0x2A6FF}, {0x2B735, 0x2B73F},
+{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
+};
+
+static std::string codepoint_to_utf8(uint32_t cp) {
+    std::string result;
+    if (/* 0x00 <= cp && */ cp <= 0x7f) {
+        result.push_back(cp);
+    }
+    else if (0x80 <= cp && cp <= 0x7ff) {
+        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x800 <= cp && cp <= 0xffff) {
+        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.push_back(0xf0 | ((cp >> 18) & 0x07));
+        result.push_back(0x80 | ((cp >> 12) & 0x3f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::string codepoints_to_utf8(const std::vector<uint32_t> & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(codepoint_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
+    assert(offset < utf8.size());
+    if (!(utf8[offset + 0] & 0x80)) {
+        auto result = utf8[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x40)) {
+        throw std::invalid_argument("invalid character");
+    }
+    else if (!(utf8[offset + 0] & 0x20)) {
+        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
+        offset += 2;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x10)) {
+        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
+        offset += 3;
+        return result;
+    }
+    else if (!(utf8[offset + 0] & 0x08)) {
+        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
+            throw std::invalid_argument("invalid character");
+        auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
+        offset += 4;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf8(const std::string & utf8) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf8.size()) {
+        result.push_back(codepoint_from_utf8(utf8, offset));
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoint_to_utf16(uint32_t cp) {
+    std::vector<uint16_t> result;
+    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
+        result.emplace_back(cp);
+    }
+    else if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
+        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+    }
+    else {
+        throw std::invalid_argument("invalid codepoint");
+    }
+    return result;
+}
+
+static std::vector<uint16_t> codepoints_to_utf16(const std::vector<uint32_t> & cps) {
+    std::vector<uint16_t> result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        auto temp = codepoint_to_utf16(cps[i]);
+        result.insert(result.end(), temp.begin(), temp.end());
+    }
+    return result;
+}
+
+static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
+    assert(offset < utf16.size());
+    if (((utf16[0] >> 10) << 10) != 0xd800) {
+        auto result = utf16[offset + 0];
+        offset += 1;
+        return result;
+    }
+    else {
+        if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
+            throw std::invalid_argument("invalid character");
+        auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
+        offset += 2;
+        return result;
+    }
+    throw std::invalid_argument("invalid string");
+}
+
+static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf16.size())
+        result.push_back(codepoint_from_utf16(utf16, offset));
+    return result;
+}
+
+#define CODEPOINT_TYPE_UNIDENTIFIED 0
+#define CODEPOINT_TYPE_DIGIT 1
+#define CODEPOINT_TYPE_LETTER 2
+#define CODEPOINT_TYPE_WHITESPACE 3
+#define CODEPOINT_TYPE_ACCENT_MARK 4
+#define CODEPOINT_TYPE_PUNCTUATION 5
+#define CODEPOINT_TYPE_SYMBOL 6
+#define CODEPOINT_TYPE_CONTROL 7
+
+static std::unordered_map<uint32_t, int> codepoint_type_map() {
+    std::unordered_map<uint32_t, int> codepoint_types;
+    for (auto p : digit_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
+    }
+    for(auto p : letter_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_LETTER;
+    }
+    for(auto p : whitespace_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
+    }
+    for(auto p : accent_mark_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
+    }
+    for(auto p : punctuation_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
+    }
+    for (auto p : symbol_ranges) {
+        for (auto i = p.first; i <= p.second; ++i)
+            codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
+    }
+    for(auto p : control_ranges) {
+        for(auto i = p.first; i <= p.second; ++ i)
+            codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
+    }
+    return codepoint_types;
+}
+
+static int codepoint_type(uint32_t cp) {
+    static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
+    return codepoint_types[cp];
+}
+
+static int codepoint_type(const std::string & utf8) {
+    if (utf8.length() == 0)
+        return CODEPOINT_TYPE_UNIDENTIFIED;
+    size_t offset = 0;
+    return codepoint_type(codepoint_from_utf8(utf8, offset));
+}
+
+static std::unordered_map<uint8_t, std::string> bytes_to_unicode_map_bpe() {
+    std::unordered_map<uint8_t, std::string> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = codepoint_to_utf8(ch);
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(ch) == map.end()) {
+            map[ch] = codepoint_to_utf8(256 + n);
+            ++n;
+        }
+    }
+    return map;
+}
+
+static std::string bytes_to_unicode_bpe(uint8_t byte) {
+    static std::unordered_map<uint8_t, std::string> map = bytes_to_unicode_map_bpe();
+    return map.at(byte);
+}
+
+static std::unordered_map<std::string, uint8_t> unicode_to_bytes_map_bpe() {
+    std::unordered_map<std::string, uint8_t> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[codepoint_to_utf8(ch)] = ch;
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(codepoint_to_utf8(ch)) == map.end()) {
+            map[codepoint_to_utf8(256 + n)] = ch;
+            ++n;
+        }
+    }
+    return map;
+}
+
+static uint8_t unicode_to_bytes_bpe(const std::string & utf8) {
+    static std::unordered_map<std::string, uint8_t> map = unicode_to_bytes_map_bpe();
+    return map.at(utf8);
+}
+

examples/talk.wasm/emscripten.cpp

@@ -271,7 +271,7 @@ EMSCRIPTEN_BINDINGS(talk) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/talk.wasm/gpt-2.cpp

@@ -155,33 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
+        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
 
-        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
-        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
+        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
+        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
 
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
 
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
+        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -191,9 +191,9 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
     // create the ggml context
     {
         struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc   = false,
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
         };
 
         model.ctx = ggml_init(params);
@@ -420,7 +420,6 @@ bool gpt2_eval(
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -442,7 +441,7 @@ bool gpt2_eval(
         // norm
         {
             // [ 768, N]
-            cur = ggml_norm(ctx0, inpL);
+            cur = ggml_norm(ctx0, inpL, 1e-5f);
 
             // cur = ln_1_g*cur + ln_1_b
             // [ 768, N]
@@ -525,8 +524,7 @@ bool gpt2_eval(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
+                        1.0f/sqrt(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // [n_past + N, N, 12]
@@ -589,7 +587,7 @@ bool gpt2_eval(
         {
             // norm
             {
-                cur = ggml_norm(ctx0, inpFF);
+                cur = ggml_norm(ctx0, inpFF, 1e-5f);
 
                 // cur = ln_2_g*cur + ln_2_b
                 // [ 768, N]
@@ -644,7 +642,7 @@ bool gpt2_eval(
     // norm
     {
         // [ 768, N]
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, 1e-5f);
 
         // inpL = ln_f_g*inpL + ln_f_b
         // [ 768, N]
@@ -664,8 +662,8 @@ bool gpt2_eval(
     //inpL = ggml_soft_max(ctx0, inpL);
 
     // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_build_forward_expand  (&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     //if (n_past%100 == 0) {
     //    ggml_graph_print   (&gf);

examples/talk/README.md

@@ -31,7 +31,7 @@ To run this, you will need a ggml GPT-2 model: [instructions](https://github.com
 Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:
 
 ```
-wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
+wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
 ```
 
 ## TTS

examples/talk/gpt-2.cpp

@@ -121,13 +121,13 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             return false;
         }
 
-        std::string word;
+        char word[129];
+
         for (int i = 0; i < n_vocab; i++) {
             uint32_t len;
             fin.read((char *) &len, sizeof(len));
-
-            word.resize(len);
-            fin.read((char *) word.data(), len);
+            word[len] = '\0';
+            fin.read((char *) word, len);
 
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;
@@ -155,33 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
+        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
 
-        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
-        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
+        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
+        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
 
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
 
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
+        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
+        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -191,9 +191,9 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
     // create the ggml context
     {
         struct ggml_init_params params = {
-            .mem_size   = ctx_size,
-            .mem_buffer = NULL,
-            .no_alloc   = false,
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
         };
 
         model.ctx = ggml_init(params);
@@ -379,6 +379,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 //   - embd_inp:  the embeddings of the tokens in the context
 //   - embd_w:    the predicted logits for the next token
 //
+// TODO: sync latest version from ggml repo
 bool gpt2_eval(
         const gpt2_model & model,
         const int n_threads,
@@ -420,7 +421,6 @@ bool gpt2_eval(
 
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@@ -442,7 +442,7 @@ bool gpt2_eval(
         // norm
         {
             // [ 768, N]
-            cur = ggml_norm(ctx0, inpL);
+            cur = ggml_norm(ctx0, inpL, 1e-5f);
 
             // cur = ln_1_g*cur + ln_1_b
             // [ 768, N]
@@ -525,8 +525,7 @@ bool gpt2_eval(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
+                        1.0f/sqrt(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // [n_past + N, N, 12]
@@ -589,7 +588,7 @@ bool gpt2_eval(
         {
             // norm
             {
-                cur = ggml_norm(ctx0, inpFF);
+                cur = ggml_norm(ctx0, inpFF, 1e-5f);
 
                 // cur = ln_2_g*cur + ln_2_b
                 // [ 768, N]
@@ -644,7 +643,7 @@ bool gpt2_eval(
     // norm
     {
         // [ 768, N]
-        inpL = ggml_norm(ctx0, inpL);
+        inpL = ggml_norm(ctx0, inpL, 1e-5f);
 
         // inpL = ln_f_g*inpL + ln_f_b
         // [ 768, N]
@@ -664,8 +663,8 @@ bool gpt2_eval(
     //inpL = ggml_soft_max(ctx0, inpL);
 
     // run the computation
-    ggml_build_forward_expand(&gf, inpL);
-    ggml_graph_compute       (ctx0, &gf);
+    ggml_build_forward_expand  (&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     //if (n_past%100 == 0) {
     //    ggml_graph_print   (&gf);

examples/talk/talk.cpp

@@ -1,8 +1,8 @@
 // Talk with AI
 //
 
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
 #include "gpt-2.h"
 
@@ -31,6 +31,7 @@ struct whisper_params {
     bool print_special = false;
     bool print_energy  = false;
     bool no_timestamps = true;
+    bool use_gpu       = true;
 
     std::string person    = "Santa";
     std::string language  = "en";
@@ -61,6 +62,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
         else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
@@ -94,6 +96,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
     fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
@@ -181,8 +184,10 @@ int main(int argc, char ** argv) {
     }
 
     // whisper init
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = params.use_gpu;
 
-    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
 
     // gpt init
 
@@ -349,7 +354,10 @@ int main(int argc, char ** argv) {
                 gpt2_set_prompt(ctx_gpt, prompt_base.c_str());
 
                 text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
-                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+                int ret = system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+                if (ret != 0) {
+                    fprintf(stderr, "%s: system() failed!\n", __func__);
+                }
 
                 audio.clear();
 

examples/twitch.sh

@@ -21,7 +21,7 @@ help()
     echo "Usage: ./twitch.sh -s [step] -m [model] -t [threads] [url]"
     echo "options:"
     echo "-s       Step in seconds (default is $step)."
-    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large' (default is '$model')."
+    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large-v2' 'large-v3' (default is '$model')."
     echo "-t       Number of threads to use."
     echo "-h       Print this help page."
     echo

examples/wchess/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(CMAKE_CXX_STANDARD 11)
+
+add_subdirectory(libwchess)
+
+if (EMSCRIPTEN)
+    add_subdirectory(wchess.wasm)
+else()
+    add_subdirectory(wchess.cmd)
+endif()

examples/wchess/README.md

@@ -0,0 +1,45 @@
+# wchess
+
+Voice-controlled chess using Whisper
+
+Online demo: https://whisper.ggerganov.com/wchess/
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/c2b2f03c-9684-49f3-8106-357d2d4e67fa
+
+## Command-line tool
+
+```bash
+mkdir build && cd build
+cmake -DWHISPER_SDL2=1 ..
+make -j
+
+./bin/wchess -m ../models/ggml-base.en.bin
+
+Move: start
+
+a b c d e f g h
+r n b q k b n r 8
+p p p p p p p p 7
+. * . * . * . * 6
+* . * . * . * . 5
+. * . * . * . * 4
+* . * . * . * . 3
+P P P P P P P P 2
+R N B Q K B N R 1
+
+White's turn
+[(l)isten/(p)ause/(q)uit]: 
+```
+
+## TODO
+
+- Fix bugs in the chess moves logic
+- Improve web-browser audio capture - sometimes it does not record the voice properly
+- Add support for more languages by making the generated grammar string multilingual
+- Explore ways to improve the dynamic grammar to be narrower
+
+PRs welcome!
+
+## Thanks
+
+- [chessboardjs](https://chessboardjs.com) for the neat chessboard JS library used in this demo

examples/wchess/libwchess/CMakeLists.txt

@@ -0,0 +1,19 @@
+add_library(wchess-core STATIC
+    WChess.cpp
+    WChess.h
+    Chessboard.cpp
+    Chessboard.h
+)
+
+target_link_libraries(wchess-core
+    PUBLIC
+    whisper
+    common
+)
+
+target_include_directories(wchess-core
+    PUBLIC
+    "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>"
+)
+
+# add_executable(test-chessboard test-chessboard.cpp Chessboard.cpp)

examples/wchess/libwchess/Chessboard.cpp

@@ -0,0 +1,803 @@
+#include "Chessboard.h"
+
+#include <array>
+#include <vector>
+#include <algorithm>
+#include <cstring>
+#include <set>
+#include <list>
+#include <chrono>
+
+namespace {
+constexpr std::array<const char*, 64> positions = {
+    "a1", "b1", "c1", "d1", "e1", "f1", "g1", "h1",
+    "a2", "b2", "c2", "d2", "e2", "f2", "g2", "h2",
+    "a3", "b3", "c3", "d3", "e3", "f3", "g3", "h3",
+    "a4", "b4", "c4", "d4", "e4", "f4", "g4", "h4",
+    "a5", "b5", "c5", "d5", "e5", "f5", "g5", "h5",
+    "a6", "b6", "c6", "d6", "e6", "f6", "g6", "h6",
+    "a7", "b7", "c7", "d7", "e7", "f7", "g7", "h7",
+    "a8", "b8", "c8", "d8", "e8", "f8", "g8", "h8",
+};
+constexpr char INVALID_POS = positions.size();
+constexpr int R = 0; // rank index
+constexpr int F = 1; // file index
+#define FILE (c[F] - '1')
+#define RANK (c[R] - 'a')
+constexpr char operator ""_P(const char * c, size_t size) {
+    return size < 2 || RANK < 0 || RANK > 7 ||
+        FILE < 0 || FILE > 7 ? INVALID_POS : FILE * 8 + RANK;
+}
+#undef FILE
+#undef RANK
+
+struct sview {
+    const char * ptr = nullptr;
+    size_t size = 0;
+
+    sview() = default;
+    sview(const char * p, size_t s) : ptr(p), size(s) {}
+    sview(const std::string& s) : ptr(s.data()), size(s.size()) {}
+
+    size_t find(char del, size_t pos) {
+        while (pos < size && ptr[pos] != del) ++pos;
+        return pos < size ? pos : std::string::npos;
+    }
+};
+
+std::vector<sview> split(sview str, char del) {
+    std::vector<sview> res;
+    size_t cur = 0;
+    size_t last = 0;
+    while (cur != std::string::npos) {
+        if (str.ptr[last] == ' ') {
+            ++last;
+            continue;
+        }
+        cur = str.find(del, last);
+        size_t len = cur == std::string::npos ? str.size - last : cur - last;
+        res.emplace_back(str.ptr + last, len);
+        last = cur + 1;
+    }
+    return res;
+}
+
+char strToPos(sview str) {
+    return operator ""_P(str.ptr, str.size);
+}
+
+constexpr std::array<const char*, 6> pieceNames =  {
+    "pawn", "knight", "bishop", "rook", "queen", "king",
+};
+
+static constexpr std::array<char, 6> blackShort =  {
+    'p', 'n', 'b', 'r', 'q', 'k',
+};
+static constexpr std::array<char, 6> whiteShort =  {
+    'P', 'N', 'B', 'R', 'Q', 'K',
+};
+
+char strToType(sview str) {
+    auto it = std::find_if(pieceNames.begin(), pieceNames.end(), [str] (const char* name) { return strncmp(name, str.ptr, str.size) == 0; });
+    return it != pieceNames.end() ? it - pieceNames.begin() : pieceNames.size();
+}
+
+// directions
+using Direction = std::array<char, 2>;
+
+constexpr Direction N   = {(char)  0, (char)  1};
+constexpr Direction NNE = {(char)  1, (char)  2};
+constexpr Direction NE  = {(char)  1, (char)  1};
+constexpr Direction ENE = {(char)  2, (char)  1};
+constexpr Direction E   = {(char)  1, (char)  0};
+constexpr Direction ESE = {(char)  2, (char) -1};
+constexpr Direction SE  = {(char)  1, (char) -1};
+constexpr Direction SSE = {(char)  1, (char) -2};
+constexpr Direction S   = {(char)  0, (char) -1};
+constexpr Direction SSW = {(char) -1, (char) -2};
+constexpr Direction SW  = {(char) -1, (char) -1};
+constexpr Direction WSW = {(char) -2, (char) -1};
+constexpr Direction W   = {(char) -1, (char)  0};
+constexpr Direction WNW = {(char) -2, (char)  1};
+constexpr Direction NW  = {(char) -1, (char)  1};
+constexpr Direction NNW = {(char) -1, (char)  2};
+
+char makeStep(char pos, const Direction& d) {
+    char next[2] = { char(positions[pos][R] + d[R]) , char(positions[pos][F] + d[F]) };
+    return strToPos(sview{next, sizeof(next)});
+}
+
+template<class Modifier>
+char traverse(char pos, const Direction& d, const Modifier& m, int count = 8) {
+    while (--count >= 0) {
+        pos = makeStep(pos, d);
+        if (pos == INVALID_POS || m(pos)) break;
+    }
+    return pos;
+}
+
+Direction normalize(const Direction& distance) {
+    //return {char((distance[R] > 0) - (distance[R] < 0)), char((distance[F] > 0) - (distance[F] < 0))};
+    const int drp = distance[R] > 0 ? 1 : 0;
+    const int drn = distance[R] < 0 ? 1 : 0;
+    const int dfp = distance[F] > 0 ? 1 : 0;
+    const int dfn = distance[F] < 0 ? 1 : 0;
+    return {char(drp - drn), char(dfp - dfn)};
+}
+
+struct Pin {
+    Direction d;
+    Piece* pinner;
+    Piece* pinned;
+};
+using Pins = std::list<Pin>;
+using Board = std::array<Piece*, 64>;
+
+std::vector<Direction> filter(const Direction& pin, std::initializer_list<Direction> directions) {
+    if (pin[R] == 0 && pin[F] == 0) return directions;
+    std::vector<Direction> result;
+    for (auto& d : directions) {
+        if ((d[R] == pin[R] || d[R] == -pin[R]) && (d[F] == pin[F] || d[F] == -pin[F])) result.push_back(d);
+    }
+    return result;
+}
+}
+
+class Piece {
+public:
+    enum Types : char {
+        Pawn,
+        Knight,
+        Bishop,
+        Rook,
+        Queen,
+        King,
+        //
+        NUM_PIECES
+    };
+
+    enum Colors : char {
+        White,
+        Black,
+    };
+
+    const char* name() const;
+    char initial() const;
+    Types type() const { return m_type; }
+    Colors color() const { return m_color; }
+    char pos() const { return m_pos; }
+    void setPos(char pos) {
+        m_pos = pos;
+        invalidate();
+    }
+    const char* coord() const;
+    const std::set<char>& allowed() const { return m_allowed; }
+    bool canReach(char pos) const;
+    virtual bool movePattern(char pos) const = 0;
+    void take();
+    virtual void reinit(const State& state) = 0;
+    void invalidate();
+protected:
+    Piece(Types type, Colors color, char pos, std::set<char> allowed)
+        : m_type(type), m_color(color), m_pos(pos), m_allowed(std::move(allowed)) {}
+    Piece(const Piece&) = delete;
+    ~Piece() = default;
+
+    const Types m_type;
+    const Colors m_color;
+    char m_pos;
+    std::set<char> m_allowed;
+    bool m_update = false;
+};
+
+struct Pawn : public Piece {
+    Pawn(Colors color, char pos, std::set<char> next) : Piece(Types::Pawn, color, pos, std::move(next)) {}
+
+    bool is_first_move() const {
+        return m_color ? coord()[F] == '7' : coord()[F] == '2';
+    }
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction distance = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        char forward = m_color ? -1 : 1;
+        return (forward == distance[F] && distance[R] * distance[R] <= 1)
+            || (is_first_move() && 2 * forward == distance[F] && distance[R] == 0);
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Knight : public Piece {
+    Knight(Colors color, char pos, std::set<char> next) : Piece(Types::Knight, color, pos, std::move(next)) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction diff = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        return diff[R]*diff[R] + diff[F]*diff[F] == 5;
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Bishop : public Piece {
+    Bishop(Colors color, char pos) : Piece(Types::Bishop, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] - cur[F] == next[R] - next[F] || cur[R] + cur[F] == next[R] + next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Rook : public Piece {
+    Rook(Colors color, char pos) : Piece(Types::Rook, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] == next[R] || cur[F] == next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Queen : public Piece {
+    Queen(Colors color, char pos) : Piece(Types::Queen, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] == next[R] || cur[F] == next[F] || cur[R] - cur[F] == next[R] - next[F] || cur[R] + cur[F] == next[R] + next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct King : public Piece {
+    King(Colors color, char pos) : Piece(Types::King, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction diff = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        return diff[R]*diff[R] + diff[F]*diff[F] <= 2;
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct PieceSet {
+    Piece* begin() { return &p1; }
+    Piece* end() { return &r2 + 1; }
+    const Piece* begin() const { return &p1; }
+    const Piece* end() const { return &r2 + 1; }
+    Piece& operator[](int i) { return *(begin() + i); }
+    const Piece& operator[](int i) const { return *(begin() + i); }
+
+    Pawn   p1;
+    Pawn   p2;
+    Pawn   p3;
+    Pawn   p4;
+    Pawn   p5;
+    Pawn   p6;
+    Pawn   p7;
+    Pawn   p8;
+    Rook   r1;
+    Knight n1;
+    Bishop b1;
+    Queen  q;
+    King   k;
+    Bishop b2;
+    Knight n2;
+    Rook   r2;
+};
+
+struct State {
+    State();
+    PieceSet blacks;
+    PieceSet whites;
+    Board board;
+    Pins blackPins;
+    Pins whitePins;
+};
+
+Direction findPin(const Piece& piece, const State& state) {
+    auto& pins = piece.color() ? state.blackPins : state.whitePins;
+    auto it = std::find_if(pins.begin(), pins.end(), [&] (const Pin& pin) { return pin.pinned == &piece; });
+    if (it != pins.end()) return it->d;
+    return {0, 0};
+}
+
+struct Find {
+    Find(const Board& board) : m_board(board) {}
+    bool operator() (char pos) const { return m_board[pos]; }
+    const Board& m_board;
+};
+
+struct Add {
+    Add(const Board& board, std::set<char>& moves, Piece::Colors color) : m_board(board), m_moves(moves), m_color(color) {}
+    bool operator() (char pos) const {
+        if (!m_board[pos] || m_board[pos]->color() != m_color) m_moves.insert(pos);
+        return m_board[pos];
+    }
+    const Board& m_board;
+    std::set<char>& m_moves;
+    Piece::Colors m_color;
+};
+
+void Pawn::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+
+    auto pin = findPin(*this, state);
+
+    auto & left = m_color ? SW : NW;
+    auto & right = m_color ? SE : NE;
+
+    for (auto& direction : filter(pin, { left, right })) {
+        auto pos = makeStep(m_pos, direction);
+        if (pos != INVALID_POS && state.board[pos] && state.board[pos]->color() != m_color) m_allowed.insert(pos);
+    }
+
+    auto & forward = m_color ? S : N;
+    if (!filter(pin, {forward}).empty()) {
+        traverse(m_pos, forward, [&] (char pos) {
+                if (!state.board[pos]) m_allowed.insert(pos);
+                return state.board[pos] || !is_first_move();
+            }, 2);
+    }
+}
+
+void Knight::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    if (pin[R] != 0 || pin[F] != 0) return;
+    for (auto& direction : { NNE, ENE, ESE, SSE, SSW, WSW, WNW, NNW }) {
+        auto pos = makeStep(m_pos, direction);
+        if (pos != INVALID_POS && (!state.board[pos] || state.board[pos]->color() != m_color)) m_allowed.insert(pos);
+    }
+}
+
+void Bishop::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { NE, SE, SW, NW })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void Rook::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { N, E, S, W })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void Queen::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { N, NE, E, SE, S, SW, W, NW })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void King::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto& enemyPieces = m_color ? state.whites : state.blacks;
+    auto& pawnAttackLeft = m_color ? SW : NW;
+    auto& pawnAttackRight = m_color ? SE : NE;
+    for (auto& direction : { N, NE, E, SE, S, SW, W, NW }) {
+        auto pos = makeStep(m_pos, direction);
+        bool accept = pos != INVALID_POS && !(state.board[pos] && state.board[pos]->color() == m_color);
+        if (accept) {
+            for (auto& p : enemyPieces) {
+                if (!p.movePattern(pos)) continue;
+                if (p.type() == Piece::Knight || p.type() == Piece::King) {
+                    accept = false;
+                    break;
+                }
+                else if (p.type() == Piece::Pawn) {
+                    auto from = positions[pos];
+                    auto to = p.coord();
+                    Direction d {char(to[R] - from[R]), char(to[F] - from[F])};
+                    if (d == pawnAttackLeft || d == pawnAttackRight) {
+                        accept = false;
+                        break;
+                    }
+                }
+                else {
+                    auto from = positions[pos];
+                    auto to = p.coord();
+                    Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+                    auto reached = traverse(pos, d, Find(state.board));
+                    if (p.pos() == reached) {
+                        accept = false;
+                        break;
+                    }
+                }
+            }
+        }
+        if (accept) m_allowed.insert(pos);
+    }
+}
+
+const char* Piece::name() const {
+    static_assert(pieceNames.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    return pieceNames[m_type];
+}
+
+char Piece::initial() const {
+    static_assert(blackShort.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    static_assert(whiteShort.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    return m_color ? blackShort[m_type] : whiteShort[m_type];
+}
+
+void Piece::invalidate() {
+    m_update = true;
+}
+
+
+const char* Piece::coord() const {
+    if (m_pos == INVALID_POS) return "";
+    return positions[m_pos];
+}
+
+bool Piece::canReach(char pos) const {
+    return movePattern(pos) && m_allowed.count(pos);
+}
+
+void Piece::take() {
+    m_pos = INVALID_POS;
+    m_allowed = {};
+}
+
+State::State()
+    : blacks {
+        {Piece::Black, "a7"_P, {"a5"_P, "a6"_P} },
+        {Piece::Black, "b7"_P, {"b5"_P, "b6"_P} },
+        {Piece::Black, "c7"_P, {"c5"_P, "c6"_P} },
+        {Piece::Black, "d7"_P, {"d5"_P, "d6"_P} },
+        {Piece::Black, "e7"_P, {"e5"_P, "e6"_P} },
+        {Piece::Black, "f7"_P, {"f5"_P, "f6"_P} },
+        {Piece::Black, "g7"_P, {"g5"_P, "g6"_P} },
+        {Piece::Black, "h7"_P, {"h5"_P, "h6"_P} },
+        {Piece::Black, "a8"_P},
+        {Piece::Black, "b8"_P, {"a6"_P, "c6"_P} },
+        {Piece::Black, "c8"_P},
+        {Piece::Black, "d8"_P},
+        {Piece::Black, "e8"_P},
+        {Piece::Black, "f8"_P},
+        {Piece::Black, "g8"_P, {"f6"_P, "h6"_P} },
+        {Piece::Black, "h8"_P},
+    }
+    , whites {
+        {Piece::White, "a2"_P, {"a3"_P, "a4"_P} },
+        {Piece::White, "b2"_P, {"b3"_P, "b4"_P} },
+        {Piece::White, "c2"_P, {"c3"_P, "c4"_P} },
+        {Piece::White, "d2"_P, {"d3"_P, "d4"_P} },
+        {Piece::White, "e2"_P, {"e3"_P, "e4"_P} },
+        {Piece::White, "f2"_P, {"f3"_P, "f4"_P} },
+        {Piece::White, "g2"_P, {"g3"_P, "g4"_P} },
+        {Piece::White, "h2"_P, {"h3"_P, "h4"_P} },
+        {Piece::White, "a1"_P},
+        {Piece::White, "b1"_P, {"a3"_P, "c3"_P} },
+        {Piece::White, "c1"_P},
+        {Piece::White, "d1"_P},
+        {Piece::White, "e1"_P},
+        {Piece::White, "f1"_P},
+        {Piece::White, "g1"_P, {"f3"_P, "h3"_P} },
+        {Piece::White, "h1"_P},
+    }
+    , board {{
+        &whites[ 8],  &whites[ 9],  &whites[10],  &whites[11],  &whites[12],  &whites[13],  &whites[14],  &whites[15],
+        &whites[ 0],  &whites[ 1],  &whites[ 2],  &whites[ 3],  &whites[ 4],  &whites[ 5],  &whites[ 6],  &whites[ 7],
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        &blacks[ 0],  &blacks[ 1],  &blacks[ 2],  &blacks[ 3],  &blacks[ 4],  &blacks[ 5],  &blacks[ 6],  &blacks[ 7],
+        &blacks[ 8],  &blacks[ 9],  &blacks[10],  &blacks[11],  &blacks[12],  &blacks[13],  &blacks[14],  &blacks[15],
+    }}
+{}
+
+Chessboard::Chessboard()
+    : m_state(new State())
+{
+    setGrammar();
+}
+
+Chessboard::~Chessboard() = default;
+
+void Chessboard::setPrompt(const std::string& prompt) {
+    m_prompt = prompt;
+    setGrammar();
+}
+
+void Chessboard::setGrammar() {
+    m_grammar.clear();
+
+    std::string result;
+    if (m_prompt.empty()) {
+        result += "move ::= \" \" ((piece | frompos) \" \" \"to \"?)? topos\n";
+        //result += "move ::= \" \" frompos \" \" \"to \"? topos\n";
+    }
+    else {
+        // result += "move ::= prompt \" \" ((piece | frompos) \" \" \"to \"?)? topos\n"
+        result += "move ::= prompt \" \" frompos \" \" \"to \"? topos\n"
+        "prompt ::= \" " + m_prompt + "\"\n";
+    }
+
+    std::set<Piece::Types> pieceTypes;
+    std::set<char> from_pos;
+    std::set<char> to_pos;
+    auto& pieces =  m_moveCounter % 2 ? m_state->blacks : m_state->whites;
+    std::set<size_t> flags;
+    for (auto& p : pieces) {
+        if (p.allowed().empty()) continue;
+        bool addPiece = false;
+        if (!m_inCheck || p.type() == Piece::King) {
+            to_pos.insert(p.allowed().begin(), p.allowed().end());
+            addPiece = !p.allowed().empty();
+        }
+        else {
+            for (auto move : p.allowed()) {
+                if (m_allowedInCheck.count(move)) {
+                    to_pos.insert(move);
+                    addPiece = true;
+                }
+            }
+        }
+        if (addPiece) {
+            pieceTypes.insert(p.type());
+            from_pos.insert(p.pos());
+        }
+    }
+    if (pieceTypes.empty()) return;
+
+    result += "piece ::= (";
+    for (auto& p : pieceTypes) result += " \"" + std::string(pieceNames[p]) + "\" |";
+    result.pop_back();
+    result += ")\n\n";
+
+    result += "frompos ::= (";
+    for (auto& p : from_pos) result += " \"" + std::string(positions[p]) + "\" |";
+    result.pop_back();
+    result += ")\n";
+
+    result += "topos ::= (";
+    for (auto& p : to_pos) result += " \"" + std::string(positions[p]) + "\" |";
+    result.pop_back();
+    result += ")\n";
+
+    m_grammar = std::move(result);
+}
+
+std::string Chessboard::stringifyBoard() {
+    std::string result;
+    result.reserve(16 + 2 * 64 + 16);
+    for (char rank = 'a'; rank <= 'h'; ++rank) {
+        result.push_back(rank);
+        result.push_back(' ');
+    }
+    result.back() = '\n';
+    for (int i = 7; i >= 0; --i) {
+        for (int j = 0; j < 8; ++j) {
+            auto p = m_state->board[i * 8 + j];
+            if (p) result.push_back(p->initial());
+            else result.push_back((i + j) % 2 ? '.' : '*');
+            result.push_back(' ');
+        }
+        result.push_back('0' + i + 1);
+        result.push_back('\n');
+    }
+    return result;
+}
+
+std::string Chessboard::process(const std::string& command) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+    auto color = Piece::Colors(m_moveCounter % 2);
+    Piece* piece = nullptr;
+    auto pos_to = INVALID_POS;
+    if (!parseCommand(command, piece, pos_to)) return "";
+
+    auto pos_from = piece->pos();
+
+    if (!move(*piece, pos_to)) return "";
+
+    flagUpdates(pos_from, pos_to);
+
+    detectChecks();
+
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    for (auto& p : enemyPieces) p.reinit(*m_state); // only enemy moves needed next
+
+    std::string result = {positions[pos_from][R], positions[pos_from][F], '-', positions[pos_to][R], positions[pos_to][F]};
+    ++m_moveCounter;
+    setGrammar();
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    auto t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+    fprintf(stdout, "%s: Move '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", result.data(), "\033[0m", (int) t_ms);
+    if (m_grammar.empty()) result.push_back('#');
+    return result;
+}
+
+bool Chessboard::parseCommand(const std::string& command, Piece*& piece, char& pos_to) {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    fprintf(stdout, "%s: Command to %s: '%s%.*s%s'\n", __func__, (color ? "Black" : "White"), "\033[1m", int(command.size()), command.data(), "\033[0m");
+
+    if (command.empty()) return false;
+    auto tokens = split(command, ' ');
+    auto pos_from = INVALID_POS;
+    auto type = Piece::Types::NUM_PIECES;
+    if (tokens.size() == 1) {
+        type = Piece::Types::Pawn;
+        pos_to = strToPos(tokens.front());
+    }
+    else {
+        pos_from = strToPos(tokens.front());
+        if (pos_from == INVALID_POS) type = Piece::Types(strToType(tokens.front()));
+        pos_to = strToPos(tokens.back());
+    }
+    if (pos_to == INVALID_POS) return false;
+    if (pos_from == INVALID_POS) {
+        if (type == Piece::Types::NUM_PIECES) return false;
+        auto& pieces = color ? m_state->blacks : m_state->whites;
+        for (auto& p : pieces) {
+            if (p.type() == type && p.canReach(pos_to)) {
+                pos_from = p.pos();
+                break;
+            }
+        }
+    }
+    if (pos_from == INVALID_POS) return false;
+    if (m_state->board[pos_from] == nullptr) return false;
+    piece = m_state->board[pos_from];
+    if (piece->color() != color) return false;
+    return true;
+}
+
+void Chessboard::flagUpdates(char pos_from, char pos_to) {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    auto& ownPieces = color ? m_state->blacks : m_state->whites;
+    for (auto& p : enemyPieces) {
+        if (p.movePattern(pos_to) || p.movePattern(pos_from)) {
+            updatePins(p);
+            p.invalidate();
+        }
+    }
+
+    for (auto& p : ownPieces) {
+        if (p.movePattern(pos_to) || p.movePattern(pos_from)) {
+            updatePins(p);
+            p.invalidate();
+        }
+    }
+}
+
+void Chessboard::updatePins(Piece& piece) {
+    if (piece.type() == Piece::Pawn || piece.type() == Piece::Knight || piece.type() == Piece::King) return;
+    auto& enemyPieces = piece.color() ? m_state->whites : m_state->blacks;
+    auto& enemyPins = piece.color() ? m_state->whitePins : m_state->blackPins;
+    auto& king = enemyPieces.k;
+    auto it = std::find_if(enemyPins.begin(), enemyPins.end(), [&] (const Pin& pin) { return pin.pinner == &piece; });
+    if (it != enemyPins.end()) {
+        it->pinned->invalidate();
+        enemyPins.erase(it);
+    }
+    if (piece.movePattern(king.pos())) {
+        auto to = positions[king.pos()];
+        auto from = piece.coord();
+        Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+
+        auto reached = traverse(piece.pos(), d, Find(m_state->board));
+        auto foundPiece = m_state->board[reached];
+        if (&king == foundPiece) {
+            // check
+            king.invalidate();
+        }
+        else if (foundPiece && foundPiece->color() != piece.color()) {
+            reached = traverse(reached, d, Find(m_state->board));
+            if (&king == m_state->board[reached]) {
+                enemyPins.push_back({d, &piece, foundPiece});
+                foundPiece->invalidate();
+            }
+        }
+    }
+}
+
+void Chessboard::detectChecks() {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    auto& ownPieces = color ? m_state->blacks : m_state->whites;
+    auto& king = enemyPieces.k;
+    auto& pawnAttackLeft = color ? SW : NW;
+    auto& pawnAttackRight = color ? SE : NE;
+    for (auto& p : ownPieces) {
+        if (!p.movePattern(king.pos())) continue;
+        auto to = positions[king.pos()];
+        auto from = p.coord();
+
+        if (p.type() == Piece::Knight) {
+            if (!m_inCheck) {
+                m_allowedInCheck = { p.pos() };
+            }
+            else {
+                m_allowedInCheck.clear();
+            }
+            m_inCheck = true;
+        }
+        else if (p.type() == Piece::Pawn) {
+            Direction d {char(to[R] - from[R]), char(to[F] - from[F])};
+            if (d == pawnAttackLeft || d == pawnAttackRight) {
+                if (!m_inCheck) {
+                    m_allowedInCheck = { p.pos() };
+                }
+                else {
+                    m_allowedInCheck.clear();
+                }
+                m_inCheck = true;
+            }
+        }
+        else {
+            Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+            std::set<char> tmp;
+            auto pos = traverse(p.pos(), d, Add(m_state->board, tmp, king.color()));
+            if (pos == king.pos()) {
+                tmp.insert(p.pos());
+                if (!m_inCheck) {
+                    m_allowedInCheck = std::move(tmp);
+                }
+                else {
+                    m_allowedInCheck.clear();
+                }
+                m_inCheck = true;
+            }
+        }
+    }
+}
+
+bool Chessboard::move(Piece& piece, char pos_to) {
+    auto& allowed = piece.allowed();
+
+    if (allowed.count(pos_to) == 0 || (m_inCheck && piece.type() != Piece::King && m_allowedInCheck.count(pos_to) == 0)) return false;
+    if (m_state->board[pos_to] && m_state->board[pos_to]->color() == piece.color()) return false;
+    if (m_state->board[pos_to]) m_state->board[pos_to]->take();
+    m_state->board[piece.pos()] = nullptr;
+    m_state->board[pos_to] = &piece;
+    piece.setPos(pos_to);
+
+    m_inCheck = false;
+    m_allowedInCheck.clear();
+
+    return true;
+}

examples/wchess/libwchess/Chessboard.h

@@ -0,0 +1,33 @@
+#pragma once
+#include <string>
+#include <set>
+#include <memory>
+
+// just basic validation
+// fixme: missing en passant, castling, promotion, etc.
+struct State;
+class Piece;
+class Chessboard {
+public:
+    Chessboard();
+    ~Chessboard();
+    std::string process(const std::string& command);
+    std::string stringifyBoard();
+    const std::string& grammar() { return m_grammar; }
+    const std::string& prompt() { return m_prompt; }
+    void setPrompt(const std::string& prompt);
+private:
+    bool parseCommand(const std::string& command, Piece*& piece, char& pos_to);
+    bool move(Piece& piece, char pos);
+    void flagUpdates(char pos_from, char pos_to);
+    void updatePins(Piece& piece);
+    void detectChecks();
+    void setGrammar();
+
+    std::unique_ptr<State> m_state;
+    std::set<char> m_allowedInCheck;
+    bool m_inCheck = false;
+    int m_moveCounter = 0;
+    std::string m_grammar;
+    std::string m_prompt;
+};

examples/wchess/libwchess/WChess.cpp

@@ -0,0 +1,193 @@
+#include "WChess.h"
+#include "Chessboard.h"
+#include "grammar-parser.h"
+#include "common.h"
+#include <thread>
+
+WChess::WChess(whisper_context * ctx,
+        const whisper_full_params & wparams,
+        callbacks cb,
+        settings s)
+        : m_ctx(ctx)
+        , m_wparams(wparams)
+        , m_cb(cb)
+        , m_settings(s)
+        , m_board(new Chessboard())
+{}
+
+WChess::~WChess() = default;
+
+void WChess::set_move(const std::string& moves, float prob) const {
+    if (m_cb.set_move) (*m_cb.set_move)(moves, prob);
+}
+
+void WChess::set_grammar(const std::string& grammar) const {
+    if (m_cb.set_grammar) (*m_cb.set_grammar)(grammar);
+}
+
+bool WChess::get_audio(std::vector<float>& pcmf32) const {
+    if (m_cb.get_audio) return (*m_cb.get_audio)(pcmf32);
+    return false;
+}
+
+std::string WChess::stringify_board() const {
+    return m_board->stringifyBoard();
+}
+
+std::string WChess::get_grammar() const {
+    return m_board->grammar();
+}
+
+void WChess::run() {
+    bool have_prompt  = true;
+    bool ask_prompt   = !have_prompt;
+
+    float logprob_min  = 0.0f;
+
+    float logprob_sum  = 0.0f;
+
+    int n_tokens  = 0;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string k_prompt = have_prompt ? "" : "rook to d4, f3";
+    int64_t t_ms = 0;
+
+    if (ask_prompt) {
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+        fprintf(stdout, "\n");
+
+        ask_prompt = false;
+    }
+
+    while (get_audio(pcmf32_cur)) {
+        if (!pcmf32_cur.empty()) {
+            // fprintf(stdout, "%s: Processing ...\n", __func__);
+
+            if (!have_prompt) {
+                const auto txt = ::trim(transcribe(pcmf32_cur, logprob_min, logprob_sum, n_tokens, t_ms));
+
+                fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+
+                const float sim = similarity(txt, k_prompt);
+
+                if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                    fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                    ask_prompt = true;
+                } else {
+                    fprintf(stdout, "\n");
+                    fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                    fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                    fprintf(stdout, "\n");
+
+                    // save the audio for the prompt
+                    pcmf32_prompt = pcmf32_cur;
+                    have_prompt = true;
+                    m_board->setPrompt(k_prompt);
+                }
+            } else {
+                if (!pcmf32_prompt.empty()) pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+                constexpr size_t MIN_SIZE = 1.2 * WHISPER_SAMPLE_RATE;
+                if (MIN_SIZE > pcmf32_cur.size()) pcmf32_cur.insert(pcmf32_cur.begin(), MIN_SIZE - pcmf32_cur.size(), 0.0f);
+
+                // fprintf(stdout, "%s: grammar rules:\n'%s'\n", __func__, m_board->grammar().c_str());
+
+                auto grammar_parsed = grammar_parser::parse(m_board->grammar().c_str());
+                auto grammar_rules  = grammar_parsed.c_rules();
+
+                m_wparams.grammar_rules   = grammar_rules.data();
+                m_wparams.n_grammar_rules = grammar_rules.size();
+
+                m_wparams.i_start_rule    = grammar_parsed.symbol_ids.at("move");
+                auto txt = ::trim(transcribe(pcmf32_cur, logprob_min, logprob_sum, n_tokens, t_ms));
+
+                const float p = 100.0f * std::exp(logprob_min);
+
+                fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                // find the prompt in the text
+                float best_sim = 0.0f;
+                size_t best_len = 0;
+                for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                    const auto prompt = txt.substr(0, n);
+
+                    const float sim = similarity(prompt, k_prompt);
+
+                    //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                    if (sim > best_sim) {
+                        best_sim = sim;
+                        best_len = n;
+                    }
+                }
+
+                fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
+                std::string command = ::trim(txt.substr(best_len));
+
+                fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                fprintf(stdout, "\n");
+
+                if (!command.empty()) {
+                    set_move(m_board->process(command), p);
+                    set_grammar(m_board->grammar());
+                }
+                if (m_board->grammar().empty()) {
+                    fprintf(stdout, "%s: No more moves possible\n", __func__);
+                    break;
+                }
+            }
+        }
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+    }
+}
+
+std::string WChess::transcribe(
+                const std::vector<float> & pcmf32,
+                float & logprob_min,
+                float & logprob_sum,
+                int & n_tokens,
+                int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    logprob_min = 0.0f;
+    logprob_sum = 0.0f;
+    n_tokens    = 0;
+    t_ms = 0;
+
+    if (whisper_full(m_ctx, m_wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return {};
+    }
+
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(m_ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(m_ctx, i);
+
+        result += text;
+
+        const int n = whisper_full_n_tokens(m_ctx, i);
+        for (int j = 0; j < n; ++j) {
+            const auto token = whisper_full_get_token_data(m_ctx, i, j);
+
+            if(token.plog > 0.0f) return {};
+            logprob_min = std::min(logprob_min, token.plog);
+            logprob_sum += token.plog;
+            ++n_tokens;
+        }
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}

examples/wchess/libwchess/WChess.h

@@ -0,0 +1,63 @@
+#pragma once
+#include "whisper.h"
+#include <string>
+#include <vector>
+#include <memory>
+
+class Chessboard;
+
+class WChess {
+public:
+    using CheckRunningCb = bool (*)();
+    using GetAudioCb = bool (*)(std::vector<float> &);
+    using SetMovesCb = void (*)(const std::string &, float);
+    using SetGrammarCb = void (*)(const std::string &);
+    using ClearAudioCb = void (*)();
+
+    struct callbacks {
+        GetAudioCb get_audio = nullptr;
+        SetMovesCb set_move = nullptr;
+        SetGrammarCb set_grammar = nullptr;
+    };
+
+    struct settings {
+        int32_t vad_ms     = 2000;
+        int32_t prompt_ms  = 5000;
+        int32_t command_ms = 4000;
+        float vad_thold    = 0.2f;
+        float freq_thold   = 100.0f;
+        bool print_energy  = false;
+    };
+
+    WChess(
+        whisper_context * ctx,
+        const whisper_full_params & wparams,
+        callbacks cb,
+        settings s
+    );
+    ~WChess();
+
+    void run();
+
+    std::string stringify_board() const;
+
+    std::string get_grammar() const;
+
+private:
+    bool get_audio(std::vector<float>& pcmf32) const;
+    void set_move(const std::string& moves, float prob) const;
+    void set_grammar(const std::string& grammar) const;
+
+    std::string transcribe(
+                    const std::vector<float> & pcmf32,
+                    float & logprob_min,
+                    float & logprob_sum,
+                    int & n_tokens,
+                    int64_t & t_ms);
+
+    whisper_context * m_ctx;
+    whisper_full_params m_wparams;
+    const callbacks m_cb;
+    const settings m_settings;
+    std::unique_ptr<Chessboard> m_board;
+};

examples/wchess/libwchess/test-chessboard.cpp

@@ -0,0 +1,117 @@
+#include "Chessboard.h"
+
+#define ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            fflush(stderr); \
+            exit(1); \
+        } \
+    } while (0)
+
+
+int main() {
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("pawn to d4") == "d2-d4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("c1 h6") == "c1-h6");
+        ASSERT(chess.process("queen h4") == "d8-h4");
+        ASSERT(chess.process("bishop to g5") == "h6-g5");
+        ASSERT(chess.process("bishop to b4") == "f8-b4");
+        ASSERT(chess.process("c4") == "");
+        ASSERT(chess.process("knight c3") == "b1-c3");
+        ASSERT(chess.process("knight c6") == "b8-c6");
+        ASSERT(chess.process("f3") == "");
+    }
+
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("d4") == "d2-d4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("queen h4") == "d8-h4");
+        ASSERT(chess.process("queen h5") == "d1-h5");
+        ASSERT(chess.process("f5") == "");
+        ASSERT(chess.process("g6") == "g7-g6");
+        ASSERT(chess.process("knight e2") == "g1-e2");
+        ASSERT(chess.process("f5") == "f7-f5");
+        ASSERT(chess.process("knight g3") == "e2-g3");
+        ASSERT(chess.process("g5") == "");
+        ASSERT(chess.process("king e7") == "e8-e7");
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("g5") == "g6-g5");
+    }
+
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("c5") == "c7-c5");
+        ASSERT(chess.process("e5") == "e4-e5");
+        ASSERT(chess.process("c4") == "c5-c4");
+        ASSERT(chess.process("e6") == "e5-e6");
+        ASSERT(chess.process("c3") == "c4-c3");
+        ASSERT(chess.process("e7") == "");
+        ASSERT(chess.process("f7") == "e6-f7");
+        ASSERT(chess.process("d2") == "");
+        ASSERT(chess.process("king to f7") == "e8-f7");
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("d2") == "c3-d2");
+        ASSERT(chess.process("f5") == "");
+        ASSERT(chess.process("king to e2") == "e1-e2");
+        ASSERT(chess.process("king to g6") == "f7-g6");
+        ASSERT(chess.process("f5") == "f4-f5");
+        ASSERT(chess.process("e6") == "");
+        ASSERT(chess.process("king to h5") == "g6-h5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("king to g5") == "h5-g5");
+        ASSERT(chess.process("h4") == "h2-h4");
+        ASSERT(chess.process("king to h5") == "");
+        ASSERT(chess.process("king to g6") == "");
+        ASSERT(chess.process("king to h6") == "g5-h6");
+        ASSERT(chess.process("bishop to d2") == "c1-d2");
+        ASSERT(chess.process("king to g5") == "");
+        ASSERT(chess.process("g5") == "g7-g5");
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("queen to h4") == "d8-h4#");
+        ASSERT(chess.process("knight f3") == "");
+        ASSERT(chess.grammar().empty());
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("d5") == "d7-d5");
+        ASSERT(chess.process("g1 f3") == "g1-f3");
+        ASSERT(chess.process("queen to h4") == "d8-h4");
+        ASSERT(!chess.grammar().empty());
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("knight c3") == "b1-c3");
+        ASSERT(chess.process("knight c6") == "b8-c6");
+        ASSERT(chess.process("knight b5") == "c3-b5");
+        ASSERT(chess.process("knight f6") == "g8-f6");
+        ASSERT(chess.process("knight d6") == "b5-d6");
+        ASSERT(chess.process("knight d4") == "");
+        ASSERT(chess.process("d6") == "c7-d6");
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("knight d4") == "c6-d4");
+        ASSERT(chess.process("d3") == "d2-d3");
+        ASSERT(chess.process("knight e4") == "f6-e4");
+        ASSERT(chess.process("king to e2") == "");
+        ASSERT(chess.process("king to d2") == "");
+    }
+}

examples/wchess/wchess.cmd/CMakeLists.txt

@@ -0,0 +1,8 @@
+if (WHISPER_SDL2)
+    set(TARGET wchess)
+    add_executable(${TARGET} wchess.cmd.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE wchess-core common-sdl ${CMAKE_THREAD_LIBS_INIT})
+endif ()