release : v1.6.0

* whisper : use flash attention in the encoder

* whisper : add kv_pad

* whisper : remove extra backend instance (huh?)

* whisper : use FA for cross-attention

* whisper : use FA for self-attention

* whisper : simplify encoder FA

* whisper : add flash_attn runtime parameter

* scripts : add bench log

* scripts : add M1 Pro bench log

.devops/main-cuda.Dockerfile

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

.devops/main.Dockerfile

@@ -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

@@ -25,6 +25,7 @@ jobs:
           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
@@ -86,9 +87,10 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
             apt install -y build-essential cmake libsdl2-dev
-            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
             make
             ctest -L gh --output-on-failure'
 
@@ -113,9 +115,10 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
-            apt install -y build-essential cmake libsdl2-dev
-            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+            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'
 
@@ -140,12 +143,171 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
             apt install -y build-essential cmake
             cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
             make
             ctest -L gh --output-on-failure'
 
+  ubuntu-22-cmake-sycl:
+    runs-on: ubuntu-22.04
+
+    strategy:
+      fail-fast: false
+      matrix:
+        dwhisper_sycl: [ON]
+        dcmake_c_compiler: [icx]
+        dcmake_cxx_compiler: [icpx]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake --build . --config Release -j $(nproc)
+
+  ubuntu-22-cmake-sycl-fp16:
+    runs-on: ubuntu-22.04
+
+    strategy:
+      fail-fast: false
+      matrix:
+        dwhisper_sycl: [ON]
+        dcmake_c_compiler: [icx]
+        dcmake_cxx_compiler: [icpx]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake --build . --config Release -j $(nproc)
+
+  windows-msys2:
+    runs-on: windows-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
+          - { sys: CLANG64, env: clang-x86_64, build: Release }
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Setup ${{ matrix.sys }}
+        uses: msys2/setup-msys2@v2
+        with:
+          update: true
+          msystem: ${{matrix.sys}}
+          install: >-
+            base-devel
+            mingw-w64-${{matrix.env}}-toolchain
+            mingw-w64-${{matrix.env}}-cmake
+            mingw-w64-${{matrix.env}}-SDL2
+            mingw-w64-${{matrix.env}}-openblas
+
+      - name: Build using make
+        shell: msys2 {0}
+        run: |
+            make -j $(nproc)
+
+      - name: Clean after building using make
+        shell: msys2 {0}
+        run: |
+            make clean
+
+      - name: Build using make w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            make WHISPER_OPENBLAS=1 -j $(nproc)
+
+      - name: Build using CMake
+        shell: msys2 {0}
+        run: |
+            cmake -B build
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
+      - name: Clean after building using CMake
+        shell: msys2 {0}
+        run: |
+            rm -rf build
+
+      - name: Build using CMake w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            cmake -B build -DWHISPER_OPENBLAS=ON
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
   windows:
     runs-on: windows-latest
 
@@ -162,7 +324,7 @@ jobs:
             s2arc: x64
             jnaPath: win32-x86-64
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
@@ -182,7 +344,7 @@ 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: |
@@ -217,13 +379,16 @@ 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
@@ -239,7 +404,7 @@ 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'
@@ -248,13 +413,26 @@ 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: |
@@ -263,17 +441,21 @@ 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:
@@ -285,11 +467,12 @@ jobs:
         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
@@ -300,7 +483,9 @@ jobs:
 
       - 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'
@@ -313,12 +498,20 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUBLAS=1
+          -DWHISPER_CUBLAS=${{ matrix.cublas }}
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
-      - name: Build
+      - 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'
@@ -328,7 +521,7 @@ jobs:
         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:
@@ -381,6 +574,14 @@ jobs:
     steps:
       - name: Clone
         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
@@ -393,9 +594,41 @@ jobs:
 
       - 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

.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 }}

.github/workflows/examples.yml

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

.gitignore

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

AUTHORS

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

CMakeLists.txt

@@ -1,6 +1,10 @@
 cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.4.2)
+# Allow for the creation of solution folders.
+set_property(GLOBAL PROPERTY USE_FOLDERS ON)
+
+project(whisper.cpp VERSION 1.6.0)
+set(SOVERSION 1)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -55,10 +59,13 @@ option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDA
 
 option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
 
-option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
-option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
-option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
-option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)
+option(WHISPER_NO_AVX                 "whisper: disable AVX"         OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2"        OFF)
+option(WHISPER_NO_AVX512              "whisper: disable AVX512"      ON)
+option(WHISPER_NO_AVX512_VBMI         "whisper: disable AVX512-VBMI" ON)
+option(WHISPER_NO_AVX512_VNNI         "whisper: disable AVX512-VNNI" ON)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"         OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c"        OFF)
 
 option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
 
@@ -68,13 +75,19 @@ if (APPLE)
     option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
     option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
     option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
+    option(WHISPER_METAL_EMBED_LIBRARY   "whisper: embed Metal library"          OFF)
 else()
-    option(WHISPER_BLAS                  "whisper: use BLAS libraries"  OFF)
-    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor" Generic)
-    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"     OFF)
-    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"  OFF)
-    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS" OFF)
-    option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
+    option(WHISPER_BLAS                  "whisper: use BLAS libraries"                        OFF)
+    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor"                       Generic)
+    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"                           OFF)
+    option(WHISPER_OPENBLAS_INTERFACE64  "whisper: use OpenBLAS w/ 64-bit interface"          OFF)
+    option(WHISPER_CUDA                  "whisper: support for CUDA"                          OFF)
+    option(WHISPER_CUBLAS                "whisper: support for CUDA (deprecated)"             OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               OFF)
+    option(WHISPER_MKL                   "whisper: use Intel Math Kernel Library (MKL)"       OFF)
+    option(WHISPER_SYCL                  "whisper: use SYCL"                                  OFF)
+    option(WHISPER_SYCL_F16              "whisper: use 16 bit floats for sycl calculations"   OFF)
 endif()
 
 option(WHISPER_PERF "whisper: enable perf timings" OFF)
@@ -105,6 +118,13 @@ endif()
 
 find_package(Threads REQUIRED)
 
+#compile flag sycl
+if (WHISPER_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
+
 # on APPLE
 if (APPLE)
     # include Accelerate framework
@@ -115,7 +135,7 @@ if (APPLE)
             message(STATUS "Accelerate framework found")
 
             set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64)
         else()
             message(FATAL_ERROR "Accelerate framework not found")
         endif()
@@ -145,8 +165,42 @@ if (APPLE)
 
         set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
 
-        # copy ggml-metal.metal to bin directory
+        # copy ggml-common.h and ggml-metal.metal to bin directory
+        configure_file(ggml-common.h    bin/ggml-common.h    COPYONLY)
         configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+        if (WHISPER_METAL_EMBED_LIBRARY)
+            enable_language(ASM)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)
+
+            set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+            set(COMMON_HEADER   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
+
+            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+            set(EMBED_METALLIB_SOURCE "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-combined.metal")
+
+            add_custom_command(
+                OUTPUT ${EMBED_METALLIB_SOURCE}
+                COMMAND sed -e "/^#include \\\"ggml-common.h\\\"/r ${COMMON_HEADER}" -e "/^#include \\\"ggml-common.h\\\"/d" ${METALLIB_SOURCE} > ${EMBED_METALLIB_SOURCE}
+                DEPENDS ${METALLIB_SOURCE} ${COMMON_HEADER}
+                COMMENT "Generating combined Metal library for embedding"
+            )
+
+            add_custom_command(
+                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".incbin \\\"${EMBED_METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
+                DEPENDS ${EMBED_METALLIB_SOURCE}
+                COMMENT "Generate assembly for embedded Metal library"
+            )
+
+            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
+        endif()
     endif()
 
     if (WHISPER_COREML)
@@ -170,30 +224,82 @@ endif()
 if (WHISPER_OPENBLAS)
     set(WHISPER_BLAS_VENDOR "OpenBLAS")
     set(WHISPER_BLAS ON)
+    # BLA_PKGCONFIG_BLAS is supported since CMake 3.25.
+    # FindBLAS.cmake pkg-config logic seems incomplete, because when
+    # BLA_SIZEOF_INTEGER is 8, then it should search for blas64 instead of blas.
+    # blas.pc/blas64.pc are not always provided, so let's be more specific
+    # and go with openblas.pc/openblas64.pc if WHISPER_OPENBLAS is on.
+    if (WHISPER_OPENBLAS_INTERFACE64)
+        set(WHISPER_BLAS_LIB "openblas64")
+    else ()
+        set(WHISPER_BLAS_LIB "openblas")
+    endif ()
+    set(BLA_PKGCONFIG_BLAS ${WHISPER_BLAS_LIB})
+    # OpenBLAS prebuilt libraries for Windows do not have "64" suffix in filename.
+    # (But .pc file has "64" suffix in filename for USE_64BITINT=1 Windows build.)
+    if (MSVC)
+        set(WHISPER_BLAS_LIB "openblas")
+    endif ()
 endif()
 
 if (WHISPER_BLAS)
-    if (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} ${BLAS_LIBRARIES})
+    if (NOT "$ENV{OPENBLAS_PATH}" STREQUAL "")
+        if (WHISPER_STATIC)
+            set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_STATIC_LIBRARY_PREFIX})
+            set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_STATIC_LIBRARY_SUFFIX})
         else ()
-            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
+            if (CMAKE_IMPORT_LIBRARY_SUFFIX)
+                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_IMPORT_LIBRARY_PREFIX})
+                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_IMPORT_LIBRARY_SUFFIX})
+            else ()
+                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_SHARED_LIBRARY_PREFIX})
+                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_SHARED_LIBRARY_SUFFIX})
+            endif ()
         endif ()
+        # OpenBLAS prebuilt libraries hardcode "lib" prefix in filename even on Windows
+        if (WHISPER_OPENBLAS)
+            set(WHISPER_BLAS_LIB_PREFIX "lib")
+        endif ()
+        message(STATUS "BLAS compatible library path provided")
+        set(BLAS_LIBRARIES "$ENV{OPENBLAS_PATH}/lib/${WHISPER_BLAS_LIB_PREFIX}${WHISPER_BLAS_LIB}${WHISPER_BLAS_LIB_SUFFIX}")
+        message(STATUS "Libraries ${BLAS_LIBRARIES}")
+        set(BLAS_INCLUDE_DIRS "$ENV{OPENBLAS_PATH}/include")
+        message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
+        if (NOT EXISTS "${BLAS_LIBRARIES}")
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH misdefined.")
+        endif ()
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+        include_directories(${BLAS_INCLUDE_DIRS})
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
     else ()
-        set(BLA_STATIC 1)
+        if (WHISPER_STATIC)
+            # FindBLAS.cmake pkg-config logic seems incomplete, because when
+            # BLA_STATIC is on, then it should use pkg_check_modules_static
+            # instead of pkg_check_modules.
+            # Some manual variable overriding may be necessary if you don't
+            # achieve desired results.
+            set(BLA_STATIC 1)
+        endif ()
         set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-        set(BLA_SIZEOF_INTEGER 8)
+        if (WHISPER_OPENBLAS_INTERFACE64)
+            set(BLA_SIZEOF_INTEGER 8)
+        else ()
+            set(BLA_SIZEOF_INTEGER 4)
+        endif()
         set(BLA_PREFER_PKGCONFIG 1)
         find_package(BLAS)
 
         if(BLAS_FOUND)
             message(STATUS "BLAS compatible library found")
             message(STATUS "Libraries ${BLAS_LIBRARIES}")
-            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
+            if (NOT DEFINED BLAS_INCLUDE_DIRS)
+                if (PKGC_BLAS_FOUND)
+                    set(BLAS_INCLUDE_DIRS "${PKGC_BLAS_INCLUDE_DIRS}")
+                else ()
+                    find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas)
+                endif()
+            endif()
+            message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
             include_directories(${BLAS_INCLUDE_DIRS})
             set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
@@ -203,7 +309,19 @@ if (WHISPER_BLAS)
     endif ()
 endif ()
 
+if (WHISPER_MKL)
+    find_package(MKL CONFIG REQUIRED PATHS $ENV{MKLROOT})
+    message(STATUS "Imported oneMKL targets: ${MKL_IMPORTED_TARGETS}")
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_BLAS_USE_MKL)
+endif()
+
 if (WHISPER_CUBLAS)
+    message(WARNING "WHISPER_CUBLAS is deprecated and will be removed in the future.\nUse WHISPER_CUDA instead")
+    set(WHISPER_CUDA ON)
+endif()
+
+if (WHISPER_CUDA)
     cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
@@ -213,16 +331,24 @@ if (WHISPER_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.h)
+        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.cu)
 
-        add_compile_definitions(GGML_USE_CUBLAS)
+        add_compile_definitions(GGML_USE_CUDA)
 
         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(FATAL_ERROR "cuBLAS not found")
     endif()
@@ -244,16 +370,18 @@ if (WHISPER_HIPBLAS)
 
     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)
+        set(GGML_HEADERS_ROCM "ggml-cuda.h")
 
+        file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
+
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
         if (WHISPER_STATIC)
             message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
         endif()
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
     else()
         message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
     endif()
@@ -278,6 +406,30 @@ if( WHISPER_OPENVINO )
     find_package(OpenVINO REQUIRED COMPONENTS Runtime)
 endif()
 
+if (WHISPER_SYCL)
+    if ( NOT DEFINED ENV{ONEAPI_ROOT})
+        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
+    endif()
+    #todo: AOT
+
+    find_package(IntelSYCL REQUIRED)
+    if (WHISPER_SYCL_F16)
+        add_compile_definitions(GGML_SYCL_F16)
+    endif()
+    add_compile_definitions(GGML_USE_SYCL)
+
+    add_compile_options(-I./) #include DPCT
+    add_compile_options(-I/${SYCL_INCLUDE_DIR})
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
+
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
+    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
+
+    set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+endif()
 # compiler flags
 
 if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
@@ -309,7 +461,8 @@ 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()
 
@@ -325,21 +478,35 @@ else()
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
         set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
         set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
-        if(NOT WHISPER_NO_AVX2)
+        if(NOT WHISPER_NO_AVX512)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX512")
+            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX512")
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX512")
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (NOT WHISPER_NO_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (NOT WHISPER_NO_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+        elseif(NOT WHISPER_NO_AVX2)
             set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
             set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
             set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-        else()
-            if(NOT WHISPER_NO_AVX)
-                set(CMAKE_CXX_FLAGS "${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()
+        elseif(NOT WHISPER_NO_AVX)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
         endif()
     else()
         if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
-            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")
@@ -347,6 +514,15 @@ else()
             if(NOT WHISPER_NO_AVX2)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
             endif()
+            if(NOT WHISPER_NO_AVX512)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw")
+                if(NOT WHISPER_NO_AVX512_VBMI)
+                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vbmi")
+                endif()
+                if(NOT WHISPER_NO_AVX512_VNNI)
+                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vnni")
+                endif()
+            endif()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
@@ -433,6 +609,7 @@ if (WHISPER_COREML)
     set_target_properties(${TARGET} PROPERTIES
         COMPILE_FLAGS "-fobjc-arc"
         )
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
 if (WHISPER_OPENVINO)
@@ -451,6 +628,7 @@ if (WHISPER_OPENVINO)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
 
     target_link_libraries(${TARGET} PRIVATE openvino::runtime)
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
 #
@@ -464,13 +642,28 @@ add_library(${TARGET}
     ggml.c
     ggml-alloc.h
     ggml-alloc.c
+    ggml-backend.h
+    ggml-backend.c
+    ggml-quants.h
+    ggml-quants.c
     ${GGML_SOURCES_METAL}
     ${GGML_SOURCES_CUDA}
     ${GGML_SOURCES_OPENCL}
+    ${GGML_SOURCES_SYCL}        ${GGML_HEADERS_SYCL}
+    ${GGML_SOURCES_ROCM}        ${GGML_HEADERS_ROCM}
     whisper.h
     whisper.cpp
     )
 
+include_directories (
+    .
+)
+# Set the version numbers
+set_target_properties(whisper PROPERTIES
+    VERSION ${PROJECT_VERSION}
+    SOVERSION ${SOVERSION}
+)
+
 include(DefaultTargetOptions)
 
 target_include_directories(${TARGET} PUBLIC
@@ -485,6 +678,10 @@ if (WHISPER_OPENVINO)
     target_link_libraries(${TARGET} PRIVATE whisper.openvino)
 endif()
 
+if (WHISPER_MKL)
+    target_link_libraries(${TARGET} PUBLIC MKL::MKL)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 
@@ -494,6 +691,7 @@ else()
 endif()
 
 if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_link_libraries(${TARGET} PUBLIC
         ${CMAKE_DL_LIBS}
         )
@@ -517,7 +715,13 @@ endif()
 
 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()
 
@@ -529,7 +733,8 @@ 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")
+set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 
 include(GNUInstallDirs)
 

LICENSE

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

Makefile

@@ -1,4 +1,4 @@
-default: main bench quantize
+default: main bench quantize server
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -18,6 +18,17 @@ ifndef NVCC_VERSION
 	endif
 endif
 
+# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
+# of non-gcc compilers don't have to provide g++ alias or wrapper.
+DEFCC  := cc
+DEFCXX := c++
+ifeq ($(origin CC),default)
+CC  := $(DEFCC)
+endif
+ifeq ($(origin CXX),default)
+CXX := $(DEFCXX)
+endif
+
 CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)
 
@@ -42,6 +53,12 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+ifdef MACOSX_DEPLOYMENT_TARGET
+	CFLAGS   += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+	CXXFLAGS += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+	LDFLAGS  += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
+endif
+
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
@@ -99,6 +116,16 @@ ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$
 	CXXFLAGS += -pthread
 endif
 
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -107,7 +134,7 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
 	else ifeq ($(UNAME_S),Linux)
 		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter MINGW32_NT% MINGW64_NT%,$(UNAME_S)))
+	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
@@ -115,42 +142,69 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 		CPUINFO_CMD := sysinfo -cpu
 	endif
 
+	# x86 ISA extensions (chronological order)
 	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')
+		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
+		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
+		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
+		AVX512F_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512F')
+		AVX512VBMI_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512VBMI')
+		AVX512VNNI_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX512_VNNI|AVX512VNNI')
+
+		# AVX-512 has many subsets, so let's make it easy to disable them all
+		ifneq ($(filter-out 0,$(WHISPER_NO_AVX512)),)
+			AVX512F_M :=
+			AVX512VBMI_M :=
+			AVX512VNNI_M :=
+		endif
+
 		ifneq (,$(SSE3_M))
 			CFLAGS   += -msse3
 			CXXFLAGS += -msse3
 		endif
 
-		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
 		ifneq (,$(SSSE3_M))
 			CFLAGS   += -mssse3
 			CXXFLAGS += -mssse3
 		endif
+
+		ifneq (,$(AVX_M))
+			CFLAGS   += -mavx
+			CXXFLAGS += -mavx
+		endif
+
+		ifneq (,$(F16C_M))
+			CFLAGS   += -mf16c
+			CXXFLAGS += -mf16c
+		endif
+
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
+		endif
+
+		ifneq (,$(AVX2_M))
+			CFLAGS   += -mavx2
+			CXXFLAGS += -mavx2
+		endif
+
+		ifneq (,$(AVX512F_M))
+			CFLAGS   += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
+			CXXFLAGS += -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw
+		endif
+
+		ifneq (,$(AVX512VBMI_M))
+			CFLAGS   += -mavx512vbmi
+			CXXFLAGS += -mavx512vbmi
+		endif
+
+		ifneq (,$(AVX512VNNI_M))
+			CFLAGS   += -mavx512vnni
+			CXXFLAGS += -mavx512vnni
+		endif
 	endif
 endif
 
@@ -169,6 +223,8 @@ ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
 		CFLAGS  += -DGGML_USE_ACCELERATE
+		CFLAGS  += -DACCELERATE_NEW_LAPACK
+		CFLAGS  += -DACCELERATE_LAPACK_ILP64
 		LDFLAGS += -framework Accelerate
 	endif
 endif
@@ -192,26 +248,54 @@ ifndef WHISPER_NO_METAL
 	endif
 endif
 
-ifdef WHISPER_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
-	LDFLAGS += -lopenblas
+ifneq ($(filter-out 0,$(WHISPER_OPENBLAS)),) # OpenBLAS
+	WHISPER_OPENBLAS_INTERFACE64 ?= 0 # use 32-bit interface by default
+	ifneq ($(filter-out 0,$(WHISPER_OPENBLAS_INTERFACE64)),)
+		WHISPER_BLAS_LIB := openblas64
+	else
+		WHISPER_BLAS_LIB := openblas
+	endif
+	ifneq ($(OPENBLAS_PATH),)
+		WHISPER_BLAS_CFLAGS  := -I$(OPENBLAS_PATH)/include
+		WHISPER_BLAS_LDFLAGS := -L$(OPENBLAS_PATH)/lib -l$(WHISPER_BLAS_LIB)
+	else
+		WHISPER_BLAS_LIB_PC_EXISTS := $(shell pkg-config --exists $(WHISPER_BLAS_LIB) && echo 1)
+		ifneq ($(filter-out 0,$(WHISPER_BLAS_LIB_PC_EXISTS)),)
+			WHISPER_BLAS_CFLAGS  := $(shell pkg-config --cflags $(WHISPER_BLAS_LIB))
+			WHISPER_BLAS_LDFLAGS := $(shell pkg-config --libs   $(WHISPER_BLAS_LIB))
+		else
+			WHISPER_BLAS_CFLAGS  := -I/usr/include/openblas
+			WHISPER_BLAS_LDFLAGS := -l$(WHISPER_BLAS_LIB)
+		endif
+	endif
+	CFLAGS  += $(WHISPER_BLAS_CFLAGS) -DGGML_USE_OPENBLAS
+	LDFLAGS += $(WHISPER_BLAS_LDFLAGS)
 endif
 
 ifdef WHISPER_CUBLAS
+# WHISPER_CUBLAS is deprecated and will be removed in the future
+	WHISPER_CUDA := 1
+endif
+
+ifdef WHISPER_CUDA
 	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
-		CUDA_ARCH_FLAG=native
+		CUDA_ARCH_FLAG ?= native
 	else
-		CUDA_ARCH_FLAG=all
+		CUDA_ARCH_FLAG ?= all
 	endif
 
-	CFLAGS      += -DGGML_USE_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
+	CFLAGS      += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	CXXFLAGS    += -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
 	WHISPER_OBJ += ggml-cuda.o
+	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
 
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -c $< -o $@
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 
@@ -219,14 +303,18 @@ ifdef WHISPER_HIPBLAS
 	ROCM_PATH   ?= /opt/rocm
 	HIPCC       ?= $(ROCM_PATH)/bin/hipcc
 	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
-	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	LDFLAGS     += -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
 	WHISPER_OBJ += ggml-cuda.o
+	WHISPER_OBJ += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 endif
 
@@ -291,6 +379,13 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
+ifdef WHISPER_CUBLAS
+$(info !!!!)
+$(info WHISPER_CUBLAS is deprecated and will be removed in the future. Use WHISPER_CUDA instead.)
+$(info !!!!)
+$(info )
+endif
+
 #
 # Build library
 #
@@ -301,7 +396,13 @@ ggml.o: ggml.c ggml.h ggml-cuda.h
 ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-WHISPER_OBJ += ggml-alloc.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 $@
@@ -323,16 +424,36 @@ ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 WHISPER_OBJ += ggml-metal.o
+
+ifdef WHISPER_METAL_EMBED_LIBRARY
+CFLAGS += -DGGML_METAL_EMBED_LIBRARY
+
+ggml-metal-embed.o: ggml-metal.metal ggml-common.h
+	@echo "Embedding Metal library"
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	$(eval TEMP_METALLIB=$(shell mktemp))
+	@sed "/^#include \"ggml-common.h\"/{r ggml-common.h"$$'\n'"d;}" ggml-metal.metal > $(TEMP_METALLIB)
+	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$(TEMP_METALLIB)\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f $(TEMP_ASSEMBLY) $(TEMP_METALLIB)
+
+WHISPER_OBJ += ggml-metal-embed.o
+endif
 endif
 
-libwhisper.a: ggml.o $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+libwhisper.a: $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a $(WHISPER_OBJ)
 
-libwhisper.so: ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
+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 lsp libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize server lsp libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -340,33 +461,36 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
+SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.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)
 
-lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-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)
+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-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)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples
@@ -411,9 +535,10 @@ 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 "==============================================="

Package.swift

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

README.md

@@ -6,7 +6,7 @@
 [![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: [v1.6.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -16,12 +16,10 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
-- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
 - Support for CPU-only inference
-- [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- [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)
 
@@ -35,11 +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)
@@ -64,22 +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
-intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
-the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
+The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
 
 ## Quick start
 
-First clone the repository.
+First clone the repository:
 
-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
@@ -94,7 +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
@@ -114,8 +111,8 @@ options:
   -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
   -ml N,     --max-len N         [0      ] maximum segment length in characters
   -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [2      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -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
@@ -132,6 +129,7 @@ options:
   -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
   -oj,       --output-json       [false  ] output result in a JSON file
+  -ojf,      --output-json-full  [false  ] include more information in the JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
@@ -143,7 +141,8 @@ options:
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
   -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of token
+  -ls,       --log-score         [false  ] log best decoder scores of tokens
+  -ng,       --no-gpu            [false  ] disable GPU
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -208,7 +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
 ```
 
@@ -234,18 +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
 
@@ -278,7 +278,8 @@ 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:
+  - MacOS Sonoma (version 14) or newer is recommended, as older versions of MacOS might experience issues with transcription hallucination.
+  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for this step:
     - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
     - To activate the environment, use: `conda activate py310-whisper`
 
@@ -304,8 +305,8 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
 - Run the examples as usual. For example:
 
-  ```bash
-  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```text
+  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -333,21 +334,23 @@ This can result in significant speedup in encoder performance. Here are the inst
 - First, setup python virtual env. and install python dependencies. Python 3.10 is recommended.
 
   Windows:
-  ```
+
+  ```powershell
   cd models
   python -m venv openvino_conv_env
   openvino_conv_env\Scripts\activate
   python -m pip install --upgrade pip
-  pip install -r openvino-conversion-requirements.txt
+  pip install -r requirements-openvino.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
+  pip install -r requirements-openvino.txt
   ```
 
 - Generate an OpenVINO encoder model. For example, to generate a `base.en` model, use:
@@ -356,7 +359,7 @@ This can result in significant speedup in encoder performance. Here are the inst
   python convert-whisper-to-openvino.py --model base.en
   ```
 
-  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as ggml models, as that
+  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as `ggml` models, as that
   is the default location that the OpenVINO extension will search at runtime.
 
 - Build `whisper.cpp` with OpenVINO support:
@@ -366,24 +369,28 @@ This can result in significant speedup in encoder performance. Here are the inst
   After downloading & extracting package onto your development system, set up required environment by sourcing setupvars script. For example:
 
   Linux:
+
   ```bash
   source /path/to/l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64/setupvars.sh
   ```
 
   Windows (cmd):
-  ```
+
+  ```powershell
   C:\Path\To\w_openvino_toolkit_windows_2023.0.0.10926.b4452d56304_x86_64\setupvars.bat
   ```
 
   And then build the project using cmake:
+
   ```bash
   cmake -B build -DWHISPER_OPENVINO=1
   cmake --build build -j --config Release
   ```
 
 - Run the examples as usual. For example:
-  ```bash
-  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+
+  ```text
+  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -399,19 +406,19 @@ This can result in significant speedup in encoder performance. Here are the inst
 
   The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
   cached for the next run.
-  
+
 For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
 
-## NVIDIA GPU support via cuBLAS
+## NVIDIA GPU support
 
-With NVIDIA cards the Encoder processing can to a large extent be offloaded to the GPU through cuBLAS.
+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:
+Now build `whisper.cpp` with CUDA support:
 
 ```
 make clean
-WHISPER_CUBLAS=1 make -j
+WHISPER_CUDA=1 make -j
 ```
 
 ## OpenCL GPU support via CLBlast
@@ -434,7 +441,6 @@ cmake -B build -DWHISPER_CLBLAST=ON
 cmake --build build -j --config Release
 ```
 
-
 Run all the examples as usual.
 
 ## BLAS CPU support via OpenBLAS
@@ -449,6 +455,53 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```
 
+## BLAS CPU support via Intel MKL
+
+Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
+First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
+
+Now build `whisper.cpp` with Intel MKL BLAS support:
+
+```
+source /opt/intel/oneapi/setvars.sh
+mkdir build
+cd build
+cmake -DWHISPER_MKL=ON ..
+WHISPER_MKL=1 make -j
+```
+
+## Docker
+
+### Prerequisites
+
+- Docker must be installed and running on your system.
+- Create a folder to store big models & intermediate files (ex. /whisper/models)
+
+### Images
+
+We have two Docker images available for this project:
+
+1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
+
+### Usage
+
+```shell
+# download model and persist it in a local folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./models/download-ggml-model.sh base /models"
+# transcribe an audio file
+docker run -it --rm \
+  -v path/to/models:/models \
+  -v path/to/audios:/audios \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
+# transcribe an audio file in samples folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
+```
+
 ## Limitations
 
 - Inference only
@@ -461,7 +514,7 @@ 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'
@@ -533,6 +586,7 @@ 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
@@ -541,7 +595,7 @@ 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
 ```
@@ -553,7 +607,7 @@ 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
 ```
 
@@ -563,8 +617,8 @@ 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'
 ...
@@ -587,8 +641,8 @@ 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'
 ...
@@ -658,7 +712,7 @@ 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
@@ -668,7 +722,7 @@ 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
@@ -678,7 +732,7 @@ 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
@@ -690,10 +744,10 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ## Video comparison of different models
 
-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:
+Use the [scripts/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/scripts/bench-wts.sh) script to generate a video in the following format:
 
-```java
-./extra/bench-wts.sh samples/jfk.wav
+```bash
+./scripts/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```
 
@@ -714,15 +768,14 @@ Additionally a script to run whisper.cpp with different models and audio files i
 You can run it with the following command, by default it will run against any standard model in the models folder.
 
 ```bash
-python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+python3 scripts/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
 ```
 
 It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
 
 It outputs a csv file with the results of the benchmarking.
 
-
-## ggml format
+## `ggml` format
 
 The original models are converted to a custom binary format. This allows to pack everything needed into a single file:
 
@@ -737,49 +790,51 @@ 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)
+  - [AIWintermuteAI/whispercpp](https://github.com/AIWintermuteAI/whispercpp) (Updated fork of aarnphm/whispercpp)
   - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
-- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
-- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
+- [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
 There are various examples of using the library for different projects in the [examples](examples) folder.
 Some of the examples are even ported to run in the browser using WebAssembly. Check them out!
 
-| Example | Web | Description |
-| ---     | --- | ---         |
-| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
-| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
-| [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
-| [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
-| [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)
 

README_sycl.md

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

bindings/go/Makefile

@@ -1,9 +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

@@ -123,6 +123,11 @@ 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
 
@@ -147,6 +152,7 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
 	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
+	str += fmt.Sprintf(" initial_prompt=%s", C.GoString(p.initial_prompt))
 	if p.translate {
 		str += " translate"
 	}

bindings/go/pkg/whisper/context.go

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

bindings/go/pkg/whisper/interface.go

@@ -38,17 +38,18 @@ 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
-	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
+	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

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

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

bindings/java/src/main/java/io/github/ggerganov/whispercpp/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;
 
@@ -145,6 +148,9 @@ public class WhisperFullParams extends Structure {
         tdrz_enable = enable ? CBool.TRUE : CBool.FALSE;
     }
 
+    /** Regular expression matching tokens to suppress. */
+    public String suppress_regex;
+
     /** Tokens to provide to the whisper decoder as an initial prompt.
      * These are prepended to any existing text context from a previous call. */
     public String initial_prompt;
@@ -304,18 +310,25 @@ 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",
+                "tdrz_enable", "suppress_regex", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
                 "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
                 "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",
                 "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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.4.2",
+  "version": "1.6.0",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/extconf.rb

@@ -3,8 +3,15 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-common.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
 
 

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

@@ -24,9 +24,9 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     // select which device to run the Core ML model on
     MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
-    config.computeUnits = MLComputeUnitsCPUAndGPU;
+    // config.computeUnits = MLComputeUnitsCPUAndGPU;
     //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
-    //config.computeUnits = MLComputeUnitsAll;
+    config.computeUnits = MLComputeUnitsAll;
 
     const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 
@@ -48,13 +48,15 @@ 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
     ];

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 +27,8 @@ add_library(${TARGET} STATIC
     common.cpp
     common-ggml.h
     common-ggml.cpp
+    grammar-parser.h
+    grammar-parser.cpp
     )
 
 include(DefaultTargetOptions)
@@ -30,6 +36,7 @@ include(DefaultTargetOptions)
 target_link_libraries(${TARGET} PRIVATE whisper)
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 
 if (WHISPER_SDL2)
     # common-sdl
@@ -47,27 +54,63 @@ if (WHISPER_SDL2)
     target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
 
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()
 
+# add json lib
+add_library(json_cpp INTERFACE)
+target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
+
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
     add_subdirectory(whisper.wasm)
+    set_target_properties(libmain PROPERTIES FOLDER "libs")
     add_subdirectory(stream.wasm)
+    set_target_properties(libstream PROPERTIES FOLDER "libs")
     add_subdirectory(command.wasm)
+    set_target_properties(libcommand PROPERTIES FOLDER "libs")
     add_subdirectory(talk.wasm)
+    set_target_properties(libtalk PROPERTIES FOLDER "libs")
     add_subdirectory(bench.wasm)
+    set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
     add_subdirectory(addon.node)
+    set_target_properties(addon.node PROPERTIES FOLDER "examples")
 else()
     add_subdirectory(main)
+    set_target_properties(main PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
     add_subdirectory(stream)
+    set_target_properties(stream PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)
+    add_subdirectory(server)
+    set_target_properties(server PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
     add_subdirectory(command)
+    set_target_properties(command PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)
     add_subdirectory(bench)
+    set_target_properties(bench PROPERTIES FOLDER "examples")
     add_subdirectory(quantize)
+    set_target_properties(quantize PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
     add_subdirectory(talk)
+    set_target_properties(talk PROPERTIES FOLDER "examples")
     add_subdirectory(talk-llama)
+    set_target_properties(talk-llama PROPERTIES FOLDER "examples")
     add_subdirectory(lsp)
+    set_target_properties(lsp PROPERTIES FOLDER "examples")
+    if (LLAMA_SYCL)
+        add_subdirectory(sycl)
+        set_target_properties(sycl PROPERTIES FOLDER "examples")
+    endif()
+endif (WHISPER_SDL2)
 endif()
+
+if (WHISPER_SDL2)
+    add_subdirectory(wchess)
+    set_target_properties(wchess PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)

examples/addon.node/CMakeLists.txt

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

examples/addon.node/README.md

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

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

@@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
   __dirname,
-  "../../../build/Release/whisper-addon"
+  "../../../build/Release/addon.node"
 ));
 const { promisify } = require("util");
 
@@ -11,6 +11,12 @@ const whisperParamsMock = {
   language: "en",
   model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+  use_gpu: true,
+  no_prints: true,
+  comma_in_time: false,
+  translate: true,
+  no_timestamps: false,
+  audio_ctx: 0,
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -19,6 +19,7 @@ struct whisper_params {
     int32_t max_len      = 0;
     int32_t best_of      = 5;
     int32_t beam_size    = -1;
+    int32_t audio_ctx    = 0;
 
     float word_thold    = 0.01f;
     float entropy_thold = 2.4f;
@@ -36,6 +37,9 @@ struct whisper_params {
     bool print_colors   = false;
     bool print_progress = false;
     bool no_timestamps  = false;
+    bool no_prints      = false;
+    bool use_gpu        = true;
+    bool comma_in_time  = true;
 
     std::string language = "en";
     std::string prompt;
@@ -43,6 +47,8 @@ struct whisper_params {
 
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
+
+    std::vector<float> pcmf32 = {}; // mono-channel F32 PCM
 };
 
 struct whisper_print_user_data {
@@ -51,27 +57,6 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma = false) {
-    int64_t msec = t * 10;
-    int64_t hr = msec / (1000 * 60 * 60);
-    msec = msec - hr * (1000 * 60 * 60);
-    int64_t min = msec / (1000 * 60);
-    msec = msec - min * (1000 * 60);
-    int64_t sec = msec / 1000;
-    msec = msec - sec * 1000;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
-
-    return std::string(buf);
-}
-
-int timestamp_to_sample(int64_t t, int n_samples) {
-    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
-}
-
 void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
@@ -103,8 +88,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
         if (params.diarize && pcmf32s.size() == 2) {
             const int64_t n_samples = pcmf32s[0].size();
 
-            const int64_t is0 = timestamp_to_sample(t0, n_samples);
-            const int64_t is1 = timestamp_to_sample(t1, n_samples);
+            const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
+            const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
 
             double energy0 = 0.0f;
             double energy1 = 0.0f;
@@ -140,9 +125,15 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
+void cb_log_disable(enum ggml_log_level, const char *, void *) {}
+
 int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
-    if (params.fname_inp.empty()) {
-        fprintf(stderr, "error: no input files specified\n");
+    if (params.no_prints) {
+        whisper_log_set(cb_log_disable, NULL);
+    }
+
+    if (params.fname_inp.empty() && params.pcmf32.empty()) {
+        fprintf(stderr, "error: no input files or audio buffer specified\n");
         return 2;
     }
 
@@ -153,13 +144,23 @@ 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");
         return 3;
     }
 
+    // if params.pcmf32 is provided, set params.fname_inp to "buffer"
+    // this is simpler than further modifications in the code
+    if (!params.pcmf32.empty()) {
+        fprintf(stderr, "info: using audio buffer as input\n");
+        params.fname_inp.clear();
+        params.fname_inp.emplace_back("buffer");
+    }
+
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
         const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -167,20 +168,25 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         std::vector<float> pcmf32; // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
-            continue;
+        // read the input audio file if params.pcmf32 is not provided
+        if (params.pcmf32.empty()) {
+            if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+                fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+                continue;
+            }
+        } else {
+            pcmf32 = params.pcmf32;
         }
 
         // print system information
-        {
+        if (!params.no_prints) {
             fprintf(stderr, "\n");
             fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
                     params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
         }
 
         // print some info about the processing
-        {
+        if (!params.no_prints) {
             fprintf(stderr, "\n");
             if (!whisper_is_multilingual(ctx)) {
                 if (params.language != "en" || params.translate) {
@@ -189,12 +195,13 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
                     fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                 }
             }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d, audio_ctx = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                     params.n_threads, params.n_processors,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
-                    params.no_timestamps ? 0 : 1);
+                    params.no_timestamps ? 0 : 1,
+                    params.audio_ctx);
 
             fprintf(stderr, "\n");
         }
@@ -221,6 +228,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+            wparams.audio_ctx        = params.audio_ctx;
 
             wparams.speed_up         = params.speed_up;
 
@@ -229,6 +237,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
             wparams.initial_prompt   = params.prompt.c_str();
 
+            wparams.no_timestamps    = params.no_timestamps;
+
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
             // this callback is called on each new segment
@@ -264,8 +274,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-        result[i].emplace_back(to_timestamp(t0, true));
-        result[i].emplace_back(to_timestamp(t1, true));
+        result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
+        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
         result[i].emplace_back(text);
     }
 
@@ -315,10 +325,32 @@ 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>();
+  bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
+  bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
+  int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
+  bool comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
+
+  Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
+  std::vector<float> pcmf32_vec;
+  if (pcmf32Value.IsTypedArray()) {
+    Napi::Float32Array pcmf32 = pcmf32Value.As<Napi::Float32Array>();
+    size_t length = pcmf32.ElementLength();
+    pcmf32_vec.reserve(length);
+    for (size_t i = 0; i < length; i++) {
+      pcmf32_vec.push_back(pcmf32[i]);
+    }
+  }
 
   params.language = language;
   params.model = model;
   params.fname_inp.emplace_back(input);
+  params.use_gpu = use_gpu;
+  params.no_prints = no_prints;
+  params.no_timestamps = no_timestamps;
+  params.audio_ctx = audio_ctx;
+  params.pcmf32 = pcmf32_vec;
+  params.comma_in_time = comma_in_time;
 
   Napi::Function callback = info[1].As<Napi::Function>();
   Worker* worker = new Worker(callback, params);

examples/addon.node/index.js

@@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
   __dirname,
-  "../../build/Release/whisper-addon"
+  "../../build/Release/addon.node"
 ));
 const { promisify } = require("util");
 
@@ -10,14 +10,26 @@ const whisperAsync = promisify(whisper);
 const whisperParams = {
   language: "en",
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
-  fname_inp: "../../samples/jfk.wav",
+  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  use_gpu: true,
+  no_prints: true,
+  comma_in_time: false,
+  translate: true,
+  no_timestamps: false,
+  audio_ctx: 0,
 };
 
 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
   arguments.reduce((pre, item) => {
     if (item.startsWith("--")) {
-      return [...pre, item.slice(2).split("=")];
+      const [key, value] = item.slice(2).split("=");
+      if (key === "audio_ctx") {
+        whisperParams[key] = parseInt(value);
+      } else {
+        whisperParams[key] = value;
+      }
+      return pre;
     }
     return pre;
   }, [])
@@ -32,5 +44,6 @@ for (const key in params) {
 console.log("whisperParams =", whisperParams);
 
 whisperAsync(whisperParams).then((result) => {
-  console.log(`Result from whisper: ${result}`);
+  console.log();
+  console.log(result);
 });

examples/addon.node/package.json

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

examples/bench.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

@@ -8,9 +8,12 @@
 // command-line parameters
 struct whisper_params {
     int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
+    int32_t what = 0; // what to benchmark: 0 - whisper encoder, 1 - memcpy, 2 - ggml_mul_mat
 
     std::string model = "models/ggml-base.en.bin";
+
+    bool use_gpu    = true;
+    bool flash_attn = false;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -23,9 +26,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
-        else if (arg == "-m" || arg == "--model")   { params.model     = argv[++i]; }
-        else if (arg == "-w" || arg == "--what")    { params.what     = atoi(argv[++i]); }
+        else if (arg == "-t"  || arg == "--threads")    { params.n_threads  = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")     { params.use_gpu    = false; }
+        else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -45,6 +50,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
+    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
     fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
@@ -54,7 +61,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
 int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams = whisper_context_default_params();
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     {
         fprintf(stderr, "\n");
@@ -66,13 +78,15 @@ int whisper_bench_full(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;
     }
 
@@ -81,13 +95,13 @@ int whisper_bench_full(const whisper_params & params) {
 
     // prompt heat
     if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        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 encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
         return 4;
     }
 
@@ -95,20 +109,30 @@ int whisper_bench_full(const whisper_params & params) {
 
     // actual run
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
         return 4;
     }
 
-    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 encode model: %d\n", ret);
+    // 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;
         }
     }
 
-    for (int i = 0; i < 256; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+    // 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;
         }
     }

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/README.md

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

examples/command/command.cpp

@@ -9,6 +9,7 @@
 #include "common-sdl.h"
 #include "common.h"
 #include "whisper.h"
+#include "grammar-parser.h"
 
 #include <sstream>
 #include <cassert>
@@ -30,20 +31,31 @@ 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;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
     std::string fname_out;
     std::string commands;
     std::string prompt;
+    std::string context;
+    std::string grammar;
+
+    // A regular expression that matches tokens to suppress
+    std::string suppress_regex;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -68,11 +80,17 @@ 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 == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
         else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
         else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
+        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -101,21 +119,38 @@ 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, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
     fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
     fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
+    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
+    fprintf(stderr, "  --suppress-regex REGEX       [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, 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 +158,43 @@ 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();
+
+    wparams.suppress_regex = params.suppress_regex.c_str();
+
+    const auto & grammar_parsed = params.grammar_parsed;
+    auto grammar_rules = grammar_parsed.c_rules();
+
+    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
+        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
+            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, grammar_rule.c_str());
+        } else {
+            wparams.grammar_rules   = grammar_rules.data();
+            wparams.n_grammar_rules = grammar_rules.size();
+            wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
+            wparams.grammar_penalty = params.grammar_penalty;
+        }
+    }
 
     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 +203,17 @@ 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 +304,7 @@ 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 +472,9 @@ 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 +512,7 @@ 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 +548,27 @@ 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 +601,11 @@ 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 +626,30 @@ 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 +662,16 @@ 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 +698,12 @@ 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;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     // print some info about the processing
     {
@@ -648,12 +741,36 @@ 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 (is_file_exist(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            ret_val = 1;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (ret_val == 0) {
+        if (!params.commands.empty()) {
+            ret_val = process_command_list(ctx, audio, params);
+        } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
+            ret_val = always_prompt_transcription(ctx, audio, params);
+        } else {
+            ret_val = process_general_transcription(ctx, audio, params);
+        }
     }
 
     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 +53,25 @@ 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_IQ2_S:
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:
+        case GGML_FTYPE_MOSTLY_IQ3_S:
+        case GGML_FTYPE_MOSTLY_IQ1_S:
+        case GGML_FTYPE_MOSTLY_IQ4_NL:
+        case GGML_FTYPE_MOSTLY_IQ4_XS:
+        case GGML_FTYPE_MOSTLY_IQ1_M:
+        case GGML_FTYPE_MOSTLY_BF16:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -77,8 +92,6 @@ bool ggml_common_quantize_0(
     std::vector<ggml_fp16_t> data_f16;
     std::vector<float>       data_f32;
 
-    std::vector<int64_t> hist_all(1 << 4, 0);
-
     while (true) {
         int32_t n_dims;
         int32_t length;
@@ -163,41 +176,39 @@ bool ggml_common_quantize_0(
             work.resize(nelements); // for quantization
 
             size_t cur_size = 0;
-            std::vector<int64_t> hist_cur(1 << 4, 0);
-
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
-                    {
-                        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], nullptr);
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_I8:
                 case GGML_TYPE_I16:
                 case GGML_TYPE_I32:
+                case GGML_TYPE_I64:
+                case GGML_TYPE_F64:
                 case GGML_TYPE_Q8_1:
-                case GGML_TYPE_Q2_K:
-                case GGML_TYPE_Q3_K:
-                case GGML_TYPE_Q4_K:
-                case GGML_TYPE_Q5_K:
-                case GGML_TYPE_Q6_K:
                 case GGML_TYPE_Q8_K:
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ2_XS:
+                case GGML_TYPE_IQ2_S:
+                case GGML_TYPE_IQ3_XXS:
+                case GGML_TYPE_IQ3_S:
+                case GGML_TYPE_IQ1_S:
+                case GGML_TYPE_IQ4_NL:
+                case GGML_TYPE_IQ4_XS:
+                case GGML_TYPE_IQ1_M:
+                case GGML_TYPE_BF16:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
@@ -208,15 +219,7 @@ bool ggml_common_quantize_0(
             fout.write(reinterpret_cast<char *>(work.data()), cur_size);
             total_size_new += cur_size;
 
-            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
-            for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                hist_all[i] += hist_cur[i];
-            }
-
-            for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                printf("%5.3f ", hist_cur[i] / (float)nelements);
-            }
-            printf("\n");
+            printf("size = %8.2f MB -> %8.2f MB\n", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
         } else {
             printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
             fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
@@ -229,18 +232,5 @@ bool ggml_common_quantize_0(
     printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
     printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
 
-    {
-        int64_t sum_all = 0;
-        for (int i = 0; i < (int) hist_all.size(); ++i) {
-            sum_all += hist_all[i];
-        }
-
-        printf("%s: hist: ", __func__);
-        for (int i = 0; i < (int) hist_all.size(); ++i) {
-            printf("%5.3f ", hist_all[i] / (float)sum_all);
-        }
-        printf("\n");
-    }
-
     return true;
 }

examples/common-sdl.cpp

@@ -139,10 +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

@@ -19,6 +19,11 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#ifdef _WIN32
+#include <fcntl.h>
+#include <io.h>
+#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] != '-') {
@@ -38,12 +43,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-t" || arg == "--threads") {
             params.n_threads = 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 == "-p" || arg == "--prompt") {
             params.prompt = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-n" || arg == "--n_predict") {
             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::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_p") {
@@ -56,6 +61,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             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(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 = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-i" || arg == "--interactive") {
@@ -97,7 +108,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
     fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
     fprintf(stderr, "                        prompt to start generation with (default: random)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
@@ -111,6 +121,9 @@ 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");
@@ -607,12 +620,31 @@ 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
 
     if (fname == "-") {
         {
+            #ifdef _WIN32
+            _setmode(_fileno(stdin), _O_BINARY);
+            #endif
+
             uint8_t buf[1024];
             while (true)
             {
@@ -631,6 +663,12 @@ 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;
@@ -638,21 +676,25 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
 
     if (wav.channels != 1 && wav.channels != 2) {
         fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
     if (stereo && wav.channels != 2) {
         fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.sampleRate != COMMON_SAMPLE_RATE) {
         fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.bitsPerSample != 16) {
         fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
@@ -807,3 +849,48 @@ void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
     fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*whisper_sample_rate)/100)));
+}
+
+bool is_file_exist(const char *fileName)
+{
+    std::ifstream infile(fileName);
+    return infile.good();
+}
+
+bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id)
+{
+    std::ofstream speak_file(path.c_str());
+    if (speak_file.fail()) {
+        fprintf(stderr, "%s: failed to open speak_file\n", __func__);
+        return false;
+    } else {
+        speak_file.write(text.c_str(), text.size());
+        speak_file.close();
+        int ret = system((command + " " + std::to_string(voice_id) + " " + path).c_str());
+        if (ret != 0) {
+            fprintf(stderr, "%s: failed to speak\n", __func__);
+            return false;
+        }
+    }
+    return true;
+}

examples/common.h

@@ -17,10 +17,15 @@
 //
 
 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;
@@ -35,8 +40,6 @@ struct gpt_params {
 
     bool    interactive      = false;
     int32_t interactive_port = -1;
-
-    int32_t n_gpu_layers     = 0;
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@@ -132,7 +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(
@@ -178,7 +185,7 @@ private:
     // It is assumed that PCM data is normalized to a range from -1 to 1
     bool write_audio(const float * data, size_t length) {
         for (size_t i = 0; i < length; ++i) {
-            const auto intSample = static_cast<const int16_t>(data[i] * 32767);
+            const int16_t intSample = data[i] * 32767;
             file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
             dataSize += sizeof(int16_t);
         }
@@ -274,3 +281,31 @@ struct sam_params {
 bool sam_params_parse(int argc, char ** argv, sam_params & params);
 
 void sam_print_usage(int argc, char ** argv, const sam_params & params);
+
+//
+// Terminal utils
+//
+
+
+// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
+// Lowest is red, middle is yellow, highest is green.
+const std::vector<std::string> k_colors = {
+    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
+    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
+};
+
+//
+// Other utils
+//
+
+// convert timestamp to string, 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false);
+
+// given a timestamp get the sample
+int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate);
+
+// check if file exists using ifstream
+bool is_file_exist(const char *fileName);
+
+// write text to file, and call system("command voice_id file")
+bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id);

examples/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 preceding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<whisper_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({WHISPER_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({WHISPER_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<whisper_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({WHISPER_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({WHISPER_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    bool is_char_element(whisper_grammar_element elem) {
+        switch (elem.type) {
+            case WHISPER_GRETYPE_CHAR:           return true;
+            case WHISPER_GRETYPE_CHAR_NOT:       return true;
+            case WHISPER_GRETYPE_CHAR_ALT:       return true;
+            case WHISPER_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
+                case WHISPER_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case WHISPER_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case WHISPER_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case WHISPER_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case WHISPER_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                case WHISPER_GRETYPE_ALT:
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                case WHISPER_GRETYPE_CHAR_NOT:
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<whisper_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != WHISPER_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with WHISPER_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            whisper_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case WHISPER_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case WHISPER_GRETYPE_CHAR_ALT:
+                    case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
+        std::vector<const whisper_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

examples/grammar-parser.h

@@ -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();
     }
 }
 
@@ -33,9 +34,6 @@ async function fetchRemote(url, cbProgress, cbPrint) {
         url,
         {
             method: 'GET',
-            headers: {
-                'Content-Type': 'application/octet-stream',
-            },
         }
     );
 

examples/livestream.sh

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

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

examples/lsp/lsp.cpp

@@ -30,6 +30,8 @@ struct whisper_params {
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
+    bool use_gpu       = true;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -72,6 +74,8 @@ 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 == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else {
@@ -102,6 +106,8 @@ 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, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "\n");
@@ -432,7 +438,12 @@ 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;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
     // init audio
 
     audio_async audio(30*1000);

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

@@ -1,10 +1,12 @@
 #include "common.h"
 
 #include "whisper.h"
+#include "grammar-parser.h"
 
 #include <cmath>
 #include <fstream>
 #include <cstdio>
+#include <regex>
 #include <string>
 #include <thread>
 #include <vector>
@@ -14,34 +16,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
-// Lowest is red, middle is yellow, highest is green.
-const std::vector<std::string> k_colors = {
-    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
-    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
-};
-
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma = false) {
-    int64_t msec = t * 10;
-    int64_t hr = msec / (1000 * 60 * 60);
-    msec = msec - hr * (1000 * 60 * 60);
-    int64_t min = msec / (1000 * 60);
-    msec = msec - min * (1000 * 60);
-    int64_t sec = msec / 1000;
-    msec = msec - sec * 1000;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
-
-    return std::string(buf);
-}
-
-int timestamp_to_sample(int64_t t, int n_samples) {
-    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
-}
-
 // helper function to replace substrings
 void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     for (size_t pos = 0; ; pos += replace.length()) {
@@ -54,20 +28,24 @@ void replace_all(std::string & s, const std::string & search, const std::string
 
 // command-line parameters
 struct whisper_params {
-    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors =  1;
-    int32_t offset_t_ms  =  0;
-    int32_t offset_n     =  0;
-    int32_t duration_ms  =  0;
-    int32_t progress_step =  5;
-    int32_t max_context  = -1;
-    int32_t max_len      =  0;
-    int32_t best_of      =  2;
-    int32_t beam_size    = -1;
+    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors  = 1;
+    int32_t offset_t_ms   = 0;
+    int32_t offset_n      = 0;
+    int32_t duration_ms   = 0;
+    int32_t progress_step = 5;
+    int32_t max_context   = -1;
+    int32_t max_len       = 0;
+    int32_t best_of       = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size     = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t audio_ctx     = 0;
 
-    float word_thold    =  0.01f;
-    float entropy_thold =  2.40f;
-    float logprob_thold = -1.00f;
+    float word_thold      =  0.01f;
+    float entropy_thold   =  2.40f;
+    float logprob_thold   = -1.00f;
+    float grammar_penalty = 100.0f;
+    float temperature     = 0.0f;
+    float temperature_inc = 0.2f;
 
     bool speed_up        = false;
     bool debug_mode      = false;
@@ -85,28 +63,48 @@ struct whisper_params {
     bool output_jsn      = false;
     bool output_jsn_full = false;
     bool output_lrc      = false;
+    bool no_prints       = false;
     bool print_special   = false;
     bool print_colors    = false;
     bool print_progress  = false;
     bool no_timestamps   = false;
     bool log_score       = false;
+    bool use_gpu         = true;
+    bool flash_attn      = false;
 
     std::string language  = "en";
     std::string prompt;
     std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model     = "models/ggml-base.en.bin";
+    std::string grammar;
+    std::string grammar_rule;
 
     // [TDRZ] speaker turn string
     std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
 
+    // A regular expression that matches tokens to suppress
+    std::string suppress_regex;
+
     std::string openvino_encode_device = "CPU";
 
+    std::string dtw = "";
+
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
+
+    grammar_parser::parse_state grammar_parsed;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
+char* whisper_param_turn_lowercase(char* in){
+    int string_len = strlen(in);
+    for(int i = 0; i < string_len; i++){
+        *(in+i) = tolower((unsigned char)*(in+i));
+    }
+    return in;
+}
+
 bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
@@ -134,9 +132,12 @@ 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-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
+        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
         // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
@@ -154,17 +155,25 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
         else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
         else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
         else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
         else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
         else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
         else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = whisper_param_turn_lowercase(argv[++i]); }
         else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
         else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
         else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
         else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
         else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score = true; }
+        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
+        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
+        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = argv[++i]; }
+        else if (                  arg == "--grammar")         { params.grammar         = argv[++i]; }
+        else if (                  arg == "--grammar-rule")    { params.grammar_rule    = argv[++i]; }
+        else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -191,9 +200,12 @@ 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, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
+    fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
     // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
@@ -210,17 +222,25 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
     fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
+    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
     fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
     fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
     fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
     fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt (max n_text_ctx/2 tokens)\n",       params.prompt.c_str());
     fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
     fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
     fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
+    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
+    fprintf(stderr, "  --grammar-rule RULE            [%-7s] top-level GBNF grammar rule name\n",               params.grammar_rule.c_str());
+    fprintf(stderr, "  --grammar-penalty N            [%-7.1f] scales down logits of nongrammar tokens\n",      params.grammar_penalty);
     fprintf(stderr, "\n");
 }
 
@@ -235,8 +255,8 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
 
-    const int64_t is0 = timestamp_to_sample(t0, n_samples);
-    const int64_t is1 = timestamp_to_sample(t1, n_samples);
+    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
 
     double energy0 = 0.0f;
     double energy1 = 0.0f;
@@ -460,6 +480,38 @@ char *escape_double_quotes_and_backslashes(const char *str) {
     return escaped;
 }
 
+// double quote should be escaped by another double quote. (rfc4180)
+char *escape_double_quotes_in_csv(const char *str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped[pos++] = '"';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
 bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
@@ -481,7 +533,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-        char * text_escaped = escape_double_quotes_and_backslashes(text);
+        char * text_escaped = escape_double_quotes_in_csv(text);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
         fout << 10 * t0 << "," << 10 * t1 << ",";
@@ -660,7 +712,8 @@ bool output_json(
                                     times_o(token.t0, token.t1, false);
                                 }
                                 value_i("id", token.id, false);
-                                value_f("p", token.p, true);
+                                value_f("p", token.p, false);
+                                value_f("t_dtw", token.t_dtw, true);
                             end_obj(j == (n - 1));
                         }
                         end_arr(!params.diarize && !params.tinydiarize);
@@ -849,14 +902,59 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
+
+void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
+    // If the only argument starts with "@", read arguments line-by-line
+    // from the given file.
+    std::vector<std::string> vec_args;
+    if (argc == 2 && argv != nullptr && argv[1] != nullptr && argv[1][0] == '@') {
+        // Save the name of the executable.
+        vec_args.push_back(argv[0]);
+
+        // Open the response file.
+        char const * rspfile = argv[1] + sizeof(char);
+        std::ifstream fin(rspfile);
+        if (fin.is_open() == false) {
+            fprintf(stderr, "error: response file '%s' not found\n", rspfile);
+            return 1;
+        }
+
+        // Read the entire response file.
+        std::string line;
+        while (std::getline(fin, line)) {
+            vec_args.push_back(line);
+        }
+
+        // Use the contents of the response file as the command-line arguments.
+        argc = static_cast<int>(vec_args.size());
+        argv = static_cast<char **>(alloca(argc * sizeof (char *)));
+        for (int i = 0; i < argc; ++i) {
+            argv[i] = const_cast<char *>(vec_args[i].c_str());
+        }
+    }
+
     if (whisper_params_parse(argc, argv, params) == false) {
         whisper_print_usage(argc, argv, params);
         return 1;
     }
 
+    // remove non-existent files
+    for (auto it = params.fname_inp.begin(); it != params.fname_inp.end();) {
+        const auto fname_inp = it->c_str();
+
+        if (*it != "-" && !is_file_exist(fname_inp)) {
+            fprintf(stderr, "error: input file not found '%s'\n", fname_inp);
+            it = params.fname_inp.erase(it);
+            continue;
+        }
+
+        it++;
+    }
+
     if (params.fname_inp.empty()) {
         fprintf(stderr, "error: no input files specified\n");
         whisper_print_usage(argc, argv, params);
@@ -875,9 +973,40 @@ 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;
+    cparams.flash_attn = params.flash_attn;
+
+    if (!params.dtw.empty()) {
+        cparams.dtw_token_timestamps = true;
+        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
+
+        if (params.dtw == "tiny")      cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
+        if (params.dtw == "tiny.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
+        if (params.dtw == "base")      cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
+        if (params.dtw == "base.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
+        if (params.dtw == "small")     cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
+        if (params.dtw == "small.en")  cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
+        if (params.dtw == "medium")    cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
+        if (params.dtw == "medium.en") cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
+        if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
+        if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
+        if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
+
+        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
+            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
+            return 3;
+        }
+    }
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     if (ctx == nullptr) {
         fprintf(stderr, "error: failed to initialize whisper context\n");
@@ -887,6 +1016,29 @@ int main(int argc, char ** argv) {
     // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
     whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
 
+    if (!params.grammar.empty()) {
+        auto & grammar = params.grammar_parsed;
+        if (is_file_exist(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            fprintf(stderr, "error: failed to parse grammar \"%s\"\n", params.grammar.c_str());
+            return 4;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -899,29 +1051,28 @@ 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, " : "",
@@ -934,7 +1085,8 @@ int main(int argc, char ** argv) {
         {
             whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
-            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+            const bool use_grammar = (!params.grammar_parsed.rules.empty() && !params.grammar_rule.empty());
+            wparams.strategy = (params.beam_size > 1 || use_grammar) ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
 
             wparams.print_realtime   = false;
             wparams.print_progress   = params.print_progress;
@@ -952,23 +1104,44 @@ int main(int argc, char ** argv) {
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
             wparams.split_on_word    = params.split_on_word;
+            wparams.audio_ctx        = params.audio_ctx;
 
             wparams.speed_up         = params.speed_up;
             wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 
+            wparams.suppress_regex   = params.suppress_regex.empty() ? nullptr : params.suppress_regex.c_str();
+
             wparams.initial_prompt   = params.prompt.c_str();
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : params.temperature_inc;
+            wparams.temperature      = params.temperature;
+
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
+            wparams.no_timestamps    = params.no_timestamps;
+
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
+            const auto & grammar_parsed = params.grammar_parsed;
+            auto grammar_rules = grammar_parsed.c_rules();
+
+            if (use_grammar) {
+                if (grammar_parsed.symbol_ids.find(params.grammar_rule) == grammar_parsed.symbol_ids.end()) {
+                    fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, params.grammar_rule.c_str());
+                } else {
+                    wparams.grammar_rules = grammar_rules.data();
+                    wparams.n_grammar_rules = grammar_rules.size();
+                    wparams.i_start_rule = grammar_parsed.symbol_ids.at(params.grammar_rule);
+                    wparams.grammar_penalty = params.grammar_penalty;
+                }
+            }
+
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
                 wparams.new_segment_callback           = whisper_print_segment_callback;
@@ -1065,7 +1238,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    whisper_print_timings(ctx);
+    if (!params.no_prints) {
+        whisper_print_timings(ctx);
+    }
     whisper_free(ctx);
 
     return 0;

examples/python/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/server/CMakeLists.txt

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

examples/server/README.md

@@ -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

@@ -103,11 +103,11 @@ void stream_main(size_t index) {
 
             {
                 const int n_segments = whisper_full_n_segments(ctx);
-                for (int i = n_segments - 1; i < n_segments; ++i) {
-                    const char * text = whisper_full_get_segment_text(ctx, i);
+                if (n_segments > 0) {
+                    const char * text = whisper_full_get_segment_text(ctx, n_segments - 1);
 
-                    const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-                    const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+                    const int64_t t0 = whisper_full_get_segment_t0(ctx, n_segments - 1);
+                    const int64_t t1 = whisper_full_get_segment_t1(ctx, n_segments - 1);
 
                     printf("transcribed: %s\n", text);
 
@@ -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
 ```
 
@@ -30,17 +30,21 @@ a transcription block that is suitable for parsing.
 The `stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
-# Install SDL2 on Linux
+# Install SDL2
+# On Debian based linux distributions:
 sudo apt-get install libsdl2-dev
 
+# On Fedora Linux:
+sudo dnf install SDL2 SDL2-devel
+
 # Install SDL2 on Mac OS
 brew install sdl2
 
 make stream
 ```
 
-Ensure you are at the root of the repo when running `make stream`.  Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`.  Attempting to compile within
+Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
+as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
 `examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
 
 ```bash

examples/stream/stream.cpp

@@ -14,20 +14,6 @@
 #include <fstream>
 
 
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t) {
-    int64_t sec = t/100;
-    int64_t msec = t - sec*100;
-    int64_t min = sec/60;
-    sec = sec - min*60;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
-
-    return std::string(buf);
-}
-
 // command-line parameters
 struct whisper_params {
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -48,11 +34,13 @@ struct whisper_params {
     bool no_context    = true;
     bool no_timestamps = false;
     bool tinydiarize   = false;
+    bool save_audio    = false; // save audio to wav file
+    bool use_gpu       = true;
+    bool flash_attn    = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
     std::string fname_out;
-    bool save_audio = false; // save audio to wav file
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -65,25 +53,27 @@ 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 == "-tdrz" || arg == "--tinydiarize")  { params.tinydiarize   = true; }
-        else if (arg == "-sa"  || arg == "--save-audio")    { params.save_audio    = true; }
+        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 if (arg == "-fa"   || arg == "--flash-attn")    { params.flash_attn    = true; }
 
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -118,8 +108,10 @@ 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, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
     fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention during inference\n",               params.flash_attn ? "true" : "false");
     fprintf(stderr, "\n");
 }
 
@@ -163,7 +155,12 @@ int main(int argc, char ** argv) {
         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;
+    cparams.flash_attn = params.flash_attn;
+
+    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;
@@ -366,7 +363,7 @@ int main(int argc, char ** argv) {
                         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-                        std::string output = "[" + to_timestamp(t0) + " --> " + to_timestamp(t1) + "]  " + text;
+                        std::string output = "[" + to_timestamp(t0, false) + " --> " + to_timestamp(t1, false) + "]  " + text;
 
                         if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
                             output += " [SPEAKER_TURN]";
@@ -424,4 +421,4 @@ int main(int argc, char ** argv) {
     whisper_free(ctx);
 
     return 0;
-}
+}

examples/sycl/CMakeLists.txt

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

examples/sycl/README.md

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

examples/sycl/build.sh

@@ -0,0 +1,19 @@
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main only
+#cmake --build . --config Release --target main
+
+#build all binary
+cmake --build . --config Release -v

examples/sycl/ls-sycl-device.cpp

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

examples/sycl/run-whisper.sh

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

examples/talk-llama/.gitignore

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

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 unicode.cpp unicode-data.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(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../ggml-alloc.c ../../whisper.cpp)
+    if (WHISPER_CLBLAST)
+        set(CLBLAST_LIBNAME clblast)
+    endif ()
+    target_link_libraries(${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

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

examples/talk-llama/eleven-labs.py

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

examples/talk-llama/speak

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

examples/talk-llama/speak.bat

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

examples/talk-llama/speak.ps1

@@ -1,12 +1,14 @@
 # Set-ExecutionPolicy -ExecutionPolicy Bypass -Scope CurrentUser
 param(
-  # voice options are David or Zira
-  [Parameter(Mandatory=$true)][string]$voice,
-  [Parameter(Mandatory=$true)][string]$text
+  [Parameter(Mandatory=$true)][int]$voicenum,
+  [Parameter(Mandatory=$true)][string]$textfile
 )
 
 Add-Type -AssemblyName System.Speech;
 $speak = New-Object System.Speech.Synthesis.SpeechSynthesizer;
-$speak.SelectVoice("Microsoft $voice Desktop");
+$voiceoptions = $speak.GetInstalledVoices("en-US");
+$voice = $voiceoptions[$voicenum % $voiceoptions.count];
+$speak.SelectVoice($voice.VoiceInfo.Name);
 $speak.Rate="0";
+$text = Get-Content -Path $textfile;
 $speak.Speak($text);

examples/talk-llama/talk-llama.cpp

@@ -14,23 +14,31 @@
 #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(ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(ctx, token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -46,6 +54,7 @@ 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;
@@ -56,12 +65,18 @@ struct whisper_params {
     bool print_energy   = false;
     bool no_timestamps  = true;
     bool verbose_prompt = false;
+    bool use_gpu        = true;
+    bool flash_attn     = false;
 
     std::string person      = "Georgi";
+    std::string bot_name    = "LLaMA";
+    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";
     std::string speak       = "./examples/talk-llama/speak";
+    std::string speak_file  = "./examples/talk-llama/to_speak.txt";
     std::string prompt      = "";
     std::string fname_out;
     std::string path_session = "";       // path to file for saving/loading model eval state
@@ -77,25 +92,32 @@ 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 == "-fa"  || arg == "--flash-attn")     { params.flash_attn     = true; }
+        else if (arg == "-p"   || arg == "--person")         { params.person         = argv[++i]; }
+        else if (arg == "-bn"   || arg == "--bot-name")      { params.bot_name       = argv[++i]; }
+        else if (arg == "--session")                         { params.path_session   = argv[++i]; }
+        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 == "-sf"  || arg == "--speak-file")     { params.speak_file     = 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') {
@@ -118,27 +140,34 @@ 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, "  -fa,      --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME    [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
+    fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
+    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, "  -sf FILE, --speak-file     [%-7s] file to pass to TTS\n",                         params.speak_file.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");
 }
 
@@ -210,6 +239,18 @@ 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}.
@@ -237,7 +278,7 @@ 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);
@@ -245,22 +286,43 @@ 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;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
+    if (!ctx_wsp) {
+        fprintf(stderr, "No whisper.cpp model specified. Please provide using -mw <modelfile>\n");
+        return 1;
+    }
 
     // llama init
 
-    llama_backend_init(true);
+    llama_backend_init();
 
-    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);
+    if (!model_llama) {
+        fprintf(stderr, "No llama.cpp model specified. Please provide using -ml <modelfile>\n");
+        return 1;
+    }
+
+    llama_context_params lcparams = llama_context_default_params();
 
     // tune these to your liking
-    lparams.n_ctx      = 2048;
-    lparams.seed       = 1;
-    lparams.f16_kv     = true;
+    lcparams.n_ctx      = 2048;
+    lcparams.seed       = 1;
+    lcparams.n_threads  = params.n_threads;
+    lcparams.flash_attn = params.flash_attn;
 
-    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lparams);
-
-    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lparams);
+    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
 
     // print some info about the processing
     {
@@ -299,12 +361,11 @@ int main(int argc, char ** argv) {
     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;
@@ -313,7 +374,7 @@ 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
@@ -343,6 +404,8 @@ int main(int argc, char ** argv) {
 
     prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
 
+    llama_batch batch = llama_batch_init(llama_n_ctx(ctx_llama), 0, 1);
+
     // init session
     std::string path_session = params.path_session;
     std::vector<llama_token> session_tokens;
@@ -356,7 +419,7 @@ 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());
@@ -378,8 +441,21 @@ 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)) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
+    // prepare batch
+    {
+        batch.n_tokens = embd_inp.size();
+
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.token[i]     = embd_inp[i];
+            batch.pos[i]       = i;
+            batch.n_seq_id[i]  = 1;
+            batch.seq_id[i][0] = 0;
+            batch.logits[i]    = i == batch.n_tokens - 1;
+        }
+    }
+
+    if (llama_decode(ctx_llama, batch)) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
         return 1;
     }
 
@@ -415,6 +491,16 @@ 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);
@@ -460,10 +546,38 @@ 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()) {
+                    speak_with_file(params.speak, params.heard_ok, params.speak_file, voice_id);
                 }
 
                 // remove text between brackets using regex
@@ -500,7 +614,7 @@ 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);
 
@@ -561,8 +675,21 @@ 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)) {
-                            fprintf(stderr, "%s : failed to eval\n", __func__);
+                        // prepare batch
+                        {
+                            batch.n_tokens = embd.size();
+
+                            for (int i = 0; i < batch.n_tokens; i++) {
+                                batch.token[i]     = embd[i];
+                                batch.pos[i]       = n_past + i;
+                                batch.n_seq_id[i]  = 1;
+                                batch.seq_id[i][0] = 0;
+                                batch.logits[i]    = i == batch.n_tokens - 1;
+                            }
+                        }
+
+                        if (llama_decode(ctx_llama, batch)) {
+                            fprintf(stderr, "%s : failed to decode\n", __func__);
                             return 1;
                         }
                     }
@@ -593,9 +720,9 @@ 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(ctx_llama)] = 0;
+                            logits[llama_token_eos(model_llama)] = 0;
 
                             std::vector<llama_token_data> candidates;
                             candidates.reserve(n_vocab);
@@ -606,13 +733,13 @@ 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(ctx_llama)];
+                            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(ctx_llama)] = nl_logit;
+                            logits[llama_token_nl(model_llama)] = nl_logit;
 
                             if (temp <= 0) {
                                 // Greedy sampling
@@ -621,18 +748,19 @@ 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(ctx_llama)) {
+                        if (id != llama_token_eos(model_llama)) {
                             // add it to the context
                             embd.push_back(id);
 
                             text_to_speak += llama_token_to_piece(ctx_llama, id);
 
                             printf("%s", llama_token_to_piece(ctx_llama, id).c_str());
+                            fflush(stdout);
                         }
                     }
 
@@ -661,11 +789,7 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                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__);
-                }
+                speak_with_file(params.speak, text_to_speak, params.speak_file, voice_id);
 
                 audio.clear();
             }

examples/talk-llama/unicode-data.h

@@ -0,0 +1,17 @@
+#pragma once
+
+#include <cstdint>
+#include <map>
+#include <utility>
+#include <vector>
+
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
+extern const std::multimap<uint32_t, uint32_t>          unicode_map_nfd;
+extern const std::map<char32_t, char32_t>               unicode_map_lowercase;

examples/talk-llama/unicode.cpp

@@ -0,0 +1,818 @@
+#include "unicode.h"
+#include "unicode-data.h"
+
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <map>
+#include <regex>
+#include <stdexcept>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+#include <locale>
+#include <codecvt>
+
+static std::string unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(unicode_cpt_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+static uint32_t unicode_cpt_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;
+    }
+    if (!(utf8[offset + 0] & 0x40)) {
+        throw std::invalid_argument("invalid character");
+    }
+    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;
+    }
+    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;
+    }
+    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("failed to convert utf8 to codepoint");
+}
+
+//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cp) {
+//    std::vector<uint16_t> result;
+//    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
+//        result.emplace_back(cp);
+//        return result;
+//    }
+//    if (0x10000 <= cp && cp <= 0x10ffff) {
+//        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
+//        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+//        return result;
+//    }
+//    throw std::invalid_argument("failed to convert codepoint to utf16");
+//}
+
+//static std::vector<uint16_t> unicode_cpts_to_utf16(const std::vector<uint32_t> & cps) {
+//    std::vector<uint16_t> result;
+//    for (size_t i = 0; i < cps.size(); ++i) {
+//        auto temp = unicode_cpt_to_utf16(cps[i]);
+//        result.insert(result.end(), temp.begin(), temp.end());
+//    }
+//    return result;
+//}
+
+//static uint32_t unicode_cpt_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;
+//    }
+//
+//    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;
+//}
+
+//static std::vector<uint32_t> unicode_cpts_from_utf16(const std::vector<uint16_t> & utf16) {
+//    std::vector<uint32_t> result;
+//    size_t offset = 0;
+//    while (offset < utf16.size()) {
+//        result.push_back(unicode_cpt_from_utf16(utf16, offset));
+//    }
+//    return result;
+//}
+
+static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
+    std::unordered_map<uint32_t, int> cpt_types;
+    for (auto p : unicode_ranges_number) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_NUMBER;
+        }
+    }
+    for (auto p : unicode_ranges_letter) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_LETTER;
+        }
+    }
+    for (auto p : unicode_ranges_separator) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_SEPARATOR;
+        }
+    }
+    for (auto p : unicode_ranges_accent_mark) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
+        }
+    }
+    for (auto p : unicode_ranges_punctuation) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_PUNCTUATION;
+        }
+    }
+    for  (auto p : unicode_ranges_symbol) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_SYMBOL;
+        }
+    }
+    for (auto p : unicode_ranges_control) {
+        for (auto i = p.first; i <= p.second; ++i) {
+            cpt_types[i] = CODEPOINT_TYPE_CONTROL;
+        }
+    }
+    return cpt_types;
+}
+
+static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
+    std::unordered_map<uint8_t, std::string> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(ch) == map.end()) {
+            map[ch] = unicode_cpt_to_utf8(256 + n);
+            ++n;
+        }
+    }
+    return map;
+}
+
+static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
+    std::unordered_map<std::string, uint8_t> map;
+    for (int ch = u'!'; ch <= u'~'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(unicode_cpt_to_utf8(ch)) == map.end()) {
+            map[unicode_cpt_to_utf8(256 + n)] = ch;
+            ++n;
+        }
+    }
+    return map;
+}
+
+static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
+    return conv.from_bytes(s);
+}
+
+static std::vector<std::string> unicode_byte_encoding_process(const std::vector<std::string> & bpe_words) {
+    std::vector<std::string> bpe_encoded_words;
+    for (const auto & word : bpe_words) {
+        std::string text_utf;
+        auto utf_word =  unicode_cpts_from_utf8(word);
+        for (size_t i = 0; i < utf_word.size(); ++i) {
+            text_utf += unicode_cpt_to_utf8(utf_word[i]);
+        }
+
+        std::string encoded_token;
+        for (char & c : text_utf) {
+            encoded_token += unicode_byte_to_utf8(c);
+        }
+        bpe_encoded_words.emplace_back(encoded_token);
+    }
+    return bpe_encoded_words;
+}
+
+// GPT2 system regex:  's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
+static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        auto _get_cpt = [&] (const size_t pos) -> char32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        };
+
+        auto _get_cpt_type = [&] (const size_t pos) -> int {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const char32_t cpt = _get_cpt(pos);
+            const int cpt_type = _get_cpt_type(pos);
+
+            // regex: 's|'t|'re|'ve|'m|'ll|'d
+            if (cpt == '\'' && pos+1 < offset_end) {
+                char32_t cpt_next = _get_cpt(pos+1);
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                    pos += _add_token(pos+2);
+                    continue;
+                }
+                if (pos+2 < offset_end) {
+                    char32_t cpt_next_next = _get_cpt(pos+2);
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
+                        pos += _add_token(pos+3);
+                        continue;
+                    }
+                }
+            }
+
+            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
+            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            // regex: <space>?\p{L}+
+            if (cpt2_type == CODEPOINT_TYPE_LETTER) {
+                pos += (cpt == ' ');
+                while (cpt2_type == CODEPOINT_TYPE_LETTER) {
+                    cpt2_type = _get_cpt_type(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: <space>?\p{N}+
+            if (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+                pos += (cpt == ' ');
+                while (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+                    cpt2_type = _get_cpt_type(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: <space>?[^\s\p{L}\p{N}]+
+            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                pos += (cpt == ' ');
+                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                    cpt2_type = _get_cpt_type(++pos);
+                    cpt2 = _get_cpt(pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            size_t num_whitespaces = 0;
+            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+                num_whitespaces++;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
+// LLAMA3 system regex: "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
+static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        auto _get_cpt = [&] (const size_t pos) -> char32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        };
+
+        auto _get_cpt_type = [&] (const size_t pos) -> int {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const char32_t cpt = _get_cpt(pos);
+            const int cpt_type = _get_cpt_type(pos);
+
+            // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
+            if (cpt == '\'' && pos+1 < offset_end) {
+                char32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                    pos += _add_token(pos+2);
+                    continue;
+                }
+                if (pos+2 < offset_end) {
+                    char32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
+                        pos += _add_token(pos+3);
+                        continue;
+                    }
+                }
+            }
+
+            // regex: [^\r\n\p{L}\p{N}]?\p{L}+  //####FIXME: the first \p{L} is correct?
+            if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) {
+                if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) {  // one or more letters
+                    pos++;
+                    while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) {
+                        pos++;
+                    }
+                    _add_token(pos);
+                    continue;
+                }
+            }
+
+            // regex: \p{N}{1,3}
+            if (cpt_type == CODEPOINT_TYPE_NUMBER) {
+                size_t ini = pos;
+                while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) {
+                    if (++pos - ini >= 3 ) {
+                        _add_token(pos);
+                        ini = pos;
+                    }
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
+            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
+            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                pos += (cpt == ' ');
+                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                    cpt2_type = _get_cpt_type(++pos);
+                    cpt2 = _get_cpt(pos);
+                }
+                while (cpt2 == '\r' || cpt2 == '\n') {
+                    cpt2 = _get_cpt(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            size_t num_whitespaces = 0;
+            size_t last_end_r_or_n = 0;
+            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+                char32_t cpt2 = _get_cpt(pos+num_whitespaces);
+                if (cpt2 == '\r' || cpt2 == '\n') {
+                    last_end_r_or_n = pos + num_whitespaces + 1;
+                }
+                num_whitespaces++;
+            }
+
+            // regex: \s*[\r\n]+
+            if (last_end_r_or_n > 0) {
+                pos = last_end_r_or_n;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
+// use std::wregex to split the text
+static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
+    std::wregex expr(regex_expr);
+    std::vector<size_t> bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+    size_t start = 0;
+    for (auto offset : offsets) {
+        std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);
+        std::wcregex_iterator end;
+
+        int64_t start_idx = 0;
+        while (it != end) {
+            std::wcmatch match = *it;
+            if (match.position() > start_idx) {
+                bpe_offsets.emplace_back(match.position() - start_idx);
+            }
+            bpe_offsets.emplace_back(match.length());
+            start_idx = match.position() + match.length();
+            ++it;
+        }
+
+        if (start_idx < (int64_t) offset) {
+            bpe_offsets.emplace_back(offset - start_idx);
+        }
+        start += offset;
+    }
+
+    return bpe_offsets;
+}
+
+// use std::regex to split the text
+static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
+    std::regex expr(regex_expr);
+    std::vector<size_t> bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+    size_t start = 0;
+    for (auto offset : offsets) {
+        std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);
+        std::cregex_iterator end;
+
+        int64_t start_idx = 0;
+        while (it != end) {
+            std::cmatch match = *it;
+            if (match.position() > start_idx) {
+                bpe_offsets.emplace_back(match.position() - start_idx);
+            }
+            bpe_offsets.emplace_back(match.length());
+            start_idx = match.position() + match.length();
+            ++it;
+        }
+
+        if (start_idx < (int64_t) offset) {
+            bpe_offsets.emplace_back(offset - start_idx);
+        }
+        start += offset;
+    }
+
+    return bpe_offsets;
+}
+
+static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets;
+
+    if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
+        bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
+    } else if (
+            regex_expr == "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" ||
+            regex_expr == "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+") {
+
+        bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
+    }
+
+    return bpe_offsets;
+}
+
+//
+// interface
+//
+
+std::string unicode_cpt_to_utf8(uint32_t cp) {
+    std::string result;
+
+    if (/* 0x00 <= cp && */ cp <= 0x7f) {
+        result.push_back(cp);
+        return result;
+    }
+    if (0x80 <= cp && cp <= 0x7ff) {
+        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+    if (0x800 <= cp && cp <= 0xffff) {
+        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+    if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.push_back(0xf0 | ((cp >> 18) & 0x07));
+        result.push_back(0x80 | ((cp >> 12) & 0x3f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+
+    throw std::invalid_argument("invalid codepoint");
+}
+
+std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
+    std::vector<uint32_t> result;
+    result.reserve(cpts.size());
+    for (size_t i = 0; i < cpts.size(); ++i) {
+        auto it = unicode_map_nfd.find(cpts[i]);
+        if (it == unicode_map_nfd.end()) {
+            result.push_back(cpts[i]);
+        } else {
+            result.push_back(it->second);
+        }
+    }
+    return result;
+}
+
+std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
+    std::vector<uint32_t> result;
+    size_t offset = 0;
+    while (offset < utf8.size()) {
+        result.push_back(unicode_cpt_from_utf8(utf8, offset));
+    }
+    return result;
+}
+
+int unicode_cpt_type(uint32_t cp) {
+    static std::unordered_map<uint32_t, int> cpt_types = unicode_cpt_type_map();
+    const auto it = cpt_types.find(cp);
+    return it == cpt_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
+}
+
+int unicode_cpt_type(const std::string & utf8) {
+    if (utf8.length() == 0) {
+        return CODEPOINT_TYPE_UNIDENTIFIED;
+    }
+    size_t offset = 0;
+    return unicode_cpt_type(unicode_cpt_from_utf8(utf8, offset));
+}
+
+bool unicode_cpt_is_whitespace(uint32_t cp) {
+    static const std::unordered_set<uint32_t> is_whitespace = [] {
+        std::unordered_set<uint32_t> is_whitespace;
+        for (auto p : unicode_ranges_whitespace) {
+            for (auto i = p.first; i <= p.second; ++i) {
+                is_whitespace.insert(i);
+            }
+        }
+        return is_whitespace;
+    }();
+    return (bool)is_whitespace.count(cp);
+}
+
+std::string unicode_byte_to_utf8(uint8_t byte) {
+    static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();
+    return map.at(byte);
+}
+
+uint8_t unicode_utf8_to_byte(const std::string & utf8) {
+    static std::unordered_map<std::string, uint8_t> map = unicode_utf8_to_byte_map();
+    return map.at(utf8);
+}
+
+char32_t unicode_tolower(char32_t cp) {
+    auto it = unicode_map_lowercase.find(cp);
+    return it == unicode_map_lowercase.end() ? cp : it->second;
+}
+
+std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
+    // unicode categories
+    static const std::map<std::string, int> k_ucat_enum = {
+        { "\\p{N}", CODEPOINT_TYPE_NUMBER },
+        { "\\p{L}", CODEPOINT_TYPE_LETTER },
+        { "\\p{P}", CODEPOINT_TYPE_PUNCTUATION },
+    };
+
+    static const std::map<int, int> k_ucat_cpt = {
+        { CODEPOINT_TYPE_NUMBER,        0xD1 },
+        { CODEPOINT_TYPE_LETTER,        0xD2 },
+        { CODEPOINT_TYPE_PUNCTUATION,   0xD3 },
+    };
+
+    static const std::map<int, std::string> k_ucat_map = {
+        { CODEPOINT_TYPE_NUMBER,        "\x30-\x39" }, // 0-9
+        { CODEPOINT_TYPE_LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { CODEPOINT_TYPE_PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+    };
+
+    // compute collapsed codepoints only if needed by at least one regex
+    bool need_collapse = false;
+    for (auto & regex_expr : regex_exprs) {
+        // search for unicode categories
+        for (const auto & ucat : k_ucat_enum) {
+            if (std::string::npos != regex_expr.find(ucat.first)) {
+                need_collapse = true;
+                break;
+            }
+        }
+    }
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
+    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    std::string text_collapsed;
+    if (need_collapse) {
+        // collapse all unicode categories
+        text_collapsed.resize(cpts.size());
+
+        for (size_t i = 0; i < cpts.size(); ++i) {
+            // keep single-byte codepoints as is
+            if (cpts[i] < 128) {
+                text_collapsed[i] = cpts[i];
+                continue;
+            }
+
+            const int cpt_type = unicode_cpt_type(cpts[i]);
+
+            if (k_ucat_cpt.find(cpt_type) != k_ucat_cpt.end()) {
+                text_collapsed[i] = k_ucat_cpt.at(cpt_type);
+            } else {
+                text_collapsed[i] = (char) 0xD0; // fallback
+            }
+        }
+    }
+
+    std::vector<size_t> bpe_offsets = { cpts.size() };
+
+    for (auto & regex_expr : regex_exprs) {
+        // first, see if we have an efficient custom regex implementation
+        auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
+
+        if (!tmp.empty()) {
+            bpe_offsets = std::move(tmp);
+            continue;
+        }
+
+        // fallback to general-purpose std::regex / std::wregex
+        try {
+            // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
+            // with the corresponding collapsed representation
+            bool use_collapsed = false;
+            for (auto & ucat : k_ucat_enum) {
+                if (std::string::npos != regex_expr.find(ucat.first)) {
+                    use_collapsed = true;
+                    break;
+                }
+            }
+
+            if (use_collapsed) {
+                // sanity-check that the original regex does not contain any non-ASCII characters
+                const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
+                for (size_t i = 0; i < cpts_regex.size(); ++i) {
+                    if (cpts_regex[i] >= 128) {
+                        throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");
+                    }
+                }
+
+                // generate a collapsed representation of the regex
+                std::string regex_expr_collapsed;
+
+                // track if we are inside [], because nested [] are not allowed
+                bool inside = false;
+                for (size_t i = 0; i < regex_expr.size(); ++i) {
+                    if (regex_expr[i] == '[' && (i == 0 || regex_expr[i - 1] != '\\')) {
+                        regex_expr_collapsed += '[';
+                        inside = true;
+                        continue;
+                    }
+
+                    if (inside && regex_expr[i] == ']' && regex_expr[i - 1] != '\\') {
+                        regex_expr_collapsed += ']';
+                        inside = false;
+                        continue;
+                    }
+
+                    if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
+                        regex_expr[i + 1] == 'p' &&
+                        regex_expr[i + 2] == '{' &&
+                        regex_expr[i + 4] == '}') {
+                        const std::string pat = regex_expr.substr(i, 5);
+                        if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
+                            if (!inside) {
+                                regex_expr_collapsed += '[';
+                            }
+                            regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
+                            regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
+                            if (!inside) {
+                                regex_expr_collapsed += ']';
+                            }
+                            i += 4;
+                            continue;
+                        }
+                    }
+
+                    regex_expr_collapsed += regex_expr[i];
+                }
+
+                //printf("text_collapsed: %s\n", text_collapsed.c_str());
+                //printf("regex_expr_collapsed: %s\n", regex_expr_collapsed.c_str());
+                bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
+            } else {
+                // no unicode category used, we can use std::wregex directly
+                const std::wstring wtext       = unicode_wstring_from_utf8(text);
+                const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
+
+                //printf("text: %s\n", text.c_str());
+                //printf("regex_expr: %s\n", regex_expr.c_str());
+                bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);
+            }
+        } catch (std::regex_error & e) {
+            fprintf(stderr, "Failed to process regex: '%s'\n", regex_expr.c_str());
+            fprintf(stderr, "Regex error: %s\n", e.what());
+            throw std::runtime_error("Failed to process regex");
+        }
+    }
+
+    std::vector<std::string> bpe_words;
+    bpe_words.reserve(bpe_offsets.size()); // reserve memory for the approximate size
+
+    size_t start = 0;
+    for (size_t & offset : bpe_offsets) {
+        bpe_words.emplace_back();
+        for (size_t i = start; i < start + offset; ++i) {
+            bpe_words.back() += unicode_cpt_to_utf8(cpts[i]);
+        }
+        start += offset;
+    }
+
+    return unicode_byte_encoding_process(bpe_words);
+}

examples/talk-llama/unicode.h

@@ -0,0 +1,31 @@
+#pragma once
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#define CODEPOINT_TYPE_UNIDENTIFIED 0
+#define CODEPOINT_TYPE_NUMBER       1
+#define CODEPOINT_TYPE_LETTER       2
+#define CODEPOINT_TYPE_SEPARATOR    3
+#define CODEPOINT_TYPE_ACCENT_MARK  4
+#define CODEPOINT_TYPE_PUNCTUATION  5
+#define CODEPOINT_TYPE_SYMBOL       6
+#define CODEPOINT_TYPE_CONTROL      7
+
+std::string unicode_cpt_to_utf8(uint32_t cp);
+std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
+
+std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
+
+int unicode_cpt_type(uint32_t cp);
+int unicode_cpt_type(const std::string & utf8);
+
+bool unicode_cpt_is_whitespace(uint32_t cp);
+
+std::string unicode_byte_to_utf8(uint8_t byte);
+uint8_t unicode_utf8_to_byte(const std::string & utf8);
+
+char32_t unicode_tolower(char32_t cp);
+
+std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

examples/talk.wasm/emscripten.cpp

@@ -29,18 +29,6 @@ std::string g_status_forced = "";
 
 std::vector<float> g_pcmf32;
 
-std::string to_timestamp(int64_t t) {
-    int64_t sec = t/100;
-    int64_t msec = t - sec*100;
-    int64_t min = sec/60;
-    sec = sec - min*60;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
-
-    return std::string(buf);
-}
-
 void talk_set_status(const std::string & status) {
     std::lock_guard<std::mutex> lock(g_mutex);
     g_status = status;
@@ -271,7 +259,7 @@ 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
 
@@ -524,8 +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]

examples/talk/.gitignore

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