whisper : add whisper_token_count helper

* README.md

fix documentaion and added fedora liunx dependencies for stream build

* fix documentaion and added fedora liunx dependencies for command build

* fix documentaion and added fedora liunx dependencies for talk build

* fix documentaion and added fedora liunx dependencies for talk-llama build

* reverted back mistakenly removed MacOS documentaion

.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

@@ -117,7 +117,6 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y clang
             apt install -y clang build-essential cmake libsdl2-dev
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
             make
@@ -151,6 +150,106 @@ jobs:
             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:
     runs-on: windows-latest
 
@@ -167,7 +266,7 @@ jobs:
             s2arc: x64
             jnaPath: win32-x86-64
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
@@ -224,11 +323,14 @@ jobs:
           - arch: Win32
             obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
+            clblast: OFF
           - arch: x64
             obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
+            clblast: ON
+            clver: 1.6.1
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
@@ -253,6 +355,18 @@ 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 ${{ matrix.arch }}
@@ -260,6 +374,7 @@ jobs:
           -DWHISPER_OPENBLAS=${{ matrix.blas }}
           -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
@@ -274,11 +389,15 @@ jobs:
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
+      - name: Copy clblast.dll
+        if: matrix.clblast == 'ON'
+        run: copy "$env:CLBlast_DIR/../../clblast.dll" build/bin/${{ matrix.build }}
+
       - name: Upload binaries
         if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@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:
@@ -295,7 +414,7 @@ jobs:
           - arch: x64
             s2arc: x64
           - sdl2: ON
-            s2ver: 2.26.0
+            s2ver: 2.28.5
 
     steps:
       - name: Clone
@@ -321,7 +440,8 @@ 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 ${{ matrix.cuda-toolkit }}
         run: |
@@ -396,6 +516,14 @@ 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
@@ -408,9 +536,15 @@ 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
 
@@ -434,7 +568,7 @@ jobs:
       - name: Build
         run: |
           cd examples/whisper.android.java
-          chmod +x ./gradlew 
+          chmod +x ./gradlew
           ./gradlew assembleRelease
 
   java:

.github/workflows/docker.yml

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

.gitignore

@@ -6,6 +6,7 @@
 .vs/
 .vscode/
 .DS_Store
+.vimspector.json
 
 build/
 build-coreml/
@@ -58,4 +59,4 @@ benchmark_results.csv
 cmake-build-debug/
 .cxx/
 .gradle/
-local.properties
+local.properties

CMakeLists.txt

@@ -1,6 +1,7 @@
 cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.5.2)
+project(whisper.cpp VERSION 1.5.4)
+set(SOVERSION 1)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -68,13 +69,16 @@ 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_CUBLAS                "whisper: support for cuBLAS"                        OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               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 +109,13 @@ 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 +126,7 @@ 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 +156,33 @@ 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_SOURCE_DIR}/ggml-metal.metal")
+
+            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+            add_custom_command(
+                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
+                DEPENDS ${METALLIB_SOURCE}
+                COMMENT "Generate assembly for embedded Metal library"
+            )
+
+            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
+        endif()
     endif()
 
     if (WHISPER_COREML)
@@ -218,11 +254,17 @@ if (WHISPER_CUBLAS)
         add_compile_definitions(GGML_USE_CUBLAS)
 
         if (WHISPER_STATIC)
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
         else()
             set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
     else()
         message(FATAL_ERROR "cuBLAS not found")
     endif()
@@ -278,6 +320,30 @@ 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 +375,8 @@ 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()
 
@@ -338,8 +405,8 @@ else()
         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 "${CMAKE_C_FLAGS} -mavx")
@@ -471,10 +538,18 @@ add_library(${TARGET}
     ${GGML_SOURCES_METAL}
     ${GGML_SOURCES_CUDA}
     ${GGML_SOURCES_OPENCL}
+    ${GGML_SOURCES_SYCL}
+    ${GGML_HEADERS_SYCL}
     whisper.h
     whisper.cpp
     )
 
+# Set the version numbers
+set_target_properties(whisper PROPERTIES
+    VERSION ${PROJECT_VERSION}
+    SOVERSION ${SOVERSION}
+)
+
 include(DefaultTargetOptions)
 
 target_include_directories(${TARGET} PUBLIC
@@ -498,6 +573,7 @@ else()
 endif()
 
 if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_link_libraries(${TARGET} PUBLIC
         ${CMAKE_DL_LIBS}
         )
@@ -521,7 +597,13 @@ 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()
 

Makefile

@@ -42,6 +42,12 @@ 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=$(MACOSX_DEPLOYMENT_TARGET)
+endif
+
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
@@ -99,6 +105,16 @@ ifeq ($(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 ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -107,7 +123,7 @@ ifeq ($(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
@@ -169,6 +185,8 @@ 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
@@ -199,14 +217,14 @@ endif
 
 ifdef WHISPER_CUBLAS
 	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
+	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
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
@@ -329,6 +347,24 @@ 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
+	@echo "Embedding Metal library"
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
+
+WHISPER_OBJ += ggml-metal-embed.o
+endif
 endif
 
 libwhisper.a: $(WHISPER_OBJ)
@@ -360,7 +396,7 @@ quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
 	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
 server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS) $(LWINSOCK2)
 
 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)
@@ -374,8 +410,8 @@ lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
 
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.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 $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples

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/)
 
-Stable: [v1.5.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -33,9 +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 high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
-The rest of the code is part of the [ggml](https://github.com/ggerganov/ggml) machine learning library.
+The 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)
@@ -60,22 +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
@@ -90,7 +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
@@ -206,7 +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
 ```
 
@@ -238,9 +239,9 @@ make large-v3
 
 ## Memory usage
 
-| Model  | Disk    | Mem      |
-| ---    | ---     | ---      |
-| tiny   |  75 MiB | ~273 MB |
+| 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 |
@@ -277,7 +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`
 
@@ -303,8 +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
 
   ...
 
@@ -332,21 +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:
@@ -355,7 +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:
@@ -365,24 +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
 
   ...
 
@@ -433,7 +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
@@ -448,6 +455,38 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```
 
+## Docker
+
+### Prerequisites
+
+- Docker must be installed and running on your system.
+- Create a folder to store big models & intermediate files (ex. /whisper/models)
+
+### Images
+
+We have two Docker images available for this project:
+
+1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
+
+### Usage
+
+```shell
+# download model and persist it in a local folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./models/download-ggml-model.sh base /models"
+# transcribe an audio file
+docker run -it --rm \
+  -v path/to/models:/models \
+  -v path/to/audios:/audios \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
+# transcribe an audio file in samples folder
+docker run -it --rm \
+  -v path/to/models:/models \
+  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
+```
+
 ## Limitations
 
 - Inference only
@@ -460,7 +499,7 @@ 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'
@@ -532,6 +571,7 @@ 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
@@ -540,7 +580,7 @@ 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
 ```
@@ -552,7 +592,7 @@ 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
 ```
 
@@ -562,8 +602,8 @@ 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'
 ...
@@ -586,8 +626,8 @@ 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'
 ...
@@ -657,7 +697,7 @@ 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
@@ -667,7 +707,7 @@ 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
@@ -677,7 +717,7 @@ 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
@@ -691,7 +731,7 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
 
-```java
+```bash
 ./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```
@@ -720,8 +760,7 @@ It is written in python with the intention of being easy to modify and extend fo
 
 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:
 
@@ -736,51 +775,50 @@ or manually from here:
 - https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
-For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
-in [models](models).
+For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or [models/README.md](models/README.md).
 
 ## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
-- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
-- [X] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [x] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
+- [x] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
   - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
-- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
-- [X] Java:
+- [x] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [x] Java:
   - [GiviMAD/whisper-jni](https://github.com/GiviMAD/whisper-jni)
-- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
-- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+- [x] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
+- [x] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
   - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
-- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+- [x] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
   - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
-- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
   - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
   - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
-- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
-- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
+- [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
 There are various examples of using the library for different projects in the [examples](examples) folder.
 Some of the examples are even ported to run in the browser using WebAssembly. Check them out!
 
-| Example | Web | Description |
-| ---     | --- | ---         |
-| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
-| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
-| [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
-| [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
-| [wchess](examples/wchess) | [wchess.wasm](examples/wchess) | Voice-controlled chess |
-| [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
-| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
-| [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
-| [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
-| [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
-| [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
-| [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
-| [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
-| [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
-| [server](examples/server) | | HTTP transcription server with OAI-like API |
+| Example                                             | Web                                   | Description                                                                                                                     |
+| --------------------------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
+| [main](examples/main)                               | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
+| [bench](examples/bench)                             | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
+| [stream](examples/stream)                           | [stream.wasm](examples/stream.wasm)   | Real-time transcription of raw microphone capture                                                                               |
+| [command](examples/command)                         | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic                                                         |
+| [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/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>
 

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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.5.2",
+  "version": "1.5.4",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/extconf.rb

@@ -9,6 +9,7 @@ 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

@@ -70,7 +70,7 @@ extern "C" {
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
         // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        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);

bindings/ruby/ext/ggml-backend.c

@@ -156,8 +156,8 @@ void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_
     backend->iface.graph_plan_compute(backend, plan);
 }
 
-void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    backend->iface.graph_compute(backend, cgraph);
+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) {

bindings/ruby/ext/ggml-backend.h

@@ -52,7 +52,7 @@ extern "C" {
 
     GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
     GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_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

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]);
 

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)
@@ -50,6 +54,9 @@ if (WHISPER_SDL2)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+# add json lib
+add_library(json_cpp INTERFACE json.hpp)
+
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -72,6 +79,9 @@ else()
     add_subdirectory(talk)
     add_subdirectory(talk-llama)
     add_subdirectory(lsp)
+    if (LLAMA_SYCL)
+        add_subdirectory(sycl)
+    endif()
 endif()
 
 add_subdirectory(wchess)

examples/addon.node/addon.cpp

@@ -52,27 +52,6 @@ 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;
@@ -104,8 +83,8 @@ 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;
@@ -154,7 +133,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     // whisper init
 
-    struct whisper_context_params cparams;
+    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);
 

examples/bench/bench.cpp

@@ -8,7 +8,7 @@
 // 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";
 
@@ -58,7 +58,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
 int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
-    struct whisper_context_params cparams;
+    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);

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

@@ -22,11 +22,6 @@
 #include <vector>
 #include <map>
 
-bool file_exists(const std::string & fname) {
-    std::ifstream f(fname.c_str());
-    return f.good();
-}
-
 // command-line parameters
 struct whisper_params {
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -693,7 +688,7 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context_params cparams;
+    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);
@@ -736,7 +731,7 @@ int main(int argc, char ** argv) {
 
     if (!params.grammar.empty()) {
         auto & grammar = params.grammar_parsed;
-        if (file_exists(params.grammar.c_str())) {
+        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>());

examples/common-ggml.cpp

@@ -62,6 +62,14 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:
+        case GGML_FTYPE_MOSTLY_IQ2_XS:
+        case GGML_FTYPE_MOSTLY_IQ2_S:
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:
+        case GGML_FTYPE_MOSTLY_IQ3_S:
+        case GGML_FTYPE_MOSTLY_IQ1_S:
+        case GGML_FTYPE_MOSTLY_IQ4_NL:
+        case GGML_FTYPE_MOSTLY_IQ4_XS:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -82,8 +90,6 @@ 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;
@@ -168,8 +174,6 @@ bool ggml_common_quantize_0(
             work.resize(nelements); // for quantization
 
             size_t cur_size = 0;
-            std::vector<int64_t> hist_cur(1 << 4, 0);
-
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
@@ -182,7 +186,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_Q5_K:
                 case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements, 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:
@@ -191,6 +195,14 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_I32:
                 case GGML_TYPE_Q8_1:
                 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_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
@@ -201,15 +213,7 @@ 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());
@@ -222,18 +226,5 @@ 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.cpp

@@ -615,6 +615,21 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 
 }
 
+bool is_wav_buffer(const std::string buf) {
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    if (buf.size() < 12 || buf.substr(0, 4) != "RIFF" || buf.substr(8, 4) != "WAVE") {
+        return false;
+    }
+
+    uint32_t chunk_size = *reinterpret_cast<const uint32_t*>(buf.data() + 4);
+    if (chunk_size + 8 != buf.size()) {
+        return false;
+    }
+
+    return true;
+}
+
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
     std::vector<uint8_t> wav_data; // used for pipe input from stdin
@@ -639,6 +654,12 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
 
         fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
     }
+    else if (is_wav_buffer(fname)) {
+        if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
+            return false;
+        }
+    }
     else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
         fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
         return false;
@@ -815,3 +836,48 @@ 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

@@ -135,7 +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(
@@ -277,3 +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/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

@@ -435,7 +435,7 @@ int main(int argc, char ** argv) {
     }
 
     // whisper init
-    struct whisper_context_params cparams;
+    struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
     // init audio

examples/main/main.cpp

@@ -14,34 +14,6 @@
 #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()) {
@@ -54,16 +26,17 @@ void replace_all(std::string & s, const std::string & search, const std::string
 
 // command-line parameters
 struct whisper_params {
-    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors =  1;
-    int32_t offset_t_ms  =  0;
-    int32_t offset_n     =  0;
-    int32_t duration_ms  =  0;
-    int32_t progress_step =  5;
-    int32_t max_context  = -1;
-    int32_t max_len      =  0;
-    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
-    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t 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;
@@ -85,6 +58,7 @@ 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;
@@ -102,12 +76,22 @@ struct whisper_params {
 
     std::string openvino_encode_device = "CPU";
 
+    std::string dtw = "";
+
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
 };
 
 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];
@@ -135,6 +119,7 @@ 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]); }
@@ -155,16 +140,18 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
         else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
         else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
         else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
         else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
         else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
         else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = 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 == "-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 {
@@ -193,6 +180,7 @@ 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);
@@ -212,16 +200,18 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
     fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
+    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
     fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
     fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
     fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
     fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\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, "\n");
@@ -238,8 +228,8 @@ 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;
@@ -663,7 +653,8 @@ 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);
@@ -852,6 +843,9 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
+
+void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -860,6 +854,19 @@ int main(int argc, char ** argv) {
         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);
@@ -878,11 +885,37 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
+    if (params.no_prints) {
+        whisper_log_set(cb_log_disable, NULL);
+    }
+
     // whisper init
 
-    struct whisper_context_params cparams;
+    struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
 
+    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) {
@@ -905,26 +938,25 @@ 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, %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.beam_size, params.best_of,
@@ -958,6 +990,7 @@ 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;
@@ -973,6 +1006,8 @@ int main(int argc, char ** argv) {
             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 };
 
             // this callback is called on each new segment

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

@@ -1,12 +1,10 @@
 set(TARGET server)
-add_executable(${TARGET} server.cpp httplib.h json.hpp)
+add_executable(${TARGET} server.cpp httplib.h)
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common json_cpp whisper ${CMAKE_THREAD_LIBS_INIT})
 
-# Check if the compiler is MinGW
-if(MINGW)
-    # Link the necessary libraries for SSL and Winsock
-    target_link_libraries(${TARGET} PRIVATE -lcrypt32 -lssl -lcrypto -lws2_32)
+if (WIN32)
+    target_link_libraries(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -46,7 +46,7 @@ options:
   --convert,                     [false  ] Convert audio to WAV, requires ffmpeg on the server
 ```
 
-> [!WARNING]  
+> [!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
@@ -56,8 +56,9 @@ options:
 curl 127.0.0.1:8080/inference \
 -H "Content-Type: multipart/form-data" \
 -F file="@<file-path>" \
--F temperature="0.2" \
--F response-format="json"
+-F temperature="0.0" \
+-F temperature_inc="0.2" \
+-F response_format="json"
 ```
 
 **/load**

examples/server/server.cpp

@@ -18,17 +18,10 @@
 #endif
 
 using namespace httplib;
-using json = nlohmann::json;
+using json = nlohmann::ordered_json;
 
 namespace {
 
-// 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",
-};
-
 // output formats
 const std::string json_format   = "json";
 const std::string text_format   = "text";
@@ -40,30 +33,33 @@ struct server_params
 {
     std::string hostname = "127.0.0.1";
     std::string public_path = "examples/server/public";
+    std::string request_path = "";
 
     int32_t port          = 8080;
     int32_t read_timeout  = 600;
     int32_t write_timeout = 600;
-    
+
     bool ffmpeg_converter = false;
 };
 
 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       = 2;
+    int32_t beam_size     = -1;
+    int32_t audio_ctx     = 0;
 
-    float word_thold    =  0.01f;
-    float entropy_thold =  2.40f;
-    float logprob_thold = -1.00f;
-    float userdef_temp  =  0.20f;
+    float word_thold      =  0.01f;
+    float entropy_thold   =  2.40f;
+    float logprob_thold   = -1.00f;
+    float temperature     =  0.00f;
+    float temperature_inc =  0.20f;
 
     bool speed_up        = false;
     bool debug_mode      = false;
@@ -93,35 +89,7 @@ struct whisper_params {
     std::string openvino_encode_device = "CPU";
 };
 
-//  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)));
-}
-
-bool is_file_exist(const char *fileName)
-{
-    std::ifstream infile(fileName);
-    return infile.good();
-}
-
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params,
-                         const server_params& sparams) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] \n", argv[0]);
     fprintf(stderr, "\n");
@@ -137,6 +105,7 @@ 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);
@@ -160,6 +129,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
     fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
     fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
+    fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
     fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
     fprintf(stderr, "\n");
 }
@@ -181,6 +151,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         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]); }
@@ -207,6 +178,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
         else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
         else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
+        else if (                  arg == "--request-path")    { sparams.request_path = argv[++i]; }
         else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -268,8 +240,8 @@ 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;
@@ -393,36 +365,106 @@ std::string output_str(struct whisper_context * ctx, const whisper_params & para
     return result.str();
 }
 
+bool parse_str_to_bool(const std::string & s) {
+    if (s == "true" || s == "1" || s == "yes" || s == "y") {
+        return true;
+    }
+    return false;
+}
+
 void get_req_parameters(const Request & req, whisper_params & params)
 {
-    // user model configu.has_fileion
-    if (req.has_file("offset-t"))
+    if (req.has_file("offset_t"))
     {
-        params.offset_t_ms = std::stoi(req.get_file_value("offset-t").content);
+        params.offset_t_ms = std::stoi(req.get_file_value("offset_t").content);
     }
-    if (req.has_file("offset-n"))
+    if (req.has_file("offset_n"))
     {
-        params.offset_n = std::stoi(req.get_file_value("offset-n").content);
+        params.offset_n = std::stoi(req.get_file_value("offset_n").content);
     }
     if (req.has_file("duration"))
     {
         params.duration_ms = std::stoi(req.get_file_value("duration").content);
     }
-    if (req.has_file("max-context"))
+    if (req.has_file("max_context"))
     {
-        params.max_context = std::stoi(req.get_file_value("max-context").content);
+        params.max_context = std::stoi(req.get_file_value("max_context").content);
+    }
+    if (req.has_file("max_len"))
+    {
+        params.max_len = std::stoi(req.get_file_value("max_len").content);
+    }
+    if (req.has_file("best_of"))
+    {
+        params.best_of = std::stoi(req.get_file_value("best_of").content);
+    }
+    if (req.has_file("beam_size"))
+    {
+        params.beam_size = std::stoi(req.get_file_value("beam_size").content);
+    }
+    if (req.has_file("audio_ctx"))
+    {
+        params.audio_ctx = std::stof(req.get_file_value("audio_ctx").content);
+    }
+    if (req.has_file("word_thold"))
+    {
+        params.word_thold = std::stof(req.get_file_value("word_thold").content);
+    }
+    if (req.has_file("entropy_thold"))
+    {
+        params.entropy_thold = std::stof(req.get_file_value("entropy_thold").content);
+    }
+    if (req.has_file("logprob_thold"))
+    {
+        params.logprob_thold = std::stof(req.get_file_value("logprob_thold").content);
+    }
+    if (req.has_file("debug_mode"))
+    {
+        params.debug_mode = parse_str_to_bool(req.get_file_value("debug_mode").content);
+    }
+    if (req.has_file("translate"))
+    {
+        params.translate = parse_str_to_bool(req.get_file_value("translate").content);
+    }
+    if (req.has_file("diarize"))
+    {
+        params.diarize = parse_str_to_bool(req.get_file_value("diarize").content);
+    }
+    if (req.has_file("tinydiarize"))
+    {
+        params.tinydiarize = parse_str_to_bool(req.get_file_value("tinydiarize").content);
+    }
+    if (req.has_file("split_on_word"))
+    {
+        params.split_on_word = parse_str_to_bool(req.get_file_value("split_on_word").content);
+    }
+    if (req.has_file("no_timestamps"))
+    {
+        params.no_timestamps = parse_str_to_bool(req.get_file_value("no_timestamps").content);
+    }
+    if (req.has_file("language"))
+    {
+        params.language = req.get_file_value("language").content;
+    }
+    if (req.has_file("detect_language"))
+    {
+        params.detect_language = parse_str_to_bool(req.get_file_value("detect_language").content);
     }
     if (req.has_file("prompt"))
     {
         params.prompt = req.get_file_value("prompt").content;
     }
-    if (req.has_file("response-format"))
+    if (req.has_file("response_format"))
     {
-        params.response_format = req.get_file_value("response-format").content;
+        params.response_format = req.get_file_value("response_format").content;
     }
     if (req.has_file("temperature"))
     {
-        params.userdef_temp = std::stof(req.get_file_value("temperature").content);
+        params.temperature = std::stof(req.get_file_value("temperature").content);
+    }
+    if (req.has_file("temperature_inc"))
+    {
+        params.temperature_inc = std::stof(req.get_file_value("temperature_inc").content);
     }
 }
 
@@ -455,7 +497,7 @@ int main(int argc, char ** argv) {
         check_ffmpeg_availibility();
     }
     // whisper init
-    struct whisper_context_params cparams;
+    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);
@@ -471,19 +513,94 @@ int main(int argc, char ** argv) {
     Server svr;
     svr.set_default_headers({{"Server", "whisper.cpp"},
                              {"Access-Control-Allow-Origin", "*"},
-                             {"Access-Control-Allow-Headers", "content-type"}});
+                             {"Access-Control-Allow-Headers", "content-type, authorization"}});
 
-    std::string const default_content = "<html>hello</html>";
+    std::string const default_content = R"(
+    <html>
+    <head>
+        <title>Whisper.cpp Server</title>
+        <meta charset="utf-8">
+        <meta name="viewport" content="width=device-width">
+        <style>
+        body {
+            font-family: sans-serif;
+        }
+        form {
+            display: flex;
+            flex-direction: column;
+            align-items: flex-start;
+        }
+        label {
+            margin-bottom: 0.5rem;
+        }
+        input, select {
+            margin-bottom: 1rem;
+        }
+        button {
+            margin-top: 1rem;
+        }
+        </style>
+    </head>
+    <body>
+        <h1>Whisper.cpp Server</h1>
+
+        <h2>/inference</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/inference \
+    -H "Content-Type: multipart/form-data" \
+    -F file="@&lt;file-path&gt;" \
+    -F temperature="0.0" \
+    -F temperature_inc="0.2" \
+    -F response_format="json"
+        </pre>
+
+        <h2>/load</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/load \
+    -H "Content-Type: multipart/form-data" \
+    -F model="&lt;path-to-model-file&gt;"
+        </pre>
+
+        <div>
+            <h2>Try it out</h2>
+            <form action="/inference" method="POST" enctype="multipart/form-data">
+                <label for="file">Choose an audio file:</label>
+                <input type="file" id="file" name="file" accept="audio/*" required><br>
+
+                <label for="temperature">Temperature:</label>
+                <input type="number" id="temperature" name="temperature" value="0.0" step="0.01" placeholder="e.g., 0.0"><br>
+
+                <label for="response_format">Response Format:</label>
+                <select id="response_format" name="response_format">
+                    <option value="verbose_json">Verbose JSON</option>
+                    <option value="json">JSON</option>
+                    <option value="text">Text</option>
+                    <option value="srt">SRT</option>
+                    <option value="vtt">VTT</option>
+                </select><br>
+
+                <button type="submit">Submit</button>
+            </form>
+        </div>
+    </body>
+    </html>
+    )";
+
+    // store default params so we can reset after each inference request
+    whisper_params default_params = params;
 
     // this is only called if no index.html is found in the public --path
-    svr.Get("/", [&default_content](const Request &, Response &res){
+    svr.Get(sparams.request_path + "/", [&default_content](const Request &, Response &res){
         res.set_content(default_content, "text/html");
         return false;
     });
 
-    svr.Post("/inference", [&](const Request &req, Response &res){
+    svr.Options(sparams.request_path + "/inference", [&](const Request &, Response &){
+    });
+
+    svr.Post(sparams.request_path + "/inference", [&](const Request &req, Response &res){
         // acquire whisper model mutex lock
-        whisper_mutex.lock();
+        std::lock_guard<std::mutex> lock(whisper_mutex);
 
         // first check user requested fields of the request
         if (!req.has_file("file"))
@@ -491,7 +608,6 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "error: no 'file' field in the request\n");
             const std::string error_resp = "{\"error\":\"no 'file' field in the request\"}";
             res.set_content(error_resp, "application/json");
-            whisper_mutex.unlock();
             return;
         }
         auto audio_file = req.get_file_value("file");
@@ -506,35 +622,42 @@ int main(int argc, char ** argv) {
         std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // write to temporary file
-        const std::string temp_filename = "whisper_server_temp_file.wav";
-        std::ofstream temp_file{temp_filename, std::ios::binary};
-        temp_file << audio_file.content;
-        temp_file.close();
-
-        // if file is not wav, convert to wav
-        
         if (sparams.ffmpeg_converter) {
+            // if file is not wav, convert to wav
+            // write to temporary file
+            const std::string temp_filename = "whisper_server_temp_file.wav";
+            std::ofstream temp_file{temp_filename, std::ios::binary};
+            temp_file << audio_file.content;
+            temp_file.close();
+
             std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
             const bool is_converted = convert_to_wav(temp_filename, error_resp);
             if (!is_converted) {
                 res.set_content(error_resp, "application/json");
-                whisper_mutex.unlock();
+                return;
+            }
+
+            // read wav content into pcmf32
+            if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize))
+            {
+                fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
+                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+                res.set_content(error_resp, "application/json");
+                std::remove(temp_filename.c_str());
+                return;
+            }
+            // remove temp file
+            std::remove(temp_filename.c_str());
+        } else {
+            if (!::read_wav(audio_file.content, pcmf32, pcmf32s, params.diarize))
+            {
+                fprintf(stderr, "error: failed to read WAV file\n");
+                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+                res.set_content(error_resp, "application/json");
                 return;
             }
         }
 
-        // read wav content into pcmf32
-        if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize)) {
-            fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
-            const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
-            res.set_content(error_resp, "application/json");
-            std::remove(temp_filename.c_str());
-            whisper_mutex.unlock();
-            return;
-        }
-        // remove temp file
-        std::remove(temp_filename.c_str());
 
         printf("Successfully loaded %s\n", filename.c_str());
 
@@ -591,6 +714,7 @@ int main(int argc, char ** argv) {
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = 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;
@@ -602,10 +726,14 @@ int main(int argc, char ** argv) {
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.temperature_inc  = params.userdef_temp;
+            wparams.temperature      = params.temperature;
+            wparams.temperature_inc  = params.temperature_inc;
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
+            wparams.no_timestamps    = params.no_timestamps;
+            wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
+
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             // this callback is called on each new segment
@@ -648,7 +776,6 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                 const std::string error_resp = "{\"error\":\"failed to process audio\"}";
                 res.set_content(error_resp, "application/json");
-                whisper_mutex.unlock();
                 return;
             }
         }
@@ -702,6 +829,59 @@ int main(int argc, char ** argv) {
                 ss << speaker << text << "\n\n";
             }
             res.set_content(ss.str(), "text/vtt");
+        } else if (params.response_format == vjson_format) {
+            /* try to match openai/whisper's Python format */
+            std::string results = output_str(ctx, params, pcmf32s);
+            json jres = json{
+                {"task", params.translate ? "translate" : "transcribe"},
+                {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
+                {"duration", float(pcmf32.size())/WHISPER_SAMPLE_RATE},
+                {"text", results},
+                {"segments", json::array()}
+            };
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i)
+            {
+                json segment = json{
+                    {"id", i},
+                    {"text", whisper_full_get_segment_text(ctx, i)},
+                };
+
+                if (!params.no_timestamps) {
+                    segment["start"] = whisper_full_get_segment_t0(ctx, i) * 0.01;
+                    segment["end"] = whisper_full_get_segment_t1(ctx, i) * 0.01;
+                }
+
+                float total_logprob = 0;
+                const int n_tokens = whisper_full_n_tokens(ctx, i);
+                for (int j = 0; j < n_tokens; ++j) {
+                    whisper_token_data token = whisper_full_get_token_data(ctx, i, j);
+                    if (token.id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+
+                    segment["tokens"].push_back(token.id);
+                    json word = json{{"word", whisper_full_get_token_text(ctx, i, j)}};
+                    if (!params.no_timestamps) {
+                        word["start"] = token.t0 * 0.01;
+                        word["end"] = token.t1 * 0.01;
+                    }
+                    word["probability"] = token.p;
+                    total_logprob += token.plog;
+                    segment["words"].push_back(word);
+                }
+
+                segment["temperature"] = params.temperature;
+                segment["avg_logprob"] = total_logprob / n_tokens;
+
+                // TODO compression_ratio and no_speech_prob are not implemented yet
+                // segment["compression_ratio"] = 0;
+                // segment["no_speech_prob"] = 0;
+
+                jres["segments"].push_back(segment);
+            }
+            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
+                            "application/json");
         }
         // TODO add more output formats
         else
@@ -714,17 +894,16 @@ int main(int argc, char ** argv) {
                             "application/json");
         }
 
-        // return whisper model mutex lock
-        whisper_mutex.unlock();
+        // reset params to thier defaults
+        params = default_params;
     });
-    svr.Post("/load", [&](const Request &req, Response &res){
-        whisper_mutex.lock();
+    svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
+        std::lock_guard<std::mutex> lock(whisper_mutex);
         if (!req.has_file("model"))
         {
             fprintf(stderr, "error: no 'model' field in the request\n");
             const std::string error_resp = "{\"error\":\"no 'model' field in the request\"}";
             res.set_content(error_resp, "application/json");
-            whisper_mutex.unlock();
             return;
         }
         std::string model = req.get_file_value("model").content;
@@ -733,7 +912,6 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "error: 'model': %s not found!\n", model.c_str());
             const std::string error_resp = "{\"error\":\"model not found!\"}";
             res.set_content(error_resp, "application/json");
-            whisper_mutex.unlock();
             return;
         }
 
@@ -756,7 +934,6 @@ int main(int argc, char ** argv) {
         res.set_content(success, "application/text");
 
         // check if the model is in the file system
-        whisper_mutex.unlock();
     });
 
     svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) {
@@ -773,11 +950,11 @@ int main(int argc, char ** argv) {
         res.status = 500;
     });
 
-    svr.set_error_handler([](const Request &, Response &res) {
+    svr.set_error_handler([](const Request &req, Response &res) {
         if (res.status == 400) {
             res.set_content("Invalid request", "text/plain");
         } else if (res.status != 500) {
-            res.set_content("File Not Found", "text/plain");
+            res.set_content("File Not Found (" + req.path + ")", "text/plain");
             res.status = 404;
         }
     });

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);
 

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());
@@ -166,7 +152,7 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
-    struct whisper_context_params cparams;
+    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);
@@ -372,7 +358,7 @@ 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]";

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

@@ -1,30 +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)
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
 
-    # TODO: this is temporary
-    #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET}
-        talk-llama.cpp
-        llama.cpp
-        ../common.cpp
-        ../common-sdl.cpp
-        ../../ggml.c
-        ../../ggml-alloc.c
-        ../../ggml-backend.c
-        ../../ggml-quants.c
-        ../../whisper.cpp)
+    if (WHISPER_CLBLAST)
+        set(CLBLAST_LIBNAME clblast)
+    endif ()
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CLBLAST_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
 
     if(WIN32)
-    # It requires Windows 8.1 or later for PrefetchVirtualMemory
-    target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
+        # It requires Windows 8.1 or later for PrefetchVirtualMemory
+        target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
     endif()
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
-
     include(DefaultTargetOptions)
 endif ()

examples/talk-llama/README.md

@@ -15,9 +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/llama.h

@@ -2,12 +2,8 @@
 #define LLAMA_H
 
 #include "ggml.h"
-#ifdef GGML_USE_CUBLAS
-#include "ggml-cuda.h"
-#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
-#else
-#define LLAMA_MAX_DEVICES 1
-#endif // GGML_USE_CUBLAS
+#include "ggml-backend.h"
+
 #include <stddef.h>
 #include <stdint.h>
 #include <stdio.h>
@@ -39,15 +35,11 @@
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
+#define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 2
-
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
-// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
-#define LLAMA_SUPPORTS_GPU_OFFLOAD
-#endif
+#define LLAMA_SESSION_VERSION 4
 
 #ifdef __cplusplus
 extern "C" {
@@ -67,8 +59,19 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
-        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
+    };
+
+    // note: these values should be synchronized with ggml_rope
+    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
+    enum llama_rope_type {
+        LLAMA_ROPE_TYPE_NONE = -1,
+        LLAMA_ROPE_TYPE_NORM =  0,
+        LLAMA_ROPE_TYPE_NEOX =  2,
+        LLAMA_ROPE_TYPE_GLM  =  4,
     };
 
     enum llama_token_type {
@@ -102,16 +105,41 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_XS        = 22, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ1_S         = 24, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ4_NL        = 25, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_S         = 26, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_M         = 27, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_S         = 28, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
 
     enum llama_rope_scaling_type {
-        LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
-        LLAMA_ROPE_SCALING_NONE        = 0,
-        LLAMA_ROPE_SCALING_LINEAR      = 1,
-        LLAMA_ROPE_SCALING_YARN        = 2,
-        LLAMA_ROPE_SCALING_MAX_VALUE   = LLAMA_ROPE_SCALING_YARN,
+        LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
+        LLAMA_ROPE_SCALING_TYPE_NONE        = 0,
+        LLAMA_ROPE_SCALING_TYPE_LINEAR      = 1,
+        LLAMA_ROPE_SCALING_TYPE_YARN        = 2,
+        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_YARN,
+    };
+
+    enum llama_pooling_type {
+        LLAMA_POOLING_TYPE_UNSPECIFIED = -1,
+        LLAMA_POOLING_TYPE_NONE = 0,
+        LLAMA_POOLING_TYPE_MEAN = 1,
+        LLAMA_POOLING_TYPE_CLS  = 2,
+    };
+
+    enum llama_split_mode {
+        LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
+        LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
+        LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
     };
 
     typedef struct llama_token_data {
@@ -126,7 +154,7 @@ extern "C" {
         bool sorted;
     } llama_token_data_array;
 
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);
 
     // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
@@ -136,7 +164,7 @@ extern "C" {
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
     // - seq_id : the sequence to which the respective token belongs
-    // - logits : if zero, the logits for the respective token will not be output
+    // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -146,7 +174,7 @@ extern "C" {
         llama_pos    *  pos;
         int32_t      *  n_seq_id;
         llama_seq_id ** seq_id;
-        int8_t       *  logits;
+        int8_t       *  logits; // TODO: rename this to "output"
 
         // NOTE: helpers for smooth API transition - can be deprecated in the future
         //       for future-proof code, use the above fields instead and ignore everything below
@@ -158,16 +186,46 @@ extern "C" {
         llama_seq_id all_seq_id; // used if seq_id == NULL
     } llama_batch;
 
+    enum llama_model_kv_override_type {
+        LLAMA_KV_OVERRIDE_TYPE_INT,
+        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
+        LLAMA_KV_OVERRIDE_TYPE_BOOL,
+    };
+
+    struct llama_model_kv_override {
+        char key[128];
+        enum llama_model_kv_override_type tag;
+        union {
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
+        };
+    };
+
     struct llama_model_params {
         int32_t n_gpu_layers; // number of layers to store in VRAM
-        int32_t main_gpu;     // the GPU that is used for scratch and small tensors
-        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
+        enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
-        // called with a progress value between 0 and 1, pass NULL to disable
+        // main_gpu interpretation depends on split_mode:
+        // LLAMA_SPLIT_NONE: the GPU that is used for the entire model
+        // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results
+        // LLAMA_SPLIT_LAYER: ignored
+        int32_t main_gpu;
+
+        // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
+        const float * tensor_split;
+
+        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
+        // If the provided progress_callback returns true, model loading continues.
+        // If it returns false, model loading is immediately aborted.
         llama_progress_callback progress_callback;
+
         // context pointer passed to the progress callback
         void * progress_callback_user_data;
 
+        // override key-value pairs of the model meta data
+        const struct llama_model_kv_override * kv_overrides;
+
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only; // only load the vocabulary, no weights
         bool use_mmap;   // use mmap if possible
@@ -177,35 +235,53 @@ extern "C" {
     struct llama_context_params {
         uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
-        uint32_t n_batch;           // prompt processing maximum batch size
+        uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
+        uint32_t n_ubatch;          // physical maximum batch size
+        uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
         uint32_t n_threads;         // number of threads to use for generation
         uint32_t n_threads_batch;   // number of threads to use for batch processing
-        int8_t   rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
+
+        enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
+        enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
+                                                        // (ignored if no pooling layer)
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
         float    rope_freq_base;   // RoPE base frequency, 0 = from model
         float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
-        float    yarn_ext_factor;  // YaRN extrapolation mix factor, NaN = from model
+        float    yarn_ext_factor;  // YaRN extrapolation mix factor, negative = from model
         float    yarn_attn_factor; // YaRN magnitude scaling factor
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
+        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+
+        ggml_backend_sched_eval_callback cb_eval;
+        void * cb_eval_user_data;
+
+        enum ggml_type type_k; // data type for K cache
+        enum ggml_type type_v; // data type for V cache
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
-        bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
-        bool embedding;  // embedding mode only
+        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
+        bool embeddings;  // if true, extract embeddings (together with logits)
+        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+
+        // Abort callback
+        // if it returns true, execution of llama_decode() will be aborted
+        // currently works only with CPU execution
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
     };
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int nthread;                 // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
         enum llama_ftype ftype;      // quantize to this llama_ftype
         bool allow_requantize;       // allow quantizing non-f32/f16 tensors
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
+        bool pure;                   // quantize all tensors to the default type
+        void * imatrix;              // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types
@@ -256,6 +332,12 @@ extern "C" {
         int32_t n_eval;
     };
 
+    // used in chat template
+    typedef struct llama_chat_message {
+        const char * role;
+        const char * content;
+    } llama_chat_message;
+
     // Helpers for getting default parameters
     LLAMA_API struct llama_model_params llama_model_default_params(void);
     LLAMA_API struct llama_context_params llama_context_default_params(void);
@@ -264,7 +346,10 @@ extern "C" {
     // Initialize the llama + ggml backend
     // If numa is true, use NUMA optimizations
     // Call once at the start of the program
-    LLAMA_API void llama_backend_init(bool numa);
+    LLAMA_API void llama_backend_init(void);
+
+    //optional:
+    LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
@@ -284,25 +369,48 @@ extern "C" {
 
     LLAMA_API int64_t llama_time_us(void);
 
-    LLAMA_API int  llama_max_devices    (void);
-    LLAMA_API bool llama_mmap_supported (void);
-    LLAMA_API bool llama_mlock_supported(void);
+    LLAMA_API size_t llama_max_devices(void);
+
+    LLAMA_API bool llama_supports_mmap       (void);
+    LLAMA_API bool llama_supports_mlock      (void);
+    LLAMA_API bool llama_supports_gpu_offload(void);
 
     LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
 
-    LLAMA_API int llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
 
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);
 
-    LLAMA_API int llama_n_vocab    (const struct llama_model * model);
-    LLAMA_API int llama_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int llama_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
+    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
 
+    // Functions to access the model's GGUF metadata scalar values
+    // - The functions return the length of the string on success, or -1 on failure
+    // - The output string is always null-terminated and cleared on failure
+    // - GGUF array values are not supported by these functions
+
+    // Get metadata value as a string by key name
+    LLAMA_API int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size);
+
+    // Get the number of metadata key/value pairs
+    LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model);
+
+    // Get metadata key name by index
+    LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
+
+    // Get metadata value as a string by index
+    LLAMA_API int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size);
+
     // Get a string describing the model type
-    LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
+    LLAMA_API int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size);
 
     // Returns the total size of all the tensors in the model in bytes
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
@@ -314,7 +422,7 @@ extern "C" {
     LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
 
     // Returns 0 on success
-    LLAMA_API int llama_model_quantize(
+    LLAMA_API uint32_t llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
             const llama_model_quantize_params * params);
@@ -325,28 +433,71 @@ extern "C" {
     // The model needs to be reloaded before applying a new adapter, otherwise the adapter
     // will be applied on top of the previous one
     // Returns 0 on success
-    LLAMA_API DEPRECATED(int llama_apply_lora_from_file(
-            struct llama_context * ctx,
-                      const char * path_lora,
-                           float   scale,
-                      const char * path_base_model,
-                             int   n_threads),
-            "use llama_model_apply_lora_from_file instead");
-
-    LLAMA_API int llama_model_apply_lora_from_file(
+    LLAMA_API int32_t llama_model_apply_lora_from_file(
             const struct llama_model * model,
                       const char * path_lora,
                            float   scale,
                       const char * path_base_model,
-                             int   n_threads);
+                         int32_t   n_threads);
 
     //
     // KV cache
     //
 
-    // Returns the number of tokens in the KV cache
-    LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx),
-            "avoid using this, it will be removed in the future, instead - count the tokens in user code");
+    // Information associated with an individual cell in the KV cache view.
+    struct llama_kv_cache_view_cell {
+        // The position for this cell. Takes KV cache shifts into account.
+        // May be negative if the cell is not populated.
+        llama_pos pos;
+    };
+
+    // An updateable view of the KV cache.
+    struct llama_kv_cache_view {
+        // Number of KV cache cells. This will be the same as the context size.
+        int32_t n_cells;
+
+        // Maximum number of sequences that can exist in a cell. It's not an error
+        // if there are more sequences in a cell than this value, however they will
+        // not be visible in the view cells_sequences.
+        int32_t n_seq_max;
+
+        // Number of tokens in the cache. For example, if there are two populated
+        // cells, the first with 1 sequence id in it and the second with 2 sequence
+        // ids then you'll have 3 tokens.
+        int32_t token_count;
+
+        // Number of populated cache cells.
+        int32_t used_cells;
+
+        // Maximum contiguous empty slots in the cache.
+        int32_t max_contiguous;
+
+        // Index to the start of the max_contiguous slot range. Can be negative
+        // when cache is full.
+        int32_t max_contiguous_idx;
+
+        // Information for an individual cell.
+        struct llama_kv_cache_view_cell * cells;
+
+        // The sequences for each cell. There will be n_seq_max items per cell.
+        llama_seq_id * cells_sequences;
+    };
+
+    // Create an empty KV cache view. (use only for debugging purposes)
+    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);
+
+    // Free a KV cache view. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
+
+    // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)
+    LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view);
+
+    // Returns the number of tokens in the KV cache (slow, use only for debug)
+    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
+    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
+
+    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
+    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
 
     // Clear the KV cache
     LLAMA_API void llama_kv_cache_clear(
@@ -356,7 +507,7 @@ extern "C" {
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_rm(
+    LLAMA_API bool llama_kv_cache_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -379,16 +530,45 @@ extern "C" {
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_shift(
+    LLAMA_API void llama_kv_cache_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
                        llama_pos   delta);
 
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
+    // Returns the largest position present in the KV cache for the specified sequence
+    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Defragment the KV cache
+    // This will be applied:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
+    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
+
+    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
+
     //
     // State / sessions
     //
@@ -408,7 +588,7 @@ extern "C" {
     // Returns the number of bytes read
     LLAMA_API size_t llama_set_state_data(
             struct llama_context * ctx,
-                         uint8_t * src);
+                   const uint8_t * src);
 
     // Save/load session file
     LLAMA_API bool llama_load_session_file(
@@ -428,27 +608,6 @@ extern "C" {
     // Decoding
     //
 
-    // Run the llama inference to obtain the logits and probabilities for the next token(s).
-    // tokens + n_tokens is the provided batch of new tokens to process
-    // n_past is the number of tokens to use from previous eval calls
-    // Returns 0 on success
-    // DEPRECATED: use llama_decode() instead
-    LLAMA_API DEPRECATED(int llama_eval(
-            struct llama_context * ctx,
-                     llama_token * tokens,
-                         int32_t   n_tokens,
-                             int   n_past),
-            "use llama_decode() instead");
-
-    // Same as llama_eval, but use float matrix input directly.
-    // DEPRECATED: use llama_decode() instead
-    LLAMA_API DEPRECATED(int llama_eval_embd(
-            struct llama_context * ctx,
-                           float * embd,
-                         int32_t   n_tokens,
-                             int   n_past),
-            "use llama_decode() instead");
-
     // Return batch for single sequence of tokens starting at pos_0
     //
     // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
@@ -478,7 +637,7 @@ extern "C" {
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
     // < 0 - error
-    LLAMA_API int llama_decode(
+    LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
@@ -487,7 +646,19 @@ extern "C" {
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
     LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
-    // Token logits obtained from the last call to llama_eval()
+    // Set whether to use causal attention or not
+    // If set to true, the model will only attend to the past tokens
+    LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
+
+    // Set abort callback
+    LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    // Wait until all computations are finished
+    // This is automatically done when using one of the functions below to obtain the computation results
+    // and is not necessary to call it explicitly in most cases
+    LLAMA_API void llama_synchronize(struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_decode()
     // The logits for the last token are stored in the last row
     // Logits for which llama_batch.logits[i] == 0 are undefined
     // Rows: n_tokens provided with llama_batch
@@ -498,10 +669,20 @@ extern "C" {
     // llama_get_logits(ctx) + i*n_vocab
     LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);
 
-    // Get the embeddings for the input
-    // shape: [n_embd] (1-dimensional)
+    // Get all output token embeddings
+    // shape: [n_tokens*n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
+    // Get the embeddings for the ith token
+    // llama_get_embeddings(ctx) + i*n_embd
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the embeddings for a sequence id
+    // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
+
     //
     // Vocab
     //
@@ -517,6 +698,12 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
 
+    // Returns -1 if unknown, 1 for true or 0 for false.
+    LLAMA_API int32_t         llama_add_bos_token(const struct llama_model * model);
+
+    // Returns -1 if unknown, 1 for true or 0 for false.
+    LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);
+
     // codellama infill tokens
     LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
     LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
@@ -529,16 +716,16 @@ extern "C" {
 
     /// @details Convert the provided text into tokens.
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
-    /// @return Returns the number of tokens on success, no more than n_max_tokens
+    /// @return Returns the number of tokens on success, no more than n_tokens_max
     /// @return Returns a negative number on failure - the number of tokens that would have been returned
     /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
     ///                Does not insert a leading space.
-    LLAMA_API int llama_tokenize(
+    LLAMA_API int32_t llama_tokenize(
         const struct llama_model * model,
                       const char * text,
-                             int   text_len,
+                         int32_t   text_len,
                      llama_token * tokens,
-                             int   n_max_tokens,
+                         int32_t   n_tokens_max,
                             bool   add_bos,
                             bool   special);
 
@@ -546,11 +733,30 @@ extern "C" {
     // Uses the vocabulary in the provided context.
     // Does not write null terminator to the buffer.
     // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
-    LLAMA_API int llama_token_to_piece(
+    LLAMA_API int32_t llama_token_to_piece(
               const struct llama_model * model,
                            llama_token   token,
                                   char * buf,
-                                  int    length);
+                               int32_t   length);
+
+    /// Apply chat template. Inspired by hf apply_chat_template() on python.
+    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
+    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+    /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
+    /// @param chat Pointer to a list of multiple llama_chat_message
+    /// @param n_msg Number of llama_chat_message in this chat
+    /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
+    /// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
+    /// @param length The size of the allocated buffer
+    /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
+    LLAMA_API int32_t llama_chat_apply_template(
+              const struct llama_model * model,
+                            const char * tmpl,
+       const struct llama_chat_message * chat,
+                                size_t   n_msg,
+                                  bool   add_ass,
+                                  char * buf,
+                               int32_t   length);
 
     //
     // Grammar
@@ -584,13 +790,13 @@ extern "C" {
                            float   penalty_present);
 
     /// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
-    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
-    /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
-    /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    LLAMA_API void llama_sample_classifier_free_guidance(
+    /// @param logits Logits extracted from the original generation context.
+    /// @param logits_guidance Logits extracted from a separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
+    /// @param scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
+    LLAMA_API void llama_sample_apply_guidance(
               struct llama_context * ctx,
-            llama_token_data_array * candidates,
-              struct llama_context * guidance_ctx,
+                             float * logits,
+                             float * logits_guidance,
                              float   scale);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
@@ -602,7 +808,7 @@ extern "C" {
     LLAMA_API void llama_sample_top_k(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
-                             int   k,
+                         int32_t   k,
                           size_t   min_keep);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
@@ -633,17 +839,19 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
+    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
+    LLAMA_API void llama_sample_entropy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates_p,
+                           float   min_temp,
+                           float   max_temp,
+                           float   exponent_val);
+
     LLAMA_API void llama_sample_temp(
             struct llama_context * ctx,
           llama_token_data_array * candidates,
                            float   temp);
 
-    LLAMA_API DEPRECATED(void llama_sample_temperature(
-                struct llama_context * ctx,
-              llama_token_data_array * candidates,
-                               float   temp),
-            "use llama_sample_temp instead");
-
     /// @details Apply constraints from grammar
     LLAMA_API void llama_sample_grammar(
             struct llama_context * ctx,
@@ -661,7 +869,7 @@ extern "C" {
           llama_token_data_array * candidates,
                            float   tau,
                            float   eta,
-                             int   m,
+                         int32_t   m,
                            float * mu);
 
     /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
@@ -734,8 +942,8 @@ extern "C" {
         llama_beam_search_callback_fn_t   callback,
                                    void * callback_data,
                                  size_t   n_beams,
-                                    int   n_past,
-                                    int   n_predict);
+                                int32_t   n_past,
+                                int32_t   n_predict);
 
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);

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

@@ -14,6 +14,7 @@
 #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) {
     auto * model = llama_get_model(ctx);
@@ -67,10 +68,14 @@ struct whisper_params {
     bool use_gpu        = true;
 
     std::string person      = "Georgi";
+    std::string bot_name    = "LLaMA";
+    std::string wake_cmd    = "";
+    std::string heard_ok    = "";
     std::string language    = "en";
     std::string model_wsp   = "models/ggml-base.en.bin";
     std::string model_llama = "models/ggml-llama-7B.bin";
     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
@@ -101,11 +106,15 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-vp"  || arg == "--verbose-prompt") { params.verbose_prompt = true; }
         else if (arg == "-ng"  || arg == "--no-gpu")         { params.use_gpu        = false; }
         else if (arg == "-p"   || arg == "--person")         { params.person         = argv[++i]; }
-        else if (arg == "--session")                         { params.path_session   = 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));
@@ -146,10 +155,14 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -vp,      --verbose-prompt [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
     fprintf(stderr, "  -ng,      --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -p NAME,  --person NAME    [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
+    fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
+    fprintf(stderr, "  -ho TEXT, --heard-ok TEXT  [%-7s] said by TTS before generating reply\n",         params.heard_ok.c_str());
     fprintf(stderr, "  -l LANG,  --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -mw FILE, --model-whisper  [%-7s] whisper model file\n",                          params.model_wsp.c_str());
     fprintf(stderr, "  -ml FILE, --model-llama    [%-7s] llama model file\n",                            params.model_llama.c_str());
     fprintf(stderr, "  -s FILE,  --speak TEXT     [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  -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());
@@ -224,6 +237,18 @@ 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}.
@@ -259,14 +284,14 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context_params cparams;
+    struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
 
     // llama init
 
-    llama_backend_init(true);
+    llama_backend_init();
 
     auto lmparams = llama_model_default_params();
     if (!params.use_gpu) {
@@ -282,7 +307,6 @@ int main(int argc, char ** argv) {
     // tune these to your liking
     lcparams.n_ctx      = 2048;
     lcparams.seed       = 1;
-    lcparams.f16_kv     = true;
     lcparams.n_threads  = params.n_threads;
 
     struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
@@ -324,12 +348,11 @@ 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;
@@ -338,7 +361,7 @@ 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
@@ -368,6 +391,8 @@ 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;
@@ -403,8 +428,21 @@ 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)) {
-        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;
     }
 
@@ -440,6 +478,16 @@ 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);
@@ -485,10 +533,38 @@ 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
@@ -525,7 +601,7 @@ 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);
 
@@ -586,8 +662,21 @@ int main(int argc, char ** argv) {
                             n_session_consumed = session_tokens.size();
                         }
 
-                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past)) {
-                            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;
                         }
                     }
@@ -658,6 +747,7 @@ int main(int argc, char ** argv) {
                             text_to_speak += llama_token_to_piece(ctx_llama, id);
 
                             printf("%s", llama_token_to_piece(ctx_llama, id).c_str());
+                            fflush(stdout);
                         }
                     }
 
@@ -686,11 +776,7 @@ 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.h

@@ -1,462 +1,26 @@
-#pragma once
+#pragma once
 
-#include <cassert>
-#include <stdexcept>
+#include <cstdint>
+#include <string>
 #include <vector>
-#include <unordered_map>
-
-static const std::vector<std::pair<uint32_t, uint32_t>> digit_ranges = {
-{0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F},
-{0xCE6, 0xCEF}, {0xD66, 0xD6F}, {0xDE6, 0xDEF}, {0xE50, 0xE59}, {0xED0, 0xED9}, {0xF20, 0xF29}, {0x1040, 0x1049}, {0x1090, 0x1099}, {0x1369, 0x1371}, {0x17E0, 0x17E9}, {0x1810, 0x1819}, {0x1946, 0x194F},
-{0x19D0, 0x19DA}, {0x1A80, 0x1A89}, {0x1A90, 0x1A99}, {0x1B50, 0x1B59}, {0x1BB0, 0x1BB9}, {0x1C40, 0x1C49}, {0x1C50, 0x1C59}, {0x2070, 0x2070}, {0x2074, 0x2079}, {0x2080, 0x2089}, {0x2460, 0x2468},
-{0x2474, 0x247C}, {0x2488, 0x2490}, {0x24EA, 0x24EA}, {0x24F5, 0x24FD}, {0x24FF, 0x24FF}, {0x2776, 0x277E}, {0x2780, 0x2788}, {0x278A, 0x2792}, {0xA620, 0xA629}, {0xA8D0, 0xA8D9}, {0xA900, 0xA909},
-{0xA9D0, 0xA9D9}, {0xA9F0, 0xA9F9}, {0xAA50, 0xAA59}, {0xABF0, 0xABF9}, {0xFF10, 0xFF19}, {0x104A0, 0x104A9}, {0x10A40, 0x10A43}, {0x10D30, 0x10D39}, {0x10E60, 0x10E68}, {0x11052, 0x1105A},
-{0x11066, 0x1106F}, {0x110F0, 0x110F9}, {0x11136, 0x1113F}, {0x111D0, 0x111D9}, {0x112F0, 0x112F9}, {0x11450, 0x11459}, {0x114D0, 0x114D9}, {0x11650, 0x11659}, {0x116C0, 0x116C9}, {0x11730, 0x11739},
-{0x118E0, 0x118E9}, {0x11950, 0x11959}, {0x11C50, 0x11C59}, {0x11D50, 0x11D59}, {0x11DA0, 0x11DA9}, {0x16A60, 0x16A69}, {0x16B50, 0x16B59}, {0x1D7CE, 0x1D7FF}, {0x1E140, 0x1E149}, {0x1E2F0, 0x1E2F9},
-{0x1E950, 0x1E959}, {0x1F100, 0x1F10A}, {0x1FBF0, 0x1FBF9},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> letter_ranges = {
-{0x41, 0x5A}, {0x61, 0x7A}, {0xAA, 0xAA}, {0xB5, 0xB5}, {0xBA, 0xBA}, {0xC0, 0xD6}, {0xD8, 0xF6}, {0xF8, 0x2C1}, {0x2C6, 0x2D1}, {0x2E0, 0x2E4}, {0x2EC, 0x2EC}, {0x2EE, 0x2EE}, {0x370, 0x374},
-{0x376, 0x377}, {0x37A, 0x37D}, {0x37F, 0x37F}, {0x386, 0x386}, {0x388, 0x38A}, {0x38C, 0x38C}, {0x38E, 0x3A1}, {0x3A3, 0x3F5}, {0x3F7, 0x481}, {0x48A, 0x52F}, {0x531, 0x556}, {0x559, 0x559},
-{0x560, 0x588}, {0x5D0, 0x5EA}, {0x5EF, 0x5F2}, {0x620, 0x64A}, {0x66E, 0x66F}, {0x671, 0x6D3}, {0x6D5, 0x6D5}, {0x6E5, 0x6E6}, {0x6EE, 0x6EF}, {0x6FA, 0x6FC}, {0x6FF, 0x6FF}, {0x710, 0x710},
-{0x712, 0x72F}, {0x74D, 0x7A5}, {0x7B1, 0x7B1}, {0x7CA, 0x7EA}, {0x7F4, 0x7F5}, {0x7FA, 0x7FA}, {0x800, 0x815}, {0x81A, 0x81A}, {0x824, 0x824}, {0x828, 0x828}, {0x840, 0x858}, {0x860, 0x86A},
-{0x8A0, 0x8B4}, {0x8B6, 0x8C7}, {0x904, 0x939}, {0x93D, 0x93D}, {0x950, 0x950}, {0x958, 0x961}, {0x971, 0x980}, {0x985, 0x98C}, {0x98F, 0x990}, {0x993, 0x9A8}, {0x9AA, 0x9B0}, {0x9B2, 0x9B2},
-{0x9B6, 0x9B9}, {0x9BD, 0x9BD}, {0x9CE, 0x9CE}, {0x9DC, 0x9DD}, {0x9DF, 0x9E1}, {0x9F0, 0x9F1}, {0x9FC, 0x9FC}, {0xA05, 0xA0A}, {0xA0F, 0xA10}, {0xA13, 0xA28}, {0xA2A, 0xA30}, {0xA32, 0xA33},
-{0xA35, 0xA36}, {0xA38, 0xA39}, {0xA59, 0xA5C}, {0xA5E, 0xA5E}, {0xA72, 0xA74}, {0xA85, 0xA8D}, {0xA8F, 0xA91}, {0xA93, 0xAA8}, {0xAAA, 0xAB0}, {0xAB2, 0xAB3}, {0xAB5, 0xAB9}, {0xABD, 0xABD},
-{0xAD0, 0xAD0}, {0xAE0, 0xAE1}, {0xAF9, 0xAF9}, {0xB05, 0xB0C}, {0xB0F, 0xB10}, {0xB13, 0xB28}, {0xB2A, 0xB30}, {0xB32, 0xB33}, {0xB35, 0xB39}, {0xB3D, 0xB3D}, {0xB5C, 0xB5D}, {0xB5F, 0xB61},
-{0xB71, 0xB71}, {0xB83, 0xB83}, {0xB85, 0xB8A}, {0xB8E, 0xB90}, {0xB92, 0xB95}, {0xB99, 0xB9A}, {0xB9C, 0xB9C}, {0xB9E, 0xB9F}, {0xBA3, 0xBA4}, {0xBA8, 0xBAA}, {0xBAE, 0xBB9}, {0xBD0, 0xBD0},
-{0xC05, 0xC0C}, {0xC0E, 0xC10}, {0xC12, 0xC28}, {0xC2A, 0xC39}, {0xC3D, 0xC3D}, {0xC58, 0xC5A}, {0xC60, 0xC61}, {0xC80, 0xC80}, {0xC85, 0xC8C}, {0xC8E, 0xC90}, {0xC92, 0xCA8}, {0xCAA, 0xCB3},
-{0xCB5, 0xCB9}, {0xCBD, 0xCBD}, {0xCDE, 0xCDE}, {0xCE0, 0xCE1}, {0xCF1, 0xCF2}, {0xD04, 0xD0C}, {0xD0E, 0xD10}, {0xD12, 0xD3A}, {0xD3D, 0xD3D}, {0xD4E, 0xD4E}, {0xD54, 0xD56}, {0xD5F, 0xD61},
-{0xD7A, 0xD7F}, {0xD85, 0xD96}, {0xD9A, 0xDB1}, {0xDB3, 0xDBB}, {0xDBD, 0xDBD}, {0xDC0, 0xDC6}, {0xE01, 0xE30}, {0xE32, 0xE33}, {0xE40, 0xE46}, {0xE81, 0xE82}, {0xE84, 0xE84}, {0xE86, 0xE8A},
-{0xE8C, 0xEA3}, {0xEA5, 0xEA5}, {0xEA7, 0xEB0}, {0xEB2, 0xEB3}, {0xEBD, 0xEBD}, {0xEC0, 0xEC4}, {0xEC6, 0xEC6}, {0xEDC, 0xEDF}, {0xF00, 0xF00}, {0xF40, 0xF47}, {0xF49, 0xF6C}, {0xF88, 0xF8C},
-{0x1000, 0x102A}, {0x103F, 0x103F}, {0x1050, 0x1055}, {0x105A, 0x105D}, {0x1061, 0x1061}, {0x1065, 0x1066}, {0x106E, 0x1070}, {0x1075, 0x1081}, {0x108E, 0x108E}, {0x10A0, 0x10C5}, {0x10C7, 0x10C7},
-{0x10CD, 0x10CD}, {0x10D0, 0x10FA}, {0x10FC, 0x1248}, {0x124A, 0x124D}, {0x1250, 0x1256}, {0x1258, 0x1258}, {0x125A, 0x125D}, {0x1260, 0x1288}, {0x128A, 0x128D}, {0x1290, 0x12B0}, {0x12B2, 0x12B5},
-{0x12B8, 0x12BE}, {0x12C0, 0x12C0}, {0x12C2, 0x12C5}, {0x12C8, 0x12D6}, {0x12D8, 0x1310}, {0x1312, 0x1315}, {0x1318, 0x135A}, {0x1380, 0x138F}, {0x13A0, 0x13F5}, {0x13F8, 0x13FD}, {0x1401, 0x166C},
-{0x166F, 0x167F}, {0x1681, 0x169A}, {0x16A0, 0x16EA}, {0x16F1, 0x16F8}, {0x1700, 0x170C}, {0x170E, 0x1711}, {0x1720, 0x1731}, {0x1740, 0x1751}, {0x1760, 0x176C}, {0x176E, 0x1770}, {0x1780, 0x17B3},
-{0x17D7, 0x17D7}, {0x17DC, 0x17DC}, {0x1820, 0x1878}, {0x1880, 0x1884}, {0x1887, 0x18A8}, {0x18AA, 0x18AA}, {0x18B0, 0x18F5}, {0x1900, 0x191E}, {0x1950, 0x196D}, {0x1970, 0x1974}, {0x1980, 0x19AB},
-{0x19B0, 0x19C9}, {0x1A00, 0x1A16}, {0x1A20, 0x1A54}, {0x1AA7, 0x1AA7}, {0x1B05, 0x1B33}, {0x1B45, 0x1B4B}, {0x1B83, 0x1BA0}, {0x1BAE, 0x1BAF}, {0x1BBA, 0x1BE5}, {0x1C00, 0x1C23}, {0x1C4D, 0x1C4F},
-{0x1C5A, 0x1C7D}, {0x1C80, 0x1C88}, {0x1C90, 0x1CBA}, {0x1CBD, 0x1CBF}, {0x1CE9, 0x1CEC}, {0x1CEE, 0x1CF3}, {0x1CF5, 0x1CF6}, {0x1CFA, 0x1CFA}, {0x1D00, 0x1DBF}, {0x1E00, 0x1F15}, {0x1F18, 0x1F1D},
-{0x1F20, 0x1F45}, {0x1F48, 0x1F4D}, {0x1F50, 0x1F57}, {0x1F59, 0x1F59}, {0x1F5B, 0x1F5B}, {0x1F5D, 0x1F5D}, {0x1F5F, 0x1F7D}, {0x1F80, 0x1FB4}, {0x1FB6, 0x1FBC}, {0x1FBE, 0x1FBE}, {0x1FC2, 0x1FC4},
-{0x1FC6, 0x1FCC}, {0x1FD0, 0x1FD3}, {0x1FD6, 0x1FDB}, {0x1FE0, 0x1FEC}, {0x1FF2, 0x1FF4}, {0x1FF6, 0x1FFC}, {0x2071, 0x2071}, {0x207F, 0x207F}, {0x2090, 0x209C}, {0x2102, 0x2102}, {0x2107, 0x2107},
-{0x210A, 0x2113}, {0x2115, 0x2115}, {0x2119, 0x211D}, {0x2124, 0x2124}, {0x2126, 0x2126}, {0x2128, 0x2128}, {0x212A, 0x212D}, {0x212F, 0x2139}, {0x213C, 0x213F}, {0x2145, 0x2149}, {0x214E, 0x214E},
-{0x2183, 0x2184}, {0x2C00, 0x2C2E}, {0x2C30, 0x2C5E}, {0x2C60, 0x2CE4}, {0x2CEB, 0x2CEE}, {0x2CF2, 0x2CF3}, {0x2D00, 0x2D25}, {0x2D27, 0x2D27}, {0x2D2D, 0x2D2D}, {0x2D30, 0x2D67}, {0x2D6F, 0x2D6F},
-{0x2D80, 0x2D96}, {0x2DA0, 0x2DA6}, {0x2DA8, 0x2DAE}, {0x2DB0, 0x2DB6}, {0x2DB8, 0x2DBE}, {0x2DC0, 0x2DC6}, {0x2DC8, 0x2DCE}, {0x2DD0, 0x2DD6}, {0x2DD8, 0x2DDE}, {0x2E2F, 0x2E2F}, {0x3005, 0x3006},
-{0x3031, 0x3035}, {0x303B, 0x303C}, {0x3041, 0x3096}, {0x309D, 0x309F}, {0x30A1, 0x30FA}, {0x30FC, 0x30FF}, {0x3105, 0x312F}, {0x3131, 0x318E}, {0x31A0, 0x31BF}, {0x31F0, 0x31FF}, {0x3400, 0x4DBF},
-{0x4E00, 0x9FFC}, {0xA000, 0xA48C}, {0xA4D0, 0xA4FD}, {0xA500, 0xA60C}, {0xA610, 0xA61F}, {0xA62A, 0xA62B}, {0xA640, 0xA66E}, {0xA67F, 0xA69D}, {0xA6A0, 0xA6E5}, {0xA717, 0xA71F}, {0xA722, 0xA788},
-{0xA78B, 0xA7BF}, {0xA7C2, 0xA7CA}, {0xA7F5, 0xA801}, {0xA803, 0xA805}, {0xA807, 0xA80A}, {0xA80C, 0xA822}, {0xA840, 0xA873}, {0xA882, 0xA8B3}, {0xA8F2, 0xA8F7}, {0xA8FB, 0xA8FB}, {0xA8FD, 0xA8FE},
-{0xA90A, 0xA925}, {0xA930, 0xA946}, {0xA960, 0xA97C}, {0xA984, 0xA9B2}, {0xA9CF, 0xA9CF}, {0xA9E0, 0xA9E4}, {0xA9E6, 0xA9EF}, {0xA9FA, 0xA9FE}, {0xAA00, 0xAA28}, {0xAA40, 0xAA42}, {0xAA44, 0xAA4B},
-{0xAA60, 0xAA76}, {0xAA7A, 0xAA7A}, {0xAA7E, 0xAAAF}, {0xAAB1, 0xAAB1}, {0xAAB5, 0xAAB6}, {0xAAB9, 0xAABD}, {0xAAC0, 0xAAC0}, {0xAAC2, 0xAAC2}, {0xAADB, 0xAADD}, {0xAAE0, 0xAAEA}, {0xAAF2, 0xAAF4},
-{0xAB01, 0xAB06}, {0xAB09, 0xAB0E}, {0xAB11, 0xAB16}, {0xAB20, 0xAB26}, {0xAB28, 0xAB2E}, {0xAB30, 0xAB5A}, {0xAB5C, 0xAB69}, {0xAB70, 0xABE2}, {0xAC00, 0xD7A3}, {0xD7B0, 0xD7C6}, {0xD7CB, 0xD7FB},
-{0xF900, 0xFA6D}, {0xFA70, 0xFAD9}, {0xFB00, 0xFB06}, {0xFB13, 0xFB17}, {0xFB1D, 0xFB1D}, {0xFB1F, 0xFB28}, {0xFB2A, 0xFB36}, {0xFB38, 0xFB3C}, {0xFB3E, 0xFB3E}, {0xFB40, 0xFB41}, {0xFB43, 0xFB44},
-{0xFB46, 0xFBB1}, {0xFBD3, 0xFD3D}, {0xFD50, 0xFD8F}, {0xFD92, 0xFDC7}, {0xFDF0, 0xFDFB}, {0xFE70, 0xFE74}, {0xFE76, 0xFEFC}, {0xFF21, 0xFF3A}, {0xFF41, 0xFF5A}, {0xFF66, 0xFFBE}, {0xFFC2, 0xFFC7},
-{0xFFCA, 0xFFCF}, {0xFFD2, 0xFFD7}, {0xFFDA, 0xFFDC}, {0x10000, 0x1000B}, {0x1000D, 0x10026}, {0x10028, 0x1003A}, {0x1003C, 0x1003D}, {0x1003F, 0x1004D}, {0x10050, 0x1005D}, {0x10080, 0x100FA},
-{0x10280, 0x1029C}, {0x102A0, 0x102D0}, {0x10300, 0x1031F}, {0x1032D, 0x10340}, {0x10342, 0x10349}, {0x10350, 0x10375}, {0x10380, 0x1039D}, {0x103A0, 0x103C3}, {0x103C8, 0x103CF}, {0x10400, 0x1049D},
-{0x104B0, 0x104D3}, {0x104D8, 0x104FB}, {0x10500, 0x10527}, {0x10530, 0x10563}, {0x10600, 0x10736}, {0x10740, 0x10755}, {0x10760, 0x10767}, {0x10800, 0x10805}, {0x10808, 0x10808}, {0x1080A, 0x10835},
-{0x10837, 0x10838}, {0x1083C, 0x1083C}, {0x1083F, 0x10855}, {0x10860, 0x10876}, {0x10880, 0x1089E}, {0x108E0, 0x108F2}, {0x108F4, 0x108F5}, {0x10900, 0x10915}, {0x10920, 0x10939}, {0x10980, 0x109B7},
-{0x109BE, 0x109BF}, {0x10A00, 0x10A00}, {0x10A10, 0x10A13}, {0x10A15, 0x10A17}, {0x10A19, 0x10A35}, {0x10A60, 0x10A7C}, {0x10A80, 0x10A9C}, {0x10AC0, 0x10AC7}, {0x10AC9, 0x10AE4}, {0x10B00, 0x10B35},
-{0x10B40, 0x10B55}, {0x10B60, 0x10B72}, {0x10B80, 0x10B91}, {0x10C00, 0x10C48}, {0x10C80, 0x10CB2}, {0x10CC0, 0x10CF2}, {0x10D00, 0x10D23}, {0x10E80, 0x10EA9}, {0x10EB0, 0x10EB1}, {0x10F00, 0x10F1C},
-{0x10F27, 0x10F27}, {0x10F30, 0x10F45}, {0x10FB0, 0x10FC4}, {0x10FE0, 0x10FF6}, {0x11003, 0x11037}, {0x11083, 0x110AF}, {0x110D0, 0x110E8}, {0x11103, 0x11126}, {0x11144, 0x11144}, {0x11147, 0x11147},
-{0x11150, 0x11172}, {0x11176, 0x11176}, {0x11183, 0x111B2}, {0x111C1, 0x111C4}, {0x111DA, 0x111DA}, {0x111DC, 0x111DC}, {0x11200, 0x11211}, {0x11213, 0x1122B}, {0x11280, 0x11286}, {0x11288, 0x11288},
-{0x1128A, 0x1128D}, {0x1128F, 0x1129D}, {0x1129F, 0x112A8}, {0x112B0, 0x112DE}, {0x11305, 0x1130C}, {0x1130F, 0x11310}, {0x11313, 0x11328}, {0x1132A, 0x11330}, {0x11332, 0x11333}, {0x11335, 0x11339},
-{0x1133D, 0x1133D}, {0x11350, 0x11350}, {0x1135D, 0x11361}, {0x11400, 0x11434}, {0x11447, 0x1144A}, {0x1145F, 0x11461}, {0x11480, 0x114AF}, {0x114C4, 0x114C5}, {0x114C7, 0x114C7}, {0x11580, 0x115AE},
-{0x115D8, 0x115DB}, {0x11600, 0x1162F}, {0x11644, 0x11644}, {0x11680, 0x116AA}, {0x116B8, 0x116B8}, {0x11700, 0x1171A}, {0x11800, 0x1182B}, {0x118A0, 0x118DF}, {0x118FF, 0x11906}, {0x11909, 0x11909},
-{0x1190C, 0x11913}, {0x11915, 0x11916}, {0x11918, 0x1192F}, {0x1193F, 0x1193F}, {0x11941, 0x11941}, {0x119A0, 0x119A7}, {0x119AA, 0x119D0}, {0x119E1, 0x119E1}, {0x119E3, 0x119E3}, {0x11A00, 0x11A00},
-{0x11A0B, 0x11A32}, {0x11A3A, 0x11A3A}, {0x11A50, 0x11A50}, {0x11A5C, 0x11A89}, {0x11A9D, 0x11A9D}, {0x11AC0, 0x11AF8}, {0x11C00, 0x11C08}, {0x11C0A, 0x11C2E}, {0x11C40, 0x11C40}, {0x11C72, 0x11C8F},
-{0x11D00, 0x11D06}, {0x11D08, 0x11D09}, {0x11D0B, 0x11D30}, {0x11D46, 0x11D46}, {0x11D60, 0x11D65}, {0x11D67, 0x11D68}, {0x11D6A, 0x11D89}, {0x11D98, 0x11D98}, {0x11EE0, 0x11EF2}, {0x11FB0, 0x11FB0},
-{0x12000, 0x12399}, {0x12480, 0x12543}, {0x13000, 0x1342E}, {0x14400, 0x14646}, {0x16800, 0x16A38}, {0x16A40, 0x16A5E}, {0x16AD0, 0x16AED}, {0x16B00, 0x16B2F}, {0x16B40, 0x16B43}, {0x16B63, 0x16B77},
-{0x16B7D, 0x16B8F}, {0x16E40, 0x16E7F}, {0x16F00, 0x16F4A}, {0x16F50, 0x16F50}, {0x16F93, 0x16F9F}, {0x16FE0, 0x16FE1}, {0x16FE3, 0x16FE3}, {0x17000, 0x187F7}, {0x18800, 0x18CD5}, {0x18D00, 0x18D08},
-{0x1B000, 0x1B11E}, {0x1B150, 0x1B152}, {0x1B164, 0x1B167}, {0x1B170, 0x1B2FB}, {0x1BC00, 0x1BC6A}, {0x1BC70, 0x1BC7C}, {0x1BC80, 0x1BC88}, {0x1BC90, 0x1BC99}, {0x1D400, 0x1D454}, {0x1D456, 0x1D49C},
-{0x1D49E, 0x1D49F}, {0x1D4A2, 0x1D4A2}, {0x1D4A5, 0x1D4A6}, {0x1D4A9, 0x1D4AC}, {0x1D4AE, 0x1D4B9}, {0x1D4BB, 0x1D4BB}, {0x1D4BD, 0x1D4C3}, {0x1D4C5, 0x1D505}, {0x1D507, 0x1D50A}, {0x1D50D, 0x1D514},
-{0x1D516, 0x1D51C}, {0x1D51E, 0x1D539}, {0x1D53B, 0x1D53E}, {0x1D540, 0x1D544}, {0x1D546, 0x1D546}, {0x1D54A, 0x1D550}, {0x1D552, 0x1D6A5}, {0x1D6A8, 0x1D6C0}, {0x1D6C2, 0x1D6DA}, {0x1D6DC, 0x1D6FA},
-{0x1D6FC, 0x1D714}, {0x1D716, 0x1D734}, {0x1D736, 0x1D74E}, {0x1D750, 0x1D76E}, {0x1D770, 0x1D788}, {0x1D78A, 0x1D7A8}, {0x1D7AA, 0x1D7C2}, {0x1D7C4, 0x1D7CB}, {0x1E100, 0x1E12C}, {0x1E137, 0x1E13D},
-{0x1E14E, 0x1E14E}, {0x1E2C0, 0x1E2EB}, {0x1E800, 0x1E8C4}, {0x1E900, 0x1E943}, {0x1E94B, 0x1E94B}, {0x1EE00, 0x1EE03}, {0x1EE05, 0x1EE1F}, {0x1EE21, 0x1EE22}, {0x1EE24, 0x1EE24}, {0x1EE27, 0x1EE27},
-{0x1EE29, 0x1EE32}, {0x1EE34, 0x1EE37}, {0x1EE39, 0x1EE39}, {0x1EE3B, 0x1EE3B}, {0x1EE42, 0x1EE42}, {0x1EE47, 0x1EE47}, {0x1EE49, 0x1EE49}, {0x1EE4B, 0x1EE4B}, {0x1EE4D, 0x1EE4F}, {0x1EE51, 0x1EE52},
-{0x1EE54, 0x1EE54}, {0x1EE57, 0x1EE57}, {0x1EE59, 0x1EE59}, {0x1EE5B, 0x1EE5B}, {0x1EE5D, 0x1EE5D}, {0x1EE5F, 0x1EE5F}, {0x1EE61, 0x1EE62}, {0x1EE64, 0x1EE64}, {0x1EE67, 0x1EE6A}, {0x1EE6C, 0x1EE72},
-{0x1EE74, 0x1EE77}, {0x1EE79, 0x1EE7C}, {0x1EE7E, 0x1EE7E}, {0x1EE80, 0x1EE89}, {0x1EE8B, 0x1EE9B}, {0x1EEA1, 0x1EEA3}, {0x1EEA5, 0x1EEA9}, {0x1EEAB, 0x1EEBB}, {0x20000, 0x2A6DD}, {0x2A700, 0x2B734},
-{0x2B740, 0x2B81D}, {0x2B820, 0x2CEA1}, {0x2CEB0, 0x2EBE0}, {0x2F800, 0x2FA1D}, {0x30000, 0x3134A},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> whitespace_ranges = {
-{0x9, 0xD}, {0x1C, 0x20}, {0x85, 0x85}, {0xA0, 0xA0}, {0x1680, 0x1680}, {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> accent_mark_ranges = {
-{0x300, 0x36F}, {0x483, 0x489}, {0x591, 0x5BD}, {0x5BF, 0x5BF}, {0x5C1, 0x5C2}, {0x5C4, 0x5C5}, {0x5C7, 0x5C7}, {0x610, 0x61A}, {0x64B, 0x65F}, {0x670, 0x670}, {0x6D6, 0x6DC}, {0x6DF, 0x6E4},
-{0x6E7, 0x6E8}, {0x6EA, 0x6ED}, {0x711, 0x711}, {0x730, 0x74A}, {0x7A6, 0x7B0}, {0x7EB, 0x7F3}, {0x7FD, 0x7FD}, {0x816, 0x819}, {0x81B, 0x823}, {0x825, 0x827}, {0x829, 0x82D}, {0x859, 0x85B},
-{0x8D3, 0x8E1}, {0x8E3, 0x903}, {0x93A, 0x93C}, {0x93E, 0x94F}, {0x951, 0x957}, {0x962, 0x963}, {0x981, 0x983}, {0x9BC, 0x9BC}, {0x9BE, 0x9C4}, {0x9C7, 0x9C8}, {0x9CB, 0x9CD}, {0x9D7, 0x9D7},
-{0x9E2, 0x9E3}, {0x9FE, 0x9FE}, {0xA01, 0xA03}, {0xA3C, 0xA3C}, {0xA3E, 0xA42}, {0xA47, 0xA48}, {0xA4B, 0xA4D}, {0xA51, 0xA51}, {0xA70, 0xA71}, {0xA75, 0xA75}, {0xA81, 0xA83}, {0xABC, 0xABC},
-{0xABE, 0xAC5}, {0xAC7, 0xAC9}, {0xACB, 0xACD}, {0xAE2, 0xAE3}, {0xAFA, 0xAFF}, {0xB01, 0xB03}, {0xB3C, 0xB3C}, {0xB3E, 0xB44}, {0xB47, 0xB48}, {0xB4B, 0xB4D}, {0xB55, 0xB57}, {0xB62, 0xB63},
-{0xB82, 0xB82}, {0xBBE, 0xBC2}, {0xBC6, 0xBC8}, {0xBCA, 0xBCD}, {0xBD7, 0xBD7}, {0xC00, 0xC04}, {0xC3E, 0xC44}, {0xC46, 0xC48}, {0xC4A, 0xC4D}, {0xC55, 0xC56}, {0xC62, 0xC63}, {0xC81, 0xC83},
-{0xCBC, 0xCBC}, {0xCBE, 0xCC4}, {0xCC6, 0xCC8}, {0xCCA, 0xCCD}, {0xCD5, 0xCD6}, {0xCE2, 0xCE3}, {0xD00, 0xD03}, {0xD3B, 0xD3C}, {0xD3E, 0xD44}, {0xD46, 0xD48}, {0xD4A, 0xD4D}, {0xD57, 0xD57},
-{0xD62, 0xD63}, {0xD81, 0xD83}, {0xDCA, 0xDCA}, {0xDCF, 0xDD4}, {0xDD6, 0xDD6}, {0xDD8, 0xDDF}, {0xDF2, 0xDF3}, {0xE31, 0xE31}, {0xE34, 0xE3A}, {0xE47, 0xE4E}, {0xEB1, 0xEB1}, {0xEB4, 0xEBC},
-{0xEC8, 0xECD}, {0xF18, 0xF19}, {0xF35, 0xF35}, {0xF37, 0xF37}, {0xF39, 0xF39}, {0xF3E, 0xF3F}, {0xF71, 0xF84}, {0xF86, 0xF87}, {0xF8D, 0xF97}, {0xF99, 0xFBC}, {0xFC6, 0xFC6}, {0x102B, 0x103E},
-{0x1056, 0x1059}, {0x105E, 0x1060}, {0x1062, 0x1064}, {0x1067, 0x106D}, {0x1071, 0x1074}, {0x1082, 0x108D}, {0x108F, 0x108F}, {0x109A, 0x109D}, {0x135D, 0x135F}, {0x1712, 0x1714}, {0x1732, 0x1734},
-{0x1752, 0x1753}, {0x1772, 0x1773}, {0x17B4, 0x17D3}, {0x17DD, 0x17DD}, {0x180B, 0x180D}, {0x1885, 0x1886}, {0x18A9, 0x18A9}, {0x1920, 0x192B}, {0x1930, 0x193B}, {0x1A17, 0x1A1B}, {0x1A55, 0x1A5E},
-{0x1A60, 0x1A7C}, {0x1A7F, 0x1A7F}, {0x1AB0, 0x1AC0}, {0x1B00, 0x1B04}, {0x1B34, 0x1B44}, {0x1B6B, 0x1B73}, {0x1B80, 0x1B82}, {0x1BA1, 0x1BAD}, {0x1BE6, 0x1BF3}, {0x1C24, 0x1C37}, {0x1CD0, 0x1CD2},
-{0x1CD4, 0x1CE8}, {0x1CED, 0x1CED}, {0x1CF4, 0x1CF4}, {0x1CF7, 0x1CF9}, {0x1DC0, 0x1DF9}, {0x1DFB, 0x1DFF}, {0x20D0, 0x20F0}, {0x2CEF, 0x2CF1}, {0x2D7F, 0x2D7F}, {0x2DE0, 0x2DFF}, {0x302A, 0x302F},
-{0x3099, 0x309A}, {0xA66F, 0xA672}, {0xA674, 0xA67D}, {0xA69E, 0xA69F}, {0xA6F0, 0xA6F1}, {0xA802, 0xA802}, {0xA806, 0xA806}, {0xA80B, 0xA80B}, {0xA823, 0xA827}, {0xA82C, 0xA82C}, {0xA880, 0xA881},
-{0xA8B4, 0xA8C5}, {0xA8E0, 0xA8F1}, {0xA8FF, 0xA8FF}, {0xA926, 0xA92D}, {0xA947, 0xA953}, {0xA980, 0xA983}, {0xA9B3, 0xA9C0}, {0xA9E5, 0xA9E5}, {0xAA29, 0xAA36}, {0xAA43, 0xAA43}, {0xAA4C, 0xAA4D},
-{0xAA7B, 0xAA7D}, {0xAAB0, 0xAAB0}, {0xAAB2, 0xAAB4}, {0xAAB7, 0xAAB8}, {0xAABE, 0xAABF}, {0xAAC1, 0xAAC1}, {0xAAEB, 0xAAEF}, {0xAAF5, 0xAAF6}, {0xABE3, 0xABEA}, {0xABEC, 0xABED}, {0xFB1E, 0xFB1E},
-{0xFE00, 0xFE0F}, {0xFE20, 0xFE2F}, {0x101FD, 0x101FD}, {0x102E0, 0x102E0}, {0x10376, 0x1037A}, {0x10A01, 0x10A03}, {0x10A05, 0x10A06}, {0x10A0C, 0x10A0F}, {0x10A38, 0x10A3A}, {0x10A3F, 0x10A3F},
-{0x10AE5, 0x10AE6}, {0x10D24, 0x10D27}, {0x10EAB, 0x10EAC}, {0x10F46, 0x10F50}, {0x11000, 0x11002}, {0x11038, 0x11046}, {0x1107F, 0x11082}, {0x110B0, 0x110BA}, {0x11100, 0x11102}, {0x11127, 0x11134},
-{0x11145, 0x11146}, {0x11173, 0x11173}, {0x11180, 0x11182}, {0x111B3, 0x111C0}, {0x111C9, 0x111CC}, {0x111CE, 0x111CF}, {0x1122C, 0x11237}, {0x1123E, 0x1123E}, {0x112DF, 0x112EA}, {0x11300, 0x11303},
-{0x1133B, 0x1133C}, {0x1133E, 0x11344}, {0x11347, 0x11348}, {0x1134B, 0x1134D}, {0x11357, 0x11357}, {0x11362, 0x11363}, {0x11366, 0x1136C}, {0x11370, 0x11374}, {0x11435, 0x11446}, {0x1145E, 0x1145E},
-{0x114B0, 0x114C3}, {0x115AF, 0x115B5}, {0x115B8, 0x115C0}, {0x115DC, 0x115DD}, {0x11630, 0x11640}, {0x116AB, 0x116B7}, {0x1171D, 0x1172B}, {0x1182C, 0x1183A}, {0x11930, 0x11935}, {0x11937, 0x11938},
-{0x1193B, 0x1193E}, {0x11940, 0x11940}, {0x11942, 0x11943}, {0x119D1, 0x119D7}, {0x119DA, 0x119E0}, {0x119E4, 0x119E4}, {0x11A01, 0x11A0A}, {0x11A33, 0x11A39}, {0x11A3B, 0x11A3E}, {0x11A47, 0x11A47},
-{0x11A51, 0x11A5B}, {0x11A8A, 0x11A99}, {0x11C2F, 0x11C36}, {0x11C38, 0x11C3F}, {0x11C92, 0x11CA7}, {0x11CA9, 0x11CB6}, {0x11D31, 0x11D36}, {0x11D3A, 0x11D3A}, {0x11D3C, 0x11D3D}, {0x11D3F, 0x11D45},
-{0x11D47, 0x11D47}, {0x11D8A, 0x11D8E}, {0x11D90, 0x11D91}, {0x11D93, 0x11D97}, {0x11EF3, 0x11EF6}, {0x16AF0, 0x16AF4}, {0x16B30, 0x16B36}, {0x16F4F, 0x16F4F}, {0x16F51, 0x16F87}, {0x16F8F, 0x16F92},
-{0x16FE4, 0x16FE4}, {0x16FF0, 0x16FF1}, {0x1BC9D, 0x1BC9E}, {0x1D165, 0x1D169}, {0x1D16D, 0x1D172}, {0x1D17B, 0x1D182}, {0x1D185, 0x1D18B}, {0x1D1AA, 0x1D1AD}, {0x1D242, 0x1D244}, {0x1DA00, 0x1DA36},
-{0x1DA3B, 0x1DA6C}, {0x1DA75, 0x1DA75}, {0x1DA84, 0x1DA84}, {0x1DA9B, 0x1DA9F}, {0x1DAA1, 0x1DAAF}, {0x1E000, 0x1E006}, {0x1E008, 0x1E018}, {0x1E01B, 0x1E021}, {0x1E023, 0x1E024}, {0x1E026, 0x1E02A},
-{0x1E130, 0x1E136}, {0x1E2EC, 0x1E2EF}, {0x1E8D0, 0x1E8D6}, {0x1E944, 0x1E94A}, {0xE0100, 0xE01EF},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> punctuation_ranges = {
-{0x21, 0x23}, {0x25, 0x2A}, {0x2C, 0x2F}, {0x3A, 0x3B}, {0x3F, 0x40}, {0x5B, 0x5D}, {0x5F, 0x5F}, {0x7B, 0x7B}, {0x7D, 0x7D}, {0xA1, 0xA1}, {0xA7, 0xA7}, {0xAB, 0xAB}, {0xB6, 0xB7}, {0xBB, 0xBB},
-{0xBF, 0xBF}, {0x37E, 0x37E}, {0x387, 0x387}, {0x55A, 0x55F}, {0x589, 0x58A}, {0x5BE, 0x5BE}, {0x5C0, 0x5C0}, {0x5C3, 0x5C3}, {0x5C6, 0x5C6}, {0x5F3, 0x5F4}, {0x609, 0x60A}, {0x60C, 0x60D},
-{0x61B, 0x61B}, {0x61E, 0x61F}, {0x66A, 0x66D}, {0x6D4, 0x6D4}, {0x700, 0x70D}, {0x7F7, 0x7F9}, {0x830, 0x83E}, {0x85E, 0x85E}, {0x964, 0x965}, {0x970, 0x970}, {0x9FD, 0x9FD}, {0xA76, 0xA76},
-{0xAF0, 0xAF0}, {0xC77, 0xC77}, {0xC84, 0xC84}, {0xDF4, 0xDF4}, {0xE4F, 0xE4F}, {0xE5A, 0xE5B}, {0xF04, 0xF12}, {0xF14, 0xF14}, {0xF3A, 0xF3D}, {0xF85, 0xF85}, {0xFD0, 0xFD4}, {0xFD9, 0xFDA},
-{0x104A, 0x104F}, {0x10FB, 0x10FB}, {0x1360, 0x1368}, {0x1400, 0x1400}, {0x166E, 0x166E}, {0x169B, 0x169C}, {0x16EB, 0x16ED}, {0x1735, 0x1736}, {0x17D4, 0x17D6}, {0x17D8, 0x17DA}, {0x1800, 0x180A},
-{0x1944, 0x1945}, {0x1A1E, 0x1A1F}, {0x1AA0, 0x1AA6}, {0x1AA8, 0x1AAD}, {0x1B5A, 0x1B60}, {0x1BFC, 0x1BFF}, {0x1C3B, 0x1C3F}, {0x1C7E, 0x1C7F}, {0x1CC0, 0x1CC7}, {0x1CD3, 0x1CD3}, {0x2010, 0x2027},
-{0x2030, 0x2043}, {0x2045, 0x2051}, {0x2053, 0x205E}, {0x207D, 0x207E}, {0x208D, 0x208E}, {0x2308, 0x230B}, {0x2329, 0x232A}, {0x2768, 0x2775}, {0x27C5, 0x27C6}, {0x27E6, 0x27EF}, {0x2983, 0x2998},
-{0x29D8, 0x29DB}, {0x29FC, 0x29FD}, {0x2CF9, 0x2CFC}, {0x2CFE, 0x2CFF}, {0x2D70, 0x2D70}, {0x2E00, 0x2E2E}, {0x2E30, 0x2E4F}, {0x2E52, 0x2E52}, {0x3001, 0x3003}, {0x3008, 0x3011}, {0x3014, 0x301F},
-{0x3030, 0x3030}, {0x303D, 0x303D}, {0x30A0, 0x30A0}, {0x30FB, 0x30FB}, {0xA4FE, 0xA4FF}, {0xA60D, 0xA60F}, {0xA673, 0xA673}, {0xA67E, 0xA67E}, {0xA6F2, 0xA6F7}, {0xA874, 0xA877}, {0xA8CE, 0xA8CF},
-{0xA8F8, 0xA8FA}, {0xA8FC, 0xA8FC}, {0xA92E, 0xA92F}, {0xA95F, 0xA95F}, {0xA9C1, 0xA9CD}, {0xA9DE, 0xA9DF}, {0xAA5C, 0xAA5F}, {0xAADE, 0xAADF}, {0xAAF0, 0xAAF1}, {0xABEB, 0xABEB}, {0xFD3E, 0xFD3F},
-{0xFE10, 0xFE19}, {0xFE30, 0xFE52}, {0xFE54, 0xFE61}, {0xFE63, 0xFE63}, {0xFE68, 0xFE68}, {0xFE6A, 0xFE6B}, {0xFF01, 0xFF03}, {0xFF05, 0xFF0A}, {0xFF0C, 0xFF0F}, {0xFF1A, 0xFF1B}, {0xFF1F, 0xFF20},
-{0xFF3B, 0xFF3D}, {0xFF3F, 0xFF3F}, {0xFF5B, 0xFF5B}, {0xFF5D, 0xFF5D}, {0xFF5F, 0xFF65}, {0x10100, 0x10102}, {0x1039F, 0x1039F}, {0x103D0, 0x103D0}, {0x1056F, 0x1056F}, {0x10857, 0x10857},
-{0x1091F, 0x1091F}, {0x1093F, 0x1093F}, {0x10A50, 0x10A58}, {0x10A7F, 0x10A7F}, {0x10AF0, 0x10AF6}, {0x10B39, 0x10B3F}, {0x10B99, 0x10B9C}, {0x10EAD, 0x10EAD}, {0x10F55, 0x10F59}, {0x11047, 0x1104D},
-{0x110BB, 0x110BC}, {0x110BE, 0x110C1}, {0x11140, 0x11143}, {0x11174, 0x11175}, {0x111C5, 0x111C8}, {0x111CD, 0x111CD}, {0x111DB, 0x111DB}, {0x111DD, 0x111DF}, {0x11238, 0x1123D}, {0x112A9, 0x112A9},
-{0x1144B, 0x1144F}, {0x1145A, 0x1145B}, {0x1145D, 0x1145D}, {0x114C6, 0x114C6}, {0x115C1, 0x115D7}, {0x11641, 0x11643}, {0x11660, 0x1166C}, {0x1173C, 0x1173E}, {0x1183B, 0x1183B}, {0x11944, 0x11946},
-{0x119E2, 0x119E2}, {0x11A3F, 0x11A46}, {0x11A9A, 0x11A9C}, {0x11A9E, 0x11AA2}, {0x11C41, 0x11C45}, {0x11C70, 0x11C71}, {0x11EF7, 0x11EF8}, {0x11FFF, 0x11FFF}, {0x12470, 0x12474}, {0x16A6E, 0x16A6F},
-{0x16AF5, 0x16AF5}, {0x16B37, 0x16B3B}, {0x16B44, 0x16B44}, {0x16E97, 0x16E9A}, {0x16FE2, 0x16FE2}, {0x1BC9F, 0x1BC9F}, {0x1DA87, 0x1DA8B}, {0x1E95E, 0x1E95F},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> symbol_ranges = {
-{0x24, 0x24}, {0x2B, 0x2B}, {0x3C, 0x3E}, {0x5E, 0x5E}, {0x60, 0x60}, {0x7C, 0x7C}, {0x7E, 0x7E}, {0xA2, 0xA6}, {0xA8, 0xA9}, {0xAC, 0xAC}, {0xAE, 0xB1}, {0xB4, 0xB4}, {0xB8, 0xB8}, {0xD7, 0xD7},
-{0xF7, 0xF7}, {0x2C2, 0x2C5}, {0x2D2, 0x2DF}, {0x2E5, 0x2EB}, {0x2ED, 0x2ED}, {0x2EF, 0x2FF}, {0x375, 0x375}, {0x384, 0x385}, {0x3F6, 0x3F6}, {0x482, 0x482}, {0x58D, 0x58F}, {0x606, 0x608},
-{0x60B, 0x60B}, {0x60E, 0x60F}, {0x6DE, 0x6DE}, {0x6E9, 0x6E9}, {0x6FD, 0x6FE}, {0x7F6, 0x7F6}, {0x7FE, 0x7FF}, {0x9F2, 0x9F3}, {0x9FA, 0x9FB}, {0xAF1, 0xAF1}, {0xB70, 0xB70}, {0xBF3, 0xBFA},
-{0xC7F, 0xC7F}, {0xD4F, 0xD4F}, {0xD79, 0xD79}, {0xE3F, 0xE3F}, {0xF01, 0xF03}, {0xF13, 0xF13}, {0xF15, 0xF17}, {0xF1A, 0xF1F}, {0xF34, 0xF34}, {0xF36, 0xF36}, {0xF38, 0xF38}, {0xFBE, 0xFC5},
-{0xFC7, 0xFCC}, {0xFCE, 0xFCF}, {0xFD5, 0xFD8}, {0x109E, 0x109F}, {0x1390, 0x1399}, {0x166D, 0x166D}, {0x17DB, 0x17DB}, {0x1940, 0x1940}, {0x19DE, 0x19FF}, {0x1B61, 0x1B6A}, {0x1B74, 0x1B7C},
-{0x1FBD, 0x1FBD}, {0x1FBF, 0x1FC1}, {0x1FCD, 0x1FCF}, {0x1FDD, 0x1FDF}, {0x1FED, 0x1FEF}, {0x1FFD, 0x1FFE}, {0x2044, 0x2044}, {0x2052, 0x2052}, {0x207A, 0x207C}, {0x208A, 0x208C}, {0x20A0, 0x20BF},
-{0x2100, 0x2101}, {0x2103, 0x2106}, {0x2108, 0x2109}, {0x2114, 0x2114}, {0x2116, 0x2118}, {0x211E, 0x2123}, {0x2125, 0x2125}, {0x2127, 0x2127}, {0x2129, 0x2129}, {0x212E, 0x212E}, {0x213A, 0x213B},
-{0x2140, 0x2144}, {0x214A, 0x214D}, {0x214F, 0x214F}, {0x218A, 0x218B}, {0x2190, 0x2307}, {0x230C, 0x2328}, {0x232B, 0x2426}, {0x2440, 0x244A}, {0x249C, 0x24E9}, {0x2500, 0x2767}, {0x2794, 0x27C4},
-{0x27C7, 0x27E5}, {0x27F0, 0x2982}, {0x2999, 0x29D7}, {0x29DC, 0x29FB}, {0x29FE, 0x2B73}, {0x2B76, 0x2B95}, {0x2B97, 0x2BFF}, {0x2CE5, 0x2CEA}, {0x2E50, 0x2E51}, {0x2E80, 0x2E99}, {0x2E9B, 0x2EF3},
-{0x2F00, 0x2FD5}, {0x2FF0, 0x2FFB}, {0x3004, 0x3004}, {0x3012, 0x3013}, {0x3020, 0x3020}, {0x3036, 0x3037}, {0x303E, 0x303F}, {0x309B, 0x309C}, {0x3190, 0x3191}, {0x3196, 0x319F}, {0x31C0, 0x31E3},
-{0x3200, 0x321E}, {0x322A, 0x3247}, {0x3250, 0x3250}, {0x3260, 0x327F}, {0x328A, 0x32B0}, {0x32C0, 0x33FF}, {0x4DC0, 0x4DFF}, {0xA490, 0xA4C6}, {0xA700, 0xA716}, {0xA720, 0xA721}, {0xA789, 0xA78A},
-{0xA828, 0xA82B}, {0xA836, 0xA839}, {0xAA77, 0xAA79}, {0xAB5B, 0xAB5B}, {0xAB6A, 0xAB6B}, {0xFB29, 0xFB29}, {0xFBB2, 0xFBC1}, {0xFDFC, 0xFDFD}, {0xFE62, 0xFE62}, {0xFE64, 0xFE66}, {0xFE69, 0xFE69},
-{0xFF04, 0xFF04}, {0xFF0B, 0xFF0B}, {0xFF1C, 0xFF1E}, {0xFF3E, 0xFF3E}, {0xFF40, 0xFF40}, {0xFF5C, 0xFF5C}, {0xFF5E, 0xFF5E}, {0xFFE0, 0xFFE6}, {0xFFE8, 0xFFEE}, {0xFFFC, 0xFFFD}, {0x10137, 0x1013F},
-{0x10179, 0x10189}, {0x1018C, 0x1018E}, {0x10190, 0x1019C}, {0x101A0, 0x101A0}, {0x101D0, 0x101FC}, {0x10877, 0x10878}, {0x10AC8, 0x10AC8}, {0x1173F, 0x1173F}, {0x11FD5, 0x11FF1}, {0x16B3C, 0x16B3F},
-{0x16B45, 0x16B45}, {0x1BC9C, 0x1BC9C}, {0x1D000, 0x1D0F5}, {0x1D100, 0x1D126}, {0x1D129, 0x1D164}, {0x1D16A, 0x1D16C}, {0x1D183, 0x1D184}, {0x1D18C, 0x1D1A9}, {0x1D1AE, 0x1D1E8}, {0x1D200, 0x1D241},
-{0x1D245, 0x1D245}, {0x1D300, 0x1D356}, {0x1D6C1, 0x1D6C1}, {0x1D6DB, 0x1D6DB}, {0x1D6FB, 0x1D6FB}, {0x1D715, 0x1D715}, {0x1D735, 0x1D735}, {0x1D74F, 0x1D74F}, {0x1D76F, 0x1D76F}, {0x1D789, 0x1D789},
-{0x1D7A9, 0x1D7A9}, {0x1D7C3, 0x1D7C3}, {0x1D800, 0x1D9FF}, {0x1DA37, 0x1DA3A}, {0x1DA6D, 0x1DA74}, {0x1DA76, 0x1DA83}, {0x1DA85, 0x1DA86}, {0x1E14F, 0x1E14F}, {0x1E2FF, 0x1E2FF}, {0x1ECAC, 0x1ECAC},
-{0x1ECB0, 0x1ECB0}, {0x1ED2E, 0x1ED2E}, {0x1EEF0, 0x1EEF1}, {0x1F000, 0x1F02B}, {0x1F030, 0x1F093}, {0x1F0A0, 0x1F0AE}, {0x1F0B1, 0x1F0BF}, {0x1F0C1, 0x1F0CF}, {0x1F0D1, 0x1F0F5}, {0x1F10D, 0x1F1AD},
-{0x1F1E6, 0x1F202}, {0x1F210, 0x1F23B}, {0x1F240, 0x1F248}, {0x1F250, 0x1F251}, {0x1F260, 0x1F265}, {0x1F300, 0x1F6D7}, {0x1F6E0, 0x1F6EC}, {0x1F6F0, 0x1F6FC}, {0x1F700, 0x1F773}, {0x1F780, 0x1F7D8},
-{0x1F7E0, 0x1F7EB}, {0x1F800, 0x1F80B}, {0x1F810, 0x1F847}, {0x1F850, 0x1F859}, {0x1F860, 0x1F887}, {0x1F890, 0x1F8AD}, {0x1F8B0, 0x1F8B1}, {0x1F900, 0x1F978}, {0x1F97A, 0x1F9CB}, {0x1F9CD, 0x1FA53},
-{0x1FA60, 0x1FA6D}, {0x1FA70, 0x1FA74}, {0x1FA78, 0x1FA7A}, {0x1FA80, 0x1FA86}, {0x1FA90, 0x1FAA8}, {0x1FAB0, 0x1FAB6}, {0x1FAC0, 0x1FAC2}, {0x1FAD0, 0x1FAD6}, {0x1FB00, 0x1FB92}, {0x1FB94, 0x1FBCA},
-};
-
-static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
-{0x0, 0x8}, {0xE, 0x1B}, {0x7F, 0x84}, {0x86, 0x9F}, {0xAD, 0xAD}, {0x378, 0x379}, {0x380, 0x383}, {0x38B, 0x38B}, {0x38D, 0x38D}, {0x3A2, 0x3A2}, {0x530, 0x530}, {0x557, 0x558}, {0x58B, 0x58C},
-{0x590, 0x590}, {0x5C8, 0x5CF}, {0x5EB, 0x5EE}, {0x5F5, 0x605}, {0x61C, 0x61D}, {0x6DD, 0x6DD}, {0x70E, 0x70F}, {0x74B, 0x74C}, {0x7B2, 0x7BF}, {0x7FB, 0x7FC}, {0x82E, 0x82F}, {0x83F, 0x83F},
-{0x85C, 0x85D}, {0x85F, 0x85F}, {0x86B, 0x89F}, {0x8B5, 0x8B5}, {0x8C8, 0x8D2}, {0x8E2, 0x8E2}, {0x984, 0x984}, {0x98D, 0x98E}, {0x991, 0x992}, {0x9A9, 0x9A9}, {0x9B1, 0x9B1}, {0x9B3, 0x9B5},
-{0x9BA, 0x9BB}, {0x9C5, 0x9C6}, {0x9C9, 0x9CA}, {0x9CF, 0x9D6}, {0x9D8, 0x9DB}, {0x9DE, 0x9DE}, {0x9E4, 0x9E5}, {0x9FF, 0xA00}, {0xA04, 0xA04}, {0xA0B, 0xA0E}, {0xA11, 0xA12}, {0xA29, 0xA29},
-{0xA31, 0xA31}, {0xA34, 0xA34}, {0xA37, 0xA37}, {0xA3A, 0xA3B}, {0xA3D, 0xA3D}, {0xA43, 0xA46}, {0xA49, 0xA4A}, {0xA4E, 0xA50}, {0xA52, 0xA58}, {0xA5D, 0xA5D}, {0xA5F, 0xA65}, {0xA77, 0xA80},
-{0xA84, 0xA84}, {0xA8E, 0xA8E}, {0xA92, 0xA92}, {0xAA9, 0xAA9}, {0xAB1, 0xAB1}, {0xAB4, 0xAB4}, {0xABA, 0xABB}, {0xAC6, 0xAC6}, {0xACA, 0xACA}, {0xACE, 0xACF}, {0xAD1, 0xADF}, {0xAE4, 0xAE5},
-{0xAF2, 0xAF8}, {0xB00, 0xB00}, {0xB04, 0xB04}, {0xB0D, 0xB0E}, {0xB11, 0xB12}, {0xB29, 0xB29}, {0xB31, 0xB31}, {0xB34, 0xB34}, {0xB3A, 0xB3B}, {0xB45, 0xB46}, {0xB49, 0xB4A}, {0xB4E, 0xB54},
-{0xB58, 0xB5B}, {0xB5E, 0xB5E}, {0xB64, 0xB65}, {0xB78, 0xB81}, {0xB84, 0xB84}, {0xB8B, 0xB8D}, {0xB91, 0xB91}, {0xB96, 0xB98}, {0xB9B, 0xB9B}, {0xB9D, 0xB9D}, {0xBA0, 0xBA2}, {0xBA5, 0xBA7},
-{0xBAB, 0xBAD}, {0xBBA, 0xBBD}, {0xBC3, 0xBC5}, {0xBC9, 0xBC9}, {0xBCE, 0xBCF}, {0xBD1, 0xBD6}, {0xBD8, 0xBE5}, {0xBFB, 0xBFF}, {0xC0D, 0xC0D}, {0xC11, 0xC11}, {0xC29, 0xC29}, {0xC3A, 0xC3C},
-{0xC45, 0xC45}, {0xC49, 0xC49}, {0xC4E, 0xC54}, {0xC57, 0xC57}, {0xC5B, 0xC5F}, {0xC64, 0xC65}, {0xC70, 0xC76}, {0xC8D, 0xC8D}, {0xC91, 0xC91}, {0xCA9, 0xCA9}, {0xCB4, 0xCB4}, {0xCBA, 0xCBB},
-{0xCC5, 0xCC5}, {0xCC9, 0xCC9}, {0xCCE, 0xCD4}, {0xCD7, 0xCDD}, {0xCDF, 0xCDF}, {0xCE4, 0xCE5}, {0xCF0, 0xCF0}, {0xCF3, 0xCFF}, {0xD0D, 0xD0D}, {0xD11, 0xD11}, {0xD45, 0xD45}, {0xD49, 0xD49},
-{0xD50, 0xD53}, {0xD64, 0xD65}, {0xD80, 0xD80}, {0xD84, 0xD84}, {0xD97, 0xD99}, {0xDB2, 0xDB2}, {0xDBC, 0xDBC}, {0xDBE, 0xDBF}, {0xDC7, 0xDC9}, {0xDCB, 0xDCE}, {0xDD5, 0xDD5}, {0xDD7, 0xDD7},
-{0xDE0, 0xDE5}, {0xDF0, 0xDF1}, {0xDF5, 0xE00}, {0xE3B, 0xE3E}, {0xE5C, 0xE80}, {0xE83, 0xE83}, {0xE85, 0xE85}, {0xE8B, 0xE8B}, {0xEA4, 0xEA4}, {0xEA6, 0xEA6}, {0xEBE, 0xEBF}, {0xEC5, 0xEC5},
-{0xEC7, 0xEC7}, {0xECE, 0xECF}, {0xEDA, 0xEDB}, {0xEE0, 0xEFF}, {0xF48, 0xF48}, {0xF6D, 0xF70}, {0xF98, 0xF98}, {0xFBD, 0xFBD}, {0xFCD, 0xFCD}, {0xFDB, 0xFFF}, {0x10C6, 0x10C6}, {0x10C8, 0x10CC},
-{0x10CE, 0x10CF}, {0x1249, 0x1249}, {0x124E, 0x124F}, {0x1257, 0x1257}, {0x1259, 0x1259}, {0x125E, 0x125F}, {0x1289, 0x1289}, {0x128E, 0x128F}, {0x12B1, 0x12B1}, {0x12B6, 0x12B7}, {0x12BF, 0x12BF},
-{0x12C1, 0x12C1}, {0x12C6, 0x12C7}, {0x12D7, 0x12D7}, {0x1311, 0x1311}, {0x1316, 0x1317}, {0x135B, 0x135C}, {0x137D, 0x137F}, {0x139A, 0x139F}, {0x13F6, 0x13F7}, {0x13FE, 0x13FF}, {0x169D, 0x169F},
-{0x16F9, 0x16FF}, {0x170D, 0x170D}, {0x1715, 0x171F}, {0x1737, 0x173F}, {0x1754, 0x175F}, {0x176D, 0x176D}, {0x1771, 0x1771}, {0x1774, 0x177F}, {0x17DE, 0x17DF}, {0x17EA, 0x17EF}, {0x17FA, 0x17FF},
-{0x180E, 0x180F}, {0x181A, 0x181F}, {0x1879, 0x187F}, {0x18AB, 0x18AF}, {0x18F6, 0x18FF}, {0x191F, 0x191F}, {0x192C, 0x192F}, {0x193C, 0x193F}, {0x1941, 0x1943}, {0x196E, 0x196F}, {0x1975, 0x197F},
-{0x19AC, 0x19AF}, {0x19CA, 0x19CF}, {0x19DB, 0x19DD}, {0x1A1C, 0x1A1D}, {0x1A5F, 0x1A5F}, {0x1A7D, 0x1A7E}, {0x1A8A, 0x1A8F}, {0x1A9A, 0x1A9F}, {0x1AAE, 0x1AAF}, {0x1AC1, 0x1AFF}, {0x1B4C, 0x1B4F},
-{0x1B7D, 0x1B7F}, {0x1BF4, 0x1BFB}, {0x1C38, 0x1C3A}, {0x1C4A, 0x1C4C}, {0x1C89, 0x1C8F}, {0x1CBB, 0x1CBC}, {0x1CC8, 0x1CCF}, {0x1CFB, 0x1CFF}, {0x1DFA, 0x1DFA}, {0x1F16, 0x1F17}, {0x1F1E, 0x1F1F},
-{0x1F46, 0x1F47}, {0x1F4E, 0x1F4F}, {0x1F58, 0x1F58}, {0x1F5A, 0x1F5A}, {0x1F5C, 0x1F5C}, {0x1F5E, 0x1F5E}, {0x1F7E, 0x1F7F}, {0x1FB5, 0x1FB5}, {0x1FC5, 0x1FC5}, {0x1FD4, 0x1FD5}, {0x1FDC, 0x1FDC},
-{0x1FF0, 0x1FF1}, {0x1FF5, 0x1FF5}, {0x1FFF, 0x1FFF}, {0x200B, 0x200F}, {0x202A, 0x202E}, {0x2060, 0x206F}, {0x2072, 0x2073}, {0x208F, 0x208F}, {0x209D, 0x209F}, {0x20C0, 0x20CF}, {0x20F1, 0x20FF},
-{0x218C, 0x218F}, {0x2427, 0x243F}, {0x244B, 0x245F}, {0x2B74, 0x2B75}, {0x2B96, 0x2B96}, {0x2C2F, 0x2C2F}, {0x2C5F, 0x2C5F}, {0x2CF4, 0x2CF8}, {0x2D26, 0x2D26}, {0x2D28, 0x2D2C}, {0x2D2E, 0x2D2F},
-{0x2D68, 0x2D6E}, {0x2D71, 0x2D7E}, {0x2D97, 0x2D9F}, {0x2DA7, 0x2DA7}, {0x2DAF, 0x2DAF}, {0x2DB7, 0x2DB7}, {0x2DBF, 0x2DBF}, {0x2DC7, 0x2DC7}, {0x2DCF, 0x2DCF}, {0x2DD7, 0x2DD7}, {0x2DDF, 0x2DDF},
-{0x2E53, 0x2E7F}, {0x2E9A, 0x2E9A}, {0x2EF4, 0x2EFF}, {0x2FD6, 0x2FEF}, {0x2FFC, 0x2FFF}, {0x3040, 0x3040}, {0x3097, 0x3098}, {0x3100, 0x3104}, {0x3130, 0x3130}, {0x318F, 0x318F}, {0x31E4, 0x31EF},
-{0x321F, 0x321F}, {0x9FFD, 0x9FFF}, {0xA48D, 0xA48F}, {0xA4C7, 0xA4CF}, {0xA62C, 0xA63F}, {0xA6F8, 0xA6FF}, {0xA7C0, 0xA7C1}, {0xA7CB, 0xA7F4}, {0xA82D, 0xA82F}, {0xA83A, 0xA83F}, {0xA878, 0xA87F},
-{0xA8C6, 0xA8CD}, {0xA8DA, 0xA8DF}, {0xA954, 0xA95E}, {0xA97D, 0xA97F}, {0xA9CE, 0xA9CE}, {0xA9DA, 0xA9DD}, {0xA9FF, 0xA9FF}, {0xAA37, 0xAA3F}, {0xAA4E, 0xAA4F}, {0xAA5A, 0xAA5B}, {0xAAC3, 0xAADA},
-{0xAAF7, 0xAB00}, {0xAB07, 0xAB08}, {0xAB0F, 0xAB10}, {0xAB17, 0xAB1F}, {0xAB27, 0xAB27}, {0xAB2F, 0xAB2F}, {0xAB6C, 0xAB6F}, {0xABEE, 0xABEF}, {0xABFA, 0xABFF}, {0xD7A4, 0xD7AF}, {0xD7C7, 0xD7CA},
-{0xD7FC, 0xF8FF}, {0xFA6E, 0xFA6F}, {0xFADA, 0xFAFF}, {0xFB07, 0xFB12}, {0xFB18, 0xFB1C}, {0xFB37, 0xFB37}, {0xFB3D, 0xFB3D}, {0xFB3F, 0xFB3F}, {0xFB42, 0xFB42}, {0xFB45, 0xFB45}, {0xFBC2, 0xFBD2},
-{0xFD40, 0xFD4F}, {0xFD90, 0xFD91}, {0xFDC8, 0xFDEF}, {0xFDFE, 0xFDFF}, {0xFE1A, 0xFE1F}, {0xFE53, 0xFE53}, {0xFE67, 0xFE67}, {0xFE6C, 0xFE6F}, {0xFE75, 0xFE75}, {0xFEFD, 0xFF00}, {0xFFBF, 0xFFC1},
-{0xFFC8, 0xFFC9}, {0xFFD0, 0xFFD1}, {0xFFD8, 0xFFD9}, {0xFFDD, 0xFFDF}, {0xFFE7, 0xFFE7}, {0xFFEF, 0xFFFB}, {0xFFFE, 0xFFFF}, {0x1000C, 0x1000C}, {0x10027, 0x10027}, {0x1003B, 0x1003B},
-{0x1003E, 0x1003E}, {0x1004E, 0x1004F}, {0x1005E, 0x1007F}, {0x100FB, 0x100FF}, {0x10103, 0x10106}, {0x10134, 0x10136}, {0x1018F, 0x1018F}, {0x1019D, 0x1019F}, {0x101A1, 0x101CF}, {0x101FE, 0x1027F},
-{0x1029D, 0x1029F}, {0x102D1, 0x102DF}, {0x102FC, 0x102FF}, {0x10324, 0x1032C}, {0x1034B, 0x1034F}, {0x1037B, 0x1037F}, {0x1039E, 0x1039E}, {0x103C4, 0x103C7}, {0x103D6, 0x103FF}, {0x1049E, 0x1049F},
-{0x104AA, 0x104AF}, {0x104D4, 0x104D7}, {0x104FC, 0x104FF}, {0x10528, 0x1052F}, {0x10564, 0x1056E}, {0x10570, 0x105FF}, {0x10737, 0x1073F}, {0x10756, 0x1075F}, {0x10768, 0x107FF}, {0x10806, 0x10807},
-{0x10809, 0x10809}, {0x10836, 0x10836}, {0x10839, 0x1083B}, {0x1083D, 0x1083E}, {0x10856, 0x10856}, {0x1089F, 0x108A6}, {0x108B0, 0x108DF}, {0x108F3, 0x108F3}, {0x108F6, 0x108FA}, {0x1091C, 0x1091E},
-{0x1093A, 0x1093E}, {0x10940, 0x1097F}, {0x109B8, 0x109BB}, {0x109D0, 0x109D1}, {0x10A04, 0x10A04}, {0x10A07, 0x10A0B}, {0x10A14, 0x10A14}, {0x10A18, 0x10A18}, {0x10A36, 0x10A37}, {0x10A3B, 0x10A3E},
-{0x10A49, 0x10A4F}, {0x10A59, 0x10A5F}, {0x10AA0, 0x10ABF}, {0x10AE7, 0x10AEA}, {0x10AF7, 0x10AFF}, {0x10B36, 0x10B38}, {0x10B56, 0x10B57}, {0x10B73, 0x10B77}, {0x10B92, 0x10B98}, {0x10B9D, 0x10BA8},
-{0x10BB0, 0x10BFF}, {0x10C49, 0x10C7F}, {0x10CB3, 0x10CBF}, {0x10CF3, 0x10CF9}, {0x10D28, 0x10D2F}, {0x10D3A, 0x10E5F}, {0x10E7F, 0x10E7F}, {0x10EAA, 0x10EAA}, {0x10EAE, 0x10EAF}, {0x10EB2, 0x10EFF},
-{0x10F28, 0x10F2F}, {0x10F5A, 0x10FAF}, {0x10FCC, 0x10FDF}, {0x10FF7, 0x10FFF}, {0x1104E, 0x11051}, {0x11070, 0x1107E}, {0x110BD, 0x110BD}, {0x110C2, 0x110CF}, {0x110E9, 0x110EF}, {0x110FA, 0x110FF},
-{0x11135, 0x11135}, {0x11148, 0x1114F}, {0x11177, 0x1117F}, {0x111E0, 0x111E0}, {0x111F5, 0x111FF}, {0x11212, 0x11212}, {0x1123F, 0x1127F}, {0x11287, 0x11287}, {0x11289, 0x11289}, {0x1128E, 0x1128E},
-{0x1129E, 0x1129E}, {0x112AA, 0x112AF}, {0x112EB, 0x112EF}, {0x112FA, 0x112FF}, {0x11304, 0x11304}, {0x1130D, 0x1130E}, {0x11311, 0x11312}, {0x11329, 0x11329}, {0x11331, 0x11331}, {0x11334, 0x11334},
-{0x1133A, 0x1133A}, {0x11345, 0x11346}, {0x11349, 0x1134A}, {0x1134E, 0x1134F}, {0x11351, 0x11356}, {0x11358, 0x1135C}, {0x11364, 0x11365}, {0x1136D, 0x1136F}, {0x11375, 0x113FF}, {0x1145C, 0x1145C},
-{0x11462, 0x1147F}, {0x114C8, 0x114CF}, {0x114DA, 0x1157F}, {0x115B6, 0x115B7}, {0x115DE, 0x115FF}, {0x11645, 0x1164F}, {0x1165A, 0x1165F}, {0x1166D, 0x1167F}, {0x116B9, 0x116BF}, {0x116CA, 0x116FF},
-{0x1171B, 0x1171C}, {0x1172C, 0x1172F}, {0x11740, 0x117FF}, {0x1183C, 0x1189F}, {0x118F3, 0x118FE}, {0x11907, 0x11908}, {0x1190A, 0x1190B}, {0x11914, 0x11914}, {0x11917, 0x11917}, {0x11936, 0x11936},
-{0x11939, 0x1193A}, {0x11947, 0x1194F}, {0x1195A, 0x1199F}, {0x119A8, 0x119A9}, {0x119D8, 0x119D9}, {0x119E5, 0x119FF}, {0x11A48, 0x11A4F}, {0x11AA3, 0x11ABF}, {0x11AF9, 0x11BFF}, {0x11C09, 0x11C09},
-{0x11C37, 0x11C37}, {0x11C46, 0x11C4F}, {0x11C6D, 0x11C6F}, {0x11C90, 0x11C91}, {0x11CA8, 0x11CA8}, {0x11CB7, 0x11CFF}, {0x11D07, 0x11D07}, {0x11D0A, 0x11D0A}, {0x11D37, 0x11D39}, {0x11D3B, 0x11D3B},
-{0x11D3E, 0x11D3E}, {0x11D48, 0x11D4F}, {0x11D5A, 0x11D5F}, {0x11D66, 0x11D66}, {0x11D69, 0x11D69}, {0x11D8F, 0x11D8F}, {0x11D92, 0x11D92}, {0x11D99, 0x11D9F}, {0x11DAA, 0x11EDF}, {0x11EF9, 0x11FAF},
-{0x11FB1, 0x11FBF}, {0x11FF2, 0x11FFE}, {0x1239A, 0x123FF}, {0x1246F, 0x1246F}, {0x12475, 0x1247F}, {0x12544, 0x12FFF}, {0x1342F, 0x143FF}, {0x14647, 0x167FF}, {0x16A39, 0x16A3F}, {0x16A5F, 0x16A5F},
-{0x16A6A, 0x16A6D}, {0x16A70, 0x16ACF}, {0x16AEE, 0x16AEF}, {0x16AF6, 0x16AFF}, {0x16B46, 0x16B4F}, {0x16B5A, 0x16B5A}, {0x16B62, 0x16B62}, {0x16B78, 0x16B7C}, {0x16B90, 0x16E3F}, {0x16E9B, 0x16EFF},
-{0x16F4B, 0x16F4E}, {0x16F88, 0x16F8E}, {0x16FA0, 0x16FDF}, {0x16FE5, 0x16FEF}, {0x16FF2, 0x16FFF}, {0x187F8, 0x187FF}, {0x18CD6, 0x18CFF}, {0x18D09, 0x1AFFF}, {0x1B11F, 0x1B14F}, {0x1B153, 0x1B163},
-{0x1B168, 0x1B16F}, {0x1B2FC, 0x1BBFF}, {0x1BC6B, 0x1BC6F}, {0x1BC7D, 0x1BC7F}, {0x1BC89, 0x1BC8F}, {0x1BC9A, 0x1BC9B}, {0x1BCA0, 0x1CFFF}, {0x1D0F6, 0x1D0FF}, {0x1D127, 0x1D128}, {0x1D173, 0x1D17A},
-{0x1D1E9, 0x1D1FF}, {0x1D246, 0x1D2DF}, {0x1D2F4, 0x1D2FF}, {0x1D357, 0x1D35F}, {0x1D379, 0x1D3FF}, {0x1D455, 0x1D455}, {0x1D49D, 0x1D49D}, {0x1D4A0, 0x1D4A1}, {0x1D4A3, 0x1D4A4}, {0x1D4A7, 0x1D4A8},
-{0x1D4AD, 0x1D4AD}, {0x1D4BA, 0x1D4BA}, {0x1D4BC, 0x1D4BC}, {0x1D4C4, 0x1D4C4}, {0x1D506, 0x1D506}, {0x1D50B, 0x1D50C}, {0x1D515, 0x1D515}, {0x1D51D, 0x1D51D}, {0x1D53A, 0x1D53A}, {0x1D53F, 0x1D53F},
-{0x1D545, 0x1D545}, {0x1D547, 0x1D549}, {0x1D551, 0x1D551}, {0x1D6A6, 0x1D6A7}, {0x1D7CC, 0x1D7CD}, {0x1DA8C, 0x1DA9A}, {0x1DAA0, 0x1DAA0}, {0x1DAB0, 0x1DFFF}, {0x1E007, 0x1E007}, {0x1E019, 0x1E01A},
-{0x1E022, 0x1E022}, {0x1E025, 0x1E025}, {0x1E02B, 0x1E0FF}, {0x1E12D, 0x1E12F}, {0x1E13E, 0x1E13F}, {0x1E14A, 0x1E14D}, {0x1E150, 0x1E2BF}, {0x1E2FA, 0x1E2FE}, {0x1E300, 0x1E7FF}, {0x1E8C5, 0x1E8C6},
-{0x1E8D7, 0x1E8FF}, {0x1E94C, 0x1E94F}, {0x1E95A, 0x1E95D}, {0x1E960, 0x1EC70}, {0x1ECB5, 0x1ED00}, {0x1ED3E, 0x1EDFF}, {0x1EE04, 0x1EE04}, {0x1EE20, 0x1EE20}, {0x1EE23, 0x1EE23}, {0x1EE25, 0x1EE26},
-{0x1EE28, 0x1EE28}, {0x1EE33, 0x1EE33}, {0x1EE38, 0x1EE38}, {0x1EE3A, 0x1EE3A}, {0x1EE3C, 0x1EE41}, {0x1EE43, 0x1EE46}, {0x1EE48, 0x1EE48}, {0x1EE4A, 0x1EE4A}, {0x1EE4C, 0x1EE4C}, {0x1EE50, 0x1EE50},
-{0x1EE53, 0x1EE53}, {0x1EE55, 0x1EE56}, {0x1EE58, 0x1EE58}, {0x1EE5A, 0x1EE5A}, {0x1EE5C, 0x1EE5C}, {0x1EE5E, 0x1EE5E}, {0x1EE60, 0x1EE60}, {0x1EE63, 0x1EE63}, {0x1EE65, 0x1EE66}, {0x1EE6B, 0x1EE6B},
-{0x1EE73, 0x1EE73}, {0x1EE78, 0x1EE78}, {0x1EE7D, 0x1EE7D}, {0x1EE7F, 0x1EE7F}, {0x1EE8A, 0x1EE8A}, {0x1EE9C, 0x1EEA0}, {0x1EEA4, 0x1EEA4}, {0x1EEAA, 0x1EEAA}, {0x1EEBC, 0x1EEEF}, {0x1EEF2, 0x1EFFF},
-{0x1F02C, 0x1F02F}, {0x1F094, 0x1F09F}, {0x1F0AF, 0x1F0B0}, {0x1F0C0, 0x1F0C0}, {0x1F0D0, 0x1F0D0}, {0x1F0F6, 0x1F0FF}, {0x1F1AE, 0x1F1E5}, {0x1F203, 0x1F20F}, {0x1F23C, 0x1F23F}, {0x1F249, 0x1F24F},
-{0x1F252, 0x1F25F}, {0x1F266, 0x1F2FF}, {0x1F6D8, 0x1F6DF}, {0x1F6ED, 0x1F6EF}, {0x1F6FD, 0x1F6FF}, {0x1F774, 0x1F77F}, {0x1F7D9, 0x1F7DF}, {0x1F7EC, 0x1F7FF}, {0x1F80C, 0x1F80F}, {0x1F848, 0x1F84F},
-{0x1F85A, 0x1F85F}, {0x1F888, 0x1F88F}, {0x1F8AE, 0x1F8AF}, {0x1F8B2, 0x1F8FF}, {0x1F979, 0x1F979}, {0x1F9CC, 0x1F9CC}, {0x1FA54, 0x1FA5F}, {0x1FA6E, 0x1FA6F}, {0x1FA75, 0x1FA77}, {0x1FA7B, 0x1FA7F},
-{0x1FA87, 0x1FA8F}, {0x1FAA9, 0x1FAAF}, {0x1FAB7, 0x1FABF}, {0x1FAC3, 0x1FACF}, {0x1FAD7, 0x1FAFF}, {0x1FB93, 0x1FB93}, {0x1FBCB, 0x1FBEF}, {0x1FBFA, 0x1FFFF}, {0x2A6DE, 0x2A6FF}, {0x2B735, 0x2B73F},
-{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
-};
-
-static std::string codepoint_to_utf8(uint32_t cp) {
-    std::string result;
-    if (/* 0x00 <= cp && */ cp <= 0x7f) {
-        result.push_back(cp);
-    }
-    else if (0x80 <= cp && cp <= 0x7ff) {
-        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
-        result.push_back(0x80 | (cp & 0x3f));
-    }
-    else if (0x800 <= cp && cp <= 0xffff) {
-        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
-    }
-    else if (0x10000 <= cp && cp <= 0x10ffff) {
-        result.push_back(0xf0 | ((cp >> 18) & 0x07));
-        result.push_back(0x80 | ((cp >> 12) & 0x3f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
-    }
-    else {
-        throw std::invalid_argument("invalid codepoint");
-    }
-    return result;
-}
-
-static std::string codepoints_to_utf8(const std::vector<uint32_t> & cps) {
-    std::string result;
-    for (size_t i = 0; i < cps.size(); ++i) {
-        result.append(codepoint_to_utf8(cps[i]));
-    }
-    return result;
-}
-
-static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
-    assert(offset < utf8.size());
-    if (!(utf8[offset + 0] & 0x80)) {
-        auto result = utf8[offset + 0];
-        offset += 1;
-        return result;
-    }
-    else if (!(utf8[offset + 0] & 0x40)) {
-        throw std::invalid_argument("invalid character");
-    }
-    else if (!(utf8[offset + 0] & 0x20)) {
-        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
-            throw std::invalid_argument("invalid character");
-        auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
-        offset += 2;
-        return result;
-    }
-    else if (!(utf8[offset + 0] & 0x10)) {
-        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
-            throw std::invalid_argument("invalid character");
-        auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
-        offset += 3;
-        return result;
-    }
-    else if (!(utf8[offset + 0] & 0x08)) {
-        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
-            throw std::invalid_argument("invalid character");
-        auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
-        offset += 4;
-        return result;
-    }
-    throw std::invalid_argument("invalid string");
-}
-
-static std::vector<uint32_t> codepoints_from_utf8(const std::string & utf8) {
-    std::vector<uint32_t> result;
-    size_t offset = 0;
-    while (offset < utf8.size()) {
-        result.push_back(codepoint_from_utf8(utf8, offset));
-    }
-    return result;
-}
-
-static std::vector<uint16_t> codepoint_to_utf16(uint32_t cp) {
-    std::vector<uint16_t> result;
-    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
-        result.emplace_back(cp);
-    }
-    else if (0x10000 <= cp && cp <= 0x10ffff) {
-        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
-        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
-    }
-    else {
-        throw std::invalid_argument("invalid codepoint");
-    }
-    return result;
-}
-
-static std::vector<uint16_t> codepoints_to_utf16(const std::vector<uint32_t> & cps) {
-    std::vector<uint16_t> result;
-    for (size_t i = 0; i < cps.size(); ++i) {
-        auto temp = codepoint_to_utf16(cps[i]);
-        result.insert(result.end(), temp.begin(), temp.end());
-    }
-    return result;
-}
-
-static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
-    assert(offset < utf16.size());
-    if (((utf16[0] >> 10) << 10) != 0xd800) {
-        auto result = utf16[offset + 0];
-        offset += 1;
-        return result;
-    }
-    else {
-        if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
-            throw std::invalid_argument("invalid character");
-        auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
-        offset += 2;
-        return result;
-    }
-    throw std::invalid_argument("invalid string");
-}
-
-static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
-    std::vector<uint32_t> result;
-    size_t offset = 0;
-    while (offset < utf16.size())
-        result.push_back(codepoint_from_utf16(utf16, offset));
-    return result;
-}
 
 #define CODEPOINT_TYPE_UNIDENTIFIED 0
-#define CODEPOINT_TYPE_DIGIT 1
-#define CODEPOINT_TYPE_LETTER 2
-#define CODEPOINT_TYPE_WHITESPACE 3
-#define CODEPOINT_TYPE_ACCENT_MARK 4
-#define CODEPOINT_TYPE_PUNCTUATION 5
-#define CODEPOINT_TYPE_SYMBOL 6
-#define CODEPOINT_TYPE_CONTROL 7
+#define CODEPOINT_TYPE_DIGIT        1
+#define CODEPOINT_TYPE_LETTER       2
+#define CODEPOINT_TYPE_WHITESPACE   3
+#define CODEPOINT_TYPE_ACCENT_MARK  4
+#define CODEPOINT_TYPE_PUNCTUATION  5
+#define CODEPOINT_TYPE_SYMBOL       6
+#define CODEPOINT_TYPE_CONTROL      7
 
-static std::unordered_map<uint32_t, int> codepoint_type_map() {
-    std::unordered_map<uint32_t, int> codepoint_types;
-    for (auto p : digit_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
-    }
-    for(auto p : letter_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_LETTER;
-    }
-    for(auto p : whitespace_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
-    }
-    for(auto p : accent_mark_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
-    }
-    for(auto p : punctuation_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
-    }
-    for (auto p : symbol_ranges) {
-        for (auto i = p.first; i <= p.second; ++i)
-            codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
-    }
-    for(auto p : control_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
-            codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
-    }
-    return codepoint_types;
-}
+std::string unicode_cpt_to_utf8(uint32_t cp);
+std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
-static int codepoint_type(uint32_t cp) {
-    static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
-    return codepoint_types[cp];
-}
+std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-static int codepoint_type(const std::string & utf8) {
-    if (utf8.length() == 0)
-        return CODEPOINT_TYPE_UNIDENTIFIED;
-    size_t offset = 0;
-    return codepoint_type(codepoint_from_utf8(utf8, offset));
-}
+int unicode_cpt_type(uint32_t cp);
+int unicode_cpt_type(const std::string & utf8);
 
-static std::unordered_map<uint8_t, std::string> bytes_to_unicode_map_bpe() {
-    std::unordered_map<uint8_t, std::string> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[ch] = codepoint_to_utf8(ch);
-    }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[ch] = codepoint_to_utf8(ch);
-    }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[ch] = codepoint_to_utf8(ch);
-    }
-    auto n = 0;
-    for (int ch = 0; ch < 256; ++ch) {
-        if (map.find(ch) == map.end()) {
-            map[ch] = codepoint_to_utf8(256 + n);
-            ++n;
-        }
-    }
-    return map;
-}
-
-static std::string bytes_to_unicode_bpe(uint8_t byte) {
-    static std::unordered_map<uint8_t, std::string> map = bytes_to_unicode_map_bpe();
-    return map.at(byte);
-}
-
-static std::unordered_map<std::string, uint8_t> unicode_to_bytes_map_bpe() {
-    std::unordered_map<std::string, uint8_t> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[codepoint_to_utf8(ch)] = ch;
-    }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[codepoint_to_utf8(ch)] = ch;
-    }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
-        assert(0 <= ch && ch < 256);
-        map[codepoint_to_utf8(ch)] = ch;
-    }
-    auto n = 0;
-    for (int ch = 0; ch < 256; ++ch) {
-        if (map.find(codepoint_to_utf8(ch)) == map.end()) {
-            map[codepoint_to_utf8(256 + n)] = ch;
-            ++n;
-        }
-    }
-    return map;
-}
-
-static uint8_t unicode_to_bytes_bpe(const std::string & utf8) {
-    static std::unordered_map<std::string, uint8_t> map = unicode_to_bytes_map_bpe();
-    return map.at(utf8);
-}
+std::string unicode_byte_to_utf8(uint8_t byte);
+uint8_t unicode_utf8_to_byte(const std::string & utf8);
 

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;

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

examples/talk/README.md

@@ -11,9 +11,13 @@ Web version: [examples/talk.wasm](/examples/talk.wasm)
 The `talk` 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/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/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
 
@@ -525,8 +525,7 @@ bool gpt2_eval(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
-                        );
+                        1.0f/sqrt(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // [n_past + N, N, 12]

examples/talk/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 175 -p 50 -a 200 -g 5 -k 5 "$2"
+function installed() { command -v $1 >/dev/null 2>&1; }
 
-# Mac OS "say" command
-say "$2"
+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
+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 without api key
-#
-#export ELEVEN_API_KEY=your_api_key
-#wd=$(dirname $0)
-#script=$wd/eleven-labs.py
-#python3 $script $1 "$2"
-#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3
+elif installed python3 && \
+  python3 -c 'import importlib.util; exit(not importlib.util.find_spec("elevenlabs"))' && \
+  installed ffplay; then
+    # It's possible to use the API for free with limited number of characters.
+    # To increase this limit register to https://beta.elevenlabs.io to get an api key
+    # and paste it after 'ELEVEN_API_KEY='
+    # Keep the line commented to use the free version without api key
+    #export ELEVEN_API_KEY=your_api_key
+    wd=$(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/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/talk.cpp

@@ -38,6 +38,7 @@ struct whisper_params {
     std::string model_wsp = "models/ggml-base.en.bin";
     std::string model_gpt = "models/ggml-gpt-2-117M.bin";
     std::string speak     = "./examples/talk/speak";
+    std::string speak_file= "./examples/talk/to_speak.txt";
     std::string fname_out;
 };
 
@@ -68,6 +69,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
         else if (arg == "-mg"  || arg == "--model-gpt")     { params.model_gpt     = 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 == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@@ -102,6 +104,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
     fprintf(stderr, "  -mg FILE, --model-gpt     [%-7s] gpt model file\n",                              params.model_gpt.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, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
@@ -184,7 +187,7 @@ int main(int argc, char ** argv) {
     }
 
     // whisper init
-    struct whisper_context_params cparams;
+    struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
 
     struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
@@ -316,7 +319,7 @@ int main(int argc, char ** argv) {
                     std::string prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);
 
                     text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
-                    text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+                    //text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
                     text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));
 
                     // remove first 2 lines of base prompt
@@ -354,10 +357,7 @@ int main(int argc, char ** argv) {
                 gpt2_set_prompt(ctx_gpt, prompt_base.c_str());
 
                 text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
-                int ret = system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
-                if (ret != 0) {
-                    fprintf(stderr, "%s: system() failed!\n", __func__);
-                }
+                speak_with_file(params.speak, text_to_speak, params.speak_file, voice_id);
 
                 audio.clear();
 

examples/wchess/README.md

@@ -33,8 +33,13 @@ White's turn
 
 ## TODO
 
-- Improve web-browser audio capture - sometimes it does not record the voice properly
-- Add support for more languages by making the generated grammar string multi-lingual
 - Fix bugs in the chess moves logic
+- Improve web-browser audio capture - sometimes it does not record the voice properly
+- Add support for more languages by making the generated grammar string multilingual
+- Explore ways to improve the dynamic grammar to be narrower
 
 PRs welcome!
+
+## Thanks
+
+- [chessboardjs](https://chessboardjs.com) for the neat chessboard JS library used in this demo

examples/wchess/wchess.cmd/wchess.cmd.cpp

@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context_params cparams;
+    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);

examples/whisper.android.java/app/src/main/jni/whisper/jni.c

@@ -245,6 +245,8 @@ Java_com_whispercpp_java_whisper_WhisperLib_benchMemcpy(JNIEnv *env, jobject thi
     UNUSED(thiz);
     const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
     jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
+
+    return string;
 }
 
 JNIEXPORT jstring JNICALL
@@ -253,5 +255,7 @@ Java_com_whispercpp_java_whisper_WhisperLib_benchGgmlMulMat(JNIEnv *env, jobject
     UNUSED(thiz);
     const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
     jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
+
+    return string;
 }
 

examples/whisper.android/README.md

@@ -12,3 +12,47 @@ To use:
 (PS: Do not move this android project folder individually to other folders, because this android project folder depends on the files of the whole project.)
 
 <img width="300" alt="image" src="https://user-images.githubusercontent.com/1670775/221613663-a17bf770-27ef-45ab-9a46-a5f99ba65d2a.jpg">
+
+## CLBlast
+
+> [!NOTE]
+> - OpenCL does not have the same level of support as CUDA or Metal.
+> - Turning on CLBlast may degrade OpenCL performance if your device isn't already tuned. See [tuning.md](https://github.com/CNugteren/CLBlast/blob/162783a414969464ce3aa5adf5c2554afa5ee93e/doc/tuning.md#already-tuned-for-devices) for a list of devices that are already tuned and what to do if yours is missing.
+
+Build CLBlast.
+
+```
+# In path/to/CLBlast (we assume OpenCL-Headers relative location)
+$ANDROID_SDK_PATH/cmake/3.22.1/bin/cmake .. \
+    -DCMAKE_SYSTEM_NAME=Android \
+    -DCMAKE_SYSTEM_VERSION=33 \
+    -DCMAKE_ANDROID_ARCH_ABI=arm64-v8a \
+    -DCMAKE_ANDROID_NDK=$ANDROID_NDK_PATH \
+    -DCMAKE_ANDROID_STL_TYPE=c++_static \
+    -DOPENCL_ROOT=$(readlink -f ../../OpenCL-Headers) \
+    -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=BOTH \
+    -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+# Build libclblast.so
+make -j4
+```
+
+Pull `libGLES_mali.so` to `libOpenCL.so`.
+
+```bash
+# In path/to/whisper.android
+mkdir lib/src/main/jniLibs/arm64-v8a
+adb pull /system/vendor/lib64/egl/libGLES_mali.so lib/src/main/jniLibs/arm64-v8a/libOpenCL.so
+```
+
+In gradle.properties, set `GGML_HOME` to the location of GGML, as well as
+required options for turning on CLBlast.
+
+```
+GGML_HOME=/path/to/ggml
+GGML_CLBLAST=ON
+CLBLAST_HOME=/path/to/CLBlast
+OPENCL_LIB=/path/to/libOpenCL.so
+OPENCL_ROOT=/path/to/OpenCL-Headers
+```
+

examples/whisper.android/lib/build.gradle

@@ -16,6 +16,28 @@ android {
         ndk {
             abiFilters 'arm64-v8a', 'armeabi-v7a', 'x86', 'x86_64'
         }
+        externalNativeBuild {
+            cmake {
+                // When set, builds whisper.android against the version located
+                // at GGML_HOME instead of the copy bundled with whisper.cpp.
+                if (
+                    project.hasProperty('GGML_HOME') &&
+                    project.findProperty('GGML_CLBLAST') == 'ON'
+                ) {
+                    // Turning on CLBlast requires GGML_HOME
+                    arguments "-DGGML_HOME=${project.property('GGML_HOME')}",
+                         "-DGGML_CLBLAST=ON",
+                         "-DOPENCL_LIB=${project.property('OPENCL_LIB')}",
+                         "-DCLBLAST_HOME=${project.property('CLBLAST_HOME')}",
+                         "-DOPENCL_ROOT=${project.property('OPENCL_ROOT')}",
+                         "-DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH",
+                         "-DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=BOTH"
+                } else if (project.hasProperty('GGML_HOME')) {
+                    arguments "-DGGML_HOME=${project.property('GGML_HOME')}"
+                }
+
+            }
+        }
     }
 
     buildTypes {

examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -3,17 +3,27 @@ cmake_minimum_required(VERSION 3.10)
 project(whisper.cpp)
 
 set(CMAKE_CXX_STANDARD 11)
-set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)
+set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../..)
+
+# Path to external GGML, otherwise uses the copy in whisper.cpp.
+option(GGML_HOME       "whisper: Path to external GGML source" OFF)
 
 set(
+    SOURCE_FILES
+    ${WHISPER_LIB_DIR}/whisper.cpp
+    ${CMAKE_SOURCE_DIR}/jni.c
+    )
+
+if (NOT GGML_HOME)
+    set(
         SOURCE_FILES
+        ${SOURCE_FILES}
         ${WHISPER_LIB_DIR}/ggml.c
         ${WHISPER_LIB_DIR}/ggml-alloc.c
         ${WHISPER_LIB_DIR}/ggml-backend.c
         ${WHISPER_LIB_DIR}/ggml-quants.c
-        ${WHISPER_LIB_DIR}/whisper.cpp
-        ${CMAKE_SOURCE_DIR}/jni.c
-)
+        )
+endif()
 
 find_library(LOG_LIB log)
 
@@ -24,16 +34,15 @@ function(build_library target_name)
         ${SOURCE_FILES}
     )
 
-    target_link_libraries(${target_name} ${LOG_LIB} android)
-
     if (${target_name} STREQUAL "whisper_v8fp16_va")
         target_compile_options(${target_name} PRIVATE -march=armv8.2-a+fp16)
+        set(GGML_COMPILE_OPTIONS                      -march=armv8.2-a+fp16)
     elseif (${target_name} STREQUAL "whisper_vfpv4")
         target_compile_options(${target_name} PRIVATE -mfpu=neon-vfpv4)
+        set(GGML_COMPILE_OPTIONS                      -mfpu=neon-vfpv4)
     endif ()
 
     if (NOT ${CMAKE_BUILD_TYPE} STREQUAL "Debug")
-
         target_compile_options(${target_name} PRIVATE -O3)
         target_compile_options(${target_name} PRIVATE -fvisibility=hidden -fvisibility-inlines-hidden)
         target_compile_options(${target_name} PRIVATE -ffunction-sections -fdata-sections)
@@ -41,11 +50,21 @@ function(build_library target_name)
         target_link_options(${target_name} PRIVATE -Wl,--gc-sections)
         target_link_options(${target_name} PRIVATE -Wl,--exclude-libs,ALL)
         target_link_options(${target_name} PRIVATE -flto)
-
     endif ()
-endfunction()
 
-build_library("whisper") # Default target
+    if (GGML_HOME)
+        include(FetchContent)
+        FetchContent_Declare(ggml SOURCE_DIR ${GGML_HOME})
+        FetchContent_MakeAvailable(ggml)
+
+        target_compile_options(ggml PRIVATE ${GGML_COMPILE_OPTIONS})
+        target_link_libraries(${target_name} ${LOG_LIB} android ggml)
+    else()
+        target_link_libraries(${target_name} ${LOG_LIB} android)
+    endif()
+
+
+endfunction()
 
 if (${ANDROID_ABI} STREQUAL "arm64-v8a")
     build_library("whisper_v8fp16_va")
@@ -53,4 +72,6 @@ elseif (${ANDROID_ABI} STREQUAL "armeabi-v7a")
     build_library("whisper_vfpv4")
 endif ()
 
+build_library("whisper") # Default target
+
 include_directories(${WHISPER_LIB_DIR})

examples/whisper.android/lib/src/main/jni/whisper/jni.c

@@ -228,6 +228,7 @@ Java_com_whispercpp_whisper_WhisperLib_00024Companion_benchMemcpy(JNIEnv *env, j
     UNUSED(thiz);
     const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
     jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
+    return string;
 }
 
 JNIEXPORT jstring JNICALL
@@ -236,4 +237,5 @@ Java_com_whispercpp_whisper_WhisperLib_00024Companion_benchGgmlMulMat(JNIEnv *en
     UNUSED(thiz);
     const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
     jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
+    return string;
 }

examples/whisper.objc/README.md

@@ -11,11 +11,11 @@ https://user-images.githubusercontent.com/1991296/204126266-ce4177c6-6eca-4bd9-b
 
 ## Usage
 
-```java
+```bash
 git clone https://github.com/ggerganov/whisper.cpp
 open whisper.cpp/examples/whisper.objc/whisper.objc.xcodeproj/
 
-// If you don't want to convert a Core ML model, you can skip this step by create dummy model
+# if you don't want to convert a Core ML model, you can skip this step by create dummy model
 mkdir models/ggml-base.en-encoder.mlmodelc
 ```
 

examples/whisper.swiftui/.gitignore

@@ -0,0 +1,2 @@
+xcuserdata
+xcshareddata

extra/bench.py

@@ -61,7 +61,9 @@ models = [
     "ggml-small.bin",
     "ggml-medium.en.bin",
     "ggml-medium.bin",
-    "ggml-large.bin",
+    "ggml-large-v1.bin",
+    "ggml-large-v2.bin",
+    "ggml-large-v3.bin",
 ]
 
 

extra/sync-ggml-am.sh

@@ -0,0 +1,180 @@
+#!/bin/bash
+#
+# Synchronize ggml changes to whisper.cpp
+#
+# Usage:
+#
+#   $ cd /path/to/whisper.cpp
+#   $ ./extra/sync-ggml-am.sh -skip hash0,hash1,hash2...
+#
+
+set -e
+
+sd=$(dirname $0)
+cd $sd/../
+
+SRC_WHISPER=$(pwd)
+SRC_GGML=$(cd ../ggml; pwd)
+
+if [ ! -d $SRC_GGML ]; then
+    echo "ggml not found at $SRC_GGML"
+    exit 1
+fi
+
+lc=$(cat $SRC_WHISPER/extra/sync-ggml.last)
+echo "Syncing ggml changes since commit $lc"
+
+to_skip=""
+if [ "$1" == "-skip" ]; then
+    to_skip=$2
+fi
+
+cd $SRC_GGML
+
+git log --oneline $lc..HEAD
+git log --oneline $lc..HEAD --reverse | grep -v "(whisper/[0-9]*)" | cut -d' ' -f1 > $SRC_WHISPER/ggml-commits
+
+if [ ! -s $SRC_WHISPER/ggml-commits ]; then
+    rm -v $SRC_WHISPER/ggml-commits
+    echo "No new commits"
+    exit 0
+fi
+
+if [ -f $SRC_WHISPER/ggml-src.patch ]; then
+    rm -v $SRC_WHISPER/ggml-src.patch
+fi
+
+while read c; do
+    if [ -n "$to_skip" ]; then
+        if [[ $to_skip == *"$c"* ]]; then
+            echo "Skipping $c"
+            continue
+        fi
+    fi
+
+    git format-patch -k $c~1..$c --stdout -- \
+        include/ggml/ggml*.h \
+        src/ggml*.h \
+        src/ggml*.c \
+        src/ggml*.cpp \
+        src/ggml*.m \
+        src/ggml*.metal \
+        src/ggml*.cu \
+        examples/common.h \
+        examples/common.cpp \
+        examples/common-ggml.h \
+        examples/common-ggml.cpp \
+        examples/whisper/whisper.h \
+        examples/whisper/whisper.cpp \
+        examples/whisper/main.cpp \
+        examples/whisper/quantize.cpp \
+        >> $SRC_WHISPER/ggml-src.patch
+done < $SRC_WHISPER/ggml-commits
+
+rm -v $SRC_WHISPER/ggml-commits
+
+# delete files if empty
+if [ ! -s $SRC_WHISPER/ggml-src.patch ]; then
+    rm -v $SRC_WHISPER/ggml-src.patch
+fi
+
+cd $SRC_WHISPER
+
+if [ -f $SRC_WHISPER/ggml-src.patch ]; then
+    # replace PR numbers
+    #
+    # Subject: some text (#1234)
+    # Subject: some text (ggml/1234)
+    cat ggml-src.patch | sed -e 's/^Subject: \(.*\) (#\([0-9]*\))/Subject: \1 (ggml\/\2)/' > ggml-src.patch.tmp
+    mv ggml-src.patch.tmp ggml-src.patch
+
+    cat ggml-src.patch | sed -e 's/^\(.*\) (#\([0-9]*\))$/\1 (ggml\/\2)/' > ggml-src.patch.tmp
+    mv ggml-src.patch.tmp ggml-src.patch
+
+    # replace filenames:
+    #
+    # src/ggml.c                  -> ggml.c
+    # src/ggml-alloc.c            -> ggml-alloc.c
+    # src/ggml-backend-impl.h     -> ggml-backend-impl.h
+    # src/ggml-backend.c          -> ggml-backend.c
+    # src/ggml-common.h           -> ggml-common.h
+    # src/ggml-cuda.cu            -> ggml-cuda.cu
+    # src/ggml-cuda.h             -> ggml-cuda.h
+    # src/ggml-impl.h             -> ggml-impl.h
+    # src/ggml-kompute.cpp        -> ggml-kompute.cpp
+    # src/ggml-kompute.h          -> ggml-kompute.h
+    # src/ggml-metal.h            -> ggml-metal.h
+    # src/ggml-metal.m            -> ggml-metal.m
+    # src/ggml-mpi.h              -> ggml-mpi.h
+    # src/ggml-mpi.c              -> ggml-mpi.c
+    # src/ggml-opencl.cpp         -> ggml-opencl.cpp
+    # src/ggml-opencl.h           -> ggml-opencl.h
+    # src/ggml-quants.c           -> ggml-quants.c
+    # src/ggml-quants.h           -> ggml-quants.h
+    # src/ggml-sycl.cpp           -> ggml-sycl.cpp
+    # src/ggml-sycl.h             -> ggml-sycl.h
+    # src/ggml-vulkan.cpp         -> ggml-vulkan.cpp
+    # src/ggml-vulkan.h           -> ggml-vulkan.h
+    # include/ggml/ggml.h         -> ggml.h
+    # include/ggml/ggml-alloc.h   -> ggml-alloc.h
+    # include/ggml/ggml-backend.h -> ggml-backend.h
+    #
+    # examples/common.h           -> examples/common.h
+    # examples/common.cpp         -> examples/common.cpp
+    # examples/common-ggml.h      -> examples/common-ggml.h
+    # examples/common-ggml.cpp    -> examples/common-ggml.cpp
+    #
+    # examples/whisper/whisper.h    -> whisper.h
+    # examples/whisper/whisper.cpp  -> whisper.cpp
+    # examples/whisper/main.cpp     -> examples/main/main.cpp
+    # examples/whisper/quantize.cpp -> examples/quantize/quantize.cpp
+
+    cat ggml-src.patch | sed \
+        -e 's/src\/ggml\.c/ggml.c/g' \
+        -e 's/src\/ggml-alloc\.c/ggml-alloc.c/g' \
+        -e 's/src\/ggml-backend-impl\.h/ggml-backend-impl.h/g' \
+        -e 's/src\/ggml-backend\.c/ggml-backend.c/g' \
+        -e 's/src\/ggml-common\.h/ggml-common.h/g' \
+        -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \
+        -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \
+        -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \
+        -e 's/src\/ggml-kompute\.cpp/ggml-kompute.cpp/g' \
+        -e 's/src\/ggml-kompute\.h/ggml-kompute.h/g' \
+        -e 's/src\/ggml-metal\.h/ggml-metal.h/g' \
+        -e 's/src\/ggml-metal\.m/ggml-metal.m/g' \
+        -e 's/src\/ggml-mpi\.h/ggml-mpi.h/g' \
+        -e 's/src\/ggml-mpi\.c/ggml-mpi.c/g' \
+        -e 's/src\/ggml-opencl\.cpp/ggml-opencl.cpp/g' \
+        -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
+        -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \
+        -e 's/src\/ggml-quants\.h/ggml-quants.h/g' \
+        -e 's/src\/ggml-sycl\.cpp/ggml-sycl.cpp/g' \
+        -e 's/src\/ggml-sycl\.h/ggml-sycl.h/g' \
+        -e 's/src\/ggml-vulkan\.cpp/ggml-vulkan.cpp/g' \
+        -e 's/src\/ggml-vulkan\.h/ggml-vulkan.h/g' \
+        -e 's/include\/ggml\/ggml\.h/ggml.h/g' \
+        -e 's/include\/ggml\/ggml-alloc\.h/ggml-alloc.h/g' \
+        -e 's/include\/ggml\/ggml-backend\.h/ggml-backend.h/g' \
+        -e 's/examples\/common\.h/examples\/common.h/g' \
+        -e 's/examples\/common\.cpp/examples\/common.cpp/g' \
+        -e 's/examples\/common-ggml\.h/examples\/common-ggml.h/g' \
+        -e 's/examples\/common-ggml\.cpp/examples\/common-ggml.cpp/g' \
+        -e 's/examples\/whisper\/whisper\.h/whisper.h/g' \
+        -e 's/examples\/whisper\/whisper\.cpp/whisper.cpp/g' \
+        -e 's/examples\/whisper\/main\.cpp/examples\/main\/main.cpp/g' \
+        -e 's/examples\/whisper\/quantize\.cpp/examples\/quantize\/quantize.cpp/g' \
+        > ggml-src.patch.tmp
+    mv ggml-src.patch.tmp ggml-src.patch
+
+    git am ggml-src.patch
+
+    rm -v $SRC_WHISPER/ggml-src.patch
+fi
+
+# update last commit
+cd $SRC_GGML
+git log -1 --format=%H > $SRC_WHISPER/extra/sync-ggml.last
+
+echo "Done"
+
+exit 0

extra/sync-ggml.last

@@ -0,0 +1 @@
+7652115c79fa0ffedb58a28c09cd2871403d5a20

extra/sync-ggml.sh

@@ -5,8 +5,11 @@ cp -rpv ../ggml/src/ggml-impl.h         ./ggml-impl.h
 cp -rpv ../ggml/src/ggml-alloc.c        ./ggml-alloc.c
 cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml-backend-impl.h
 cp -rpv ../ggml/src/ggml-backend.c      ./ggml-backend.c
+cp -rpv ../ggml/src/ggml-common.h       ./ggml-common.h
 cp -rpv ../ggml/src/ggml-cuda.cu        ./ggml-cuda.cu
 cp -rpv ../ggml/src/ggml-cuda.h         ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-kompute.cpp    ./ggml-kompute.cpp
+cp -rpv ../ggml/src/ggml-kompute.h      ./ggml-kompute.h
 cp -rpv ../ggml/src/ggml-metal.h        ./ggml-metal.h
 cp -rpv ../ggml/src/ggml-metal.m        ./ggml-metal.m
 cp -rpv ../ggml/src/ggml-metal.metal    ./ggml-metal.metal
@@ -16,6 +19,10 @@ cp -rpv ../ggml/src/ggml-opencl.cpp     ./ggml-opencl.cpp
 cp -rpv ../ggml/src/ggml-opencl.h       ./ggml-opencl.h
 cp -rpv ../ggml/src/ggml-quants.c       ./ggml-quants.c
 cp -rpv ../ggml/src/ggml-quants.h       ./ggml-quants.h
+cp -rpv ../ggml/src/ggml-sycl.cpp       ./ggml-sycl.cpp
+cp -rpv ../ggml/src/ggml-sycl.h         ./ggml-sycl.h
+cp -rpv ../ggml/src/ggml-vulkan.cpp     ./ggml-vulkan.cpp
+cp -rpv ../ggml/src/ggml-vulkan.h       ./ggml-vulkan.h
 
 cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h

extra/sync-llama.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+cp -rpv ../llama.cpp/llama.h     ./examples/talk-llama/llama.h
+cp -rpv ../llama.cpp/llama.cpp   ./examples/talk-llama/llama.cpp
+cp -rpv ../llama.cpp/unicode.h   ./examples/talk-llama/unicode.h
+cp -rpv ../llama.cpp/unicode.cpp ./examples/talk-llama/unicode.cpp

ggml-alloc.h

@@ -6,86 +6,70 @@
 extern "C" {
 #endif
 
-struct ggml_backend;
-struct ggml_backend_buffer;
-struct ggml_backend_buffer_type;
-
-//
-// Legacy API
-//
-
-typedef struct ggml_allocr * ggml_allocr_t;
-
-// initialize allocator for use with CPU backend only
-GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
-
-// initialize allocator for use with ggml-backend
-GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
-
-// tell the allocator to parse nodes following the order described in the list
-// you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
-
-GGML_API void   ggml_allocr_free       (ggml_allocr_t alloc);
-GGML_API bool   ggml_allocr_is_measure (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_reset      (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_alloc      (ggml_allocr_t alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_max_size   (ggml_allocr_t alloc);
-
-GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
-
-//
-// ggml-backend v2 API
-//
-
-// Separate tensor and graph allocator objects
-// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
-// The original API is kept as a wrapper around the new API
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct ggml_backend * ggml_backend_t;
 
 // Tensor allocator
-typedef struct ggml_tallocr * ggml_tallocr_t;
-
-GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
-
-GGML_API void   ggml_tallocr_free       (ggml_tallocr_t talloc);
-GGML_API bool   ggml_tallocr_is_measure (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_reset      (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_alloc      (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_tallocr_max_size   (ggml_tallocr_t talloc);
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
 
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
 
 // Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
-GGML_API ggml_gallocr_t ggml_gallocr_new(void);
-GGML_API void   ggml_gallocr_free(ggml_gallocr_t galloc);
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 
-GGML_API void   ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
-GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
 
-// Allocate tensors from the allocators given by the hash table
-GGML_API void   ggml_gallocr_alloc_graph_n(
-                    ggml_gallocr_t galloc,
-                    struct ggml_cgraph * graph,
-                    struct ggml_hash_set hash_set,
-                    ggml_tallocr_t * hash_node_talloc);
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
 
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
 
 #ifdef  __cplusplus
 }

ggml-backend-impl.h

@@ -16,10 +16,15 @@ extern "C" {
     typedef void * ggml_backend_buffer_type_context_t;
 
     struct ggml_backend_buffer_type_i {
-        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
+        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
@@ -31,15 +36,15 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void     (*free_buffer)(ggml_backend_buffer_t buffer);
-        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void *   (*get_base)   (ggml_backend_buffer_t buffer);
-        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
+        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
+        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
+        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
+        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
     };
 
     struct ggml_backend_buffer {
@@ -47,14 +52,22 @@ extern "C" {
         ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
         size_t size;
+        enum ggml_backend_buffer_usage usage;
     };
 
-    ggml_backend_buffer_t ggml_backend_buffer_init(
+    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
                    ggml_backend_buffer_type_t      buft,
             struct ggml_backend_buffer_i           iface,
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend
@@ -63,49 +76,60 @@ extern "C" {
     typedef void * ggml_backend_context_t;
 
     struct ggml_backend_i {
-        const char * (*get_name)(ggml_backend_t backend);
+        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
 
-        void (*free)(ggml_backend_t backend);
+        void (*GGML_CALL free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
+        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
 
-        // (optional) asynchroneous tensor data access
-        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);
+        // (optional) asynchronous tensor data access
+        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        // (optional) complete all pending operations
+        void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        void (*synchronize)     (ggml_backend_t backend);
+        // compute graph with a plan (not used currently)
+        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
         // 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
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // compute graph without a plan (async)
+        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
-        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
+        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
     };
 
     struct ggml_backend {
-        struct ggml_backend_i iface;
+        ggml_guid_t guid;
 
+        struct ggml_backend_i iface;
         ggml_backend_context_t context;
     };
 
+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
 
     //
     // Backend registry
     //
 
-    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
 
-    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
 
 #ifdef  __cplusplus
 }

ggml-backend.h

@@ -9,6 +9,7 @@ extern "C" {
 
     typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
     typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
     typedef struct ggml_backend * ggml_backend_t;
     typedef void * ggml_backend_graph_plan_t;
 
@@ -17,50 +18,79 @@ extern "C" {
     //
 
     // buffer type
-    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
-    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+    };
+
+    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     //
     // Backend
     //
 
-
+    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
     GGML_API const char * ggml_backend_name(ggml_backend_t backend);
     GGML_API void         ggml_backend_free(ggml_backend_t backend);
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
     GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
 
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    GGML_API 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 GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
     GGML_API void ggml_backend_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 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_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
 
     //
     // CPU backend
@@ -68,13 +98,18 @@ extern "C" {
 
     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);
+    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
     // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
 
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
 
     //
     // Backend registry
@@ -102,54 +137,71 @@ extern "C" {
     /*
       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
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be asigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
 
-        // 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);
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);
 
-            // 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);
-        }
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
 
-        // allocate backend buffers from measure graph
-        ggml_backend_sched_init_measure(sched, measure_graph);
-
-        // the scheduler is now ready to compute graphs
+        ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
         graph = build_graph(sched);
         ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
     */
 
     struct ggml_backend_sched;
     typedef struct ggml_backend_sched * ggml_backend_sched_t;
 
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+    // when ask == true, the scheduler wants to know if the user wants to observe this node
+    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
+    //
+    // when ask == false, the scheduler is passing the node tensor to the user for observation
+    // if the user returns false, the scheduler will cancel the graph compute
+    //
+    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // 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 bool                 ggml_backend_sched_reserve(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);
+    // Get the number of splits of the last graph
+    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
 
-    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, 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);
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
+    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
+
+    // Reset all assignments and allocators - must be called before changing the node backends
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+
+    // Set a callback to be called for each resulting node during graph compute
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
 
     //
     // Utils
@@ -166,10 +218,10 @@ extern "C" {
     GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
     GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
 
-    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml-cuda.h

@@ -18,46 +18,34 @@ extern "C" {
 #define GGML_CUDA_MAX_DEVICES       16
 
 // Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
-GGML_API void   ggml_init_cublas(void);
+GGML_API GGML_CALL void   ggml_init_cublas(void);
 
 // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
-GGML_API bool   ggml_cublas_loaded(void);
+GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
 
-GGML_API void * ggml_cuda_host_malloc(size_t size);
-GGML_API void   ggml_cuda_host_free(void * ptr);
+GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
+GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
 
-GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
-GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
+GGML_API GGML_CALL bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API GGML_CALL bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
-GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
-GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_set_main_device(int main_device);
-GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);
-GGML_API void   ggml_cuda_free_scratch(void);
-GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-GGML_API int    ggml_cuda_get_device_count(void);
-GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL int    ggml_cuda_get_device_count(void);
+GGML_API GGML_CALL void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
-GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
+GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
 #ifdef  __cplusplus
 }

ggml-impl.h

@@ -5,6 +5,7 @@
 // GGML internal header
 
 #include <assert.h>
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
@@ -18,6 +19,7 @@ extern "C" {
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef __cplusplus
 #ifndef static_assert
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
 #define static_assert(cond, msg) _Static_assert(cond, msg)
@@ -25,6 +27,7 @@ extern "C" {
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
 #endif
+#endif
 
 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
@@ -50,14 +53,30 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+typedef __fp16 ggml_fp16_internal_t;
 
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    ggml_fp16_internal_t tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    ggml_fp16_internal_t tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
 
 #else
 
+typedef uint16_t ggml_fp16_internal_t;
+
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -211,8 +230,7 @@ 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)
-
+#if !defined(GGML_FP16_TO_FP32)
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
@@ -220,13 +238,17 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #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
 
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 #endif
 
 #define GGML_HASHTABLE_FULL ((size_t)-1)
 #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
 
+struct ggml_hash_set ggml_hash_set_new(size_t size);
+
 bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted

ggml-kompute.h

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

ggml-metal.h

@@ -27,7 +27,6 @@
 
 // max memory buffers that can be mapped to the device
 #define GGML_METAL_MAX_BUFFERS 64
-#define GGML_METAL_MAX_COMMAND_BUFFERS 32
 
 struct ggml_tensor;
 struct ggml_cgraph;
@@ -36,76 +35,31 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-//
-// internal API
-// temporary exposed to user-code
-//
-
-struct ggml_metal_context;
-
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-// number of command buffers to use
-struct ggml_metal_context * ggml_metal_init(int n_cb);
-void ggml_metal_free(struct ggml_metal_context * ctx);
-
-void * ggml_metal_host_malloc(size_t n);
-void   ggml_metal_host_free  (void * data);
-
-// set the number of command buffers to use
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
-
-// creates a mapping between a host memory buffer and a device memory buffer
-// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
-// - the mapping is used during computation to determine the arguments of the compute kernels
-// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
-// - max_size specifies the maximum size of a tensor and is used to create shared views such
-//   that it is guaranteed that the tensor will fit in at least one of the views
-//
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                       const char * name,
-                             void * data,
-                           size_t   size,
-                           size_t   max_size);
-
-// set data from host memory into the device
-void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// get data from the device into host memory
-void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// try to find operations that can be run concurrently in the graph
-// you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
-
-// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
-int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
-
-// output the concur_list for ggml_alloc
-int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
-
-// same as ggml_graph_compute but uses Metal
-// creates gf->n_threads command buffers in parallel
-void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
 //
 // backend API
 // user-code should use only these functions
 //
 
+GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
+GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
 GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
+// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
+GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
 #ifdef __cplusplus
 }
 #endif

ggml-opencl.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-opencl.h"
+#include "ggml-backend-impl.h"
 
 #include <array>
 #include <atomic>
@@ -10,7 +11,7 @@
 #include <sstream>
 #include <vector>
 
-#define CL_TARGET_OPENCL_VERSION 110
+#define CL_TARGET_OPENCL_VERSION 120
 #include <clblast.h>
 
 #if defined(_MSC_VER)
@@ -713,7 +714,6 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx,
         dst[row] = tmp[0];
     }
 }
-
 );
 
 
@@ -783,6 +783,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
         dst[row] = tmp[0];
     }
 }
+
 );
 
 
@@ -798,6 +799,18 @@ __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y
 }
 );
 
+std::string add_template = MULTILINE_QUOTE(
+__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) {
+    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
+
+    if (i >= get_global_size(0)) {
+        return;
+    }
+
+    dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky];
+}
+);
+
 #define CL_CHECK(err)                                               \
     do {                                                            \
         cl_int err_ = (err);                                        \
@@ -877,6 +890,7 @@ static std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
+    src << add_template << '\n';
 
     return src.str();
 }
@@ -892,6 +906,7 @@ static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl,
 static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
 static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
 static cl_kernel mul_f32_cl;
+static cl_kernel add_f32_cl;
 static bool fp16_support;
 
 static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
@@ -929,6 +944,12 @@ static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, co
 }
 
 void ggml_cl_init(void) {
+    static bool initialized = false;
+    if (initialized) {
+        return;
+    }
+    initialized = true;
+
     cl_int err;
 
     struct cl_device;
@@ -1093,9 +1114,10 @@ void ggml_cl_init(void) {
     char *ext_buffer = (char *)alloca(ext_str_size + 1);
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
     ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
+    // Disabled due to faulty outputs
     // Check if ext_buffer contains cl_khr_fp16
-    fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
-    fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
+    fp16_support = false;  // strstr(ext_buffer, "cl_khr_fp16") != NULL;
+    // fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
 
     cl_context_properties properties[] = {
         (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
@@ -1143,6 +1165,8 @@ void ggml_cl_init(void) {
 
     // mul kernel
     CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
+
+    CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err));
 }
 
 static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
@@ -1330,7 +1354,7 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
 }
 
 void ggml_cl_free_data(const struct ggml_tensor* tensor) {
-    if (tensor->backend != GGML_BACKEND_GPU) {
+    if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
         return;
     }
 
@@ -1388,7 +1412,7 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
@@ -1451,6 +1475,70 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
     ggml_cl_mul_f32(src0, src1, dst);
 }
 
+static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+    size_t x_size;
+    size_t d_size;
+
+    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
+    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
+
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            cl_event ev;
+
+            // copy src0 to device
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
+
+            const int64_t i13 = i03%ne13;
+            const int64_t i12 = i02%ne12;
+            const int i1 = i13*ne12*ne11 + i12*ne11;
+
+            cl_int x_offset = 0;
+            cl_int y_offset = i1*ne10;
+            cl_int d_offset = 0;
+
+            size_t global = ne00 * ne01;
+            cl_int ky = ne10 * ne11;
+
+            CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky));
+            CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
+
+            CL_CHECK(clReleaseEvent(ev));
+            CL_CHECK(clFinish(queue));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
+        }
+    }
+    ggml_cl_pool_free(d_X, x_size);
+    ggml_cl_pool_free(d_D, d_size);
+}
+
+void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cl_add_f32(src0, src1, dst);
+}
+
 static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -1478,13 +1566,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
+    cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D =  dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
 
@@ -1492,7 +1580,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_GPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
@@ -1501,7 +1589,9 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // copy src1 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    if (src1->backend == GGML_BACKEND_TYPE_CPU) {
+                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
+                    }
 
                     CL_CHECK(clFinish(queue));
 
@@ -1522,18 +1612,24 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
+                    }
                 }
             }
         }
     }
 
-    if (src0->backend != GGML_BACKEND_GPU) {
+    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
+    if (src1->backend != GGML_BACKEND_TYPE_GPU) {
+        ggml_cl_pool_free(d_Y, y_size);
+    }
+    if (dst->backend != GGML_BACKEND_TYPE_GPU) {
+        ggml_cl_pool_free(d_D, d_size);
+    }
 }
 
 static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
@@ -1574,7 +1670,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
@@ -1591,13 +1687,15 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_GPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
                 }
 
+                // FIXME: convert on device
+
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // convert src1 to fp16
                     // TODO: use multiple threads
@@ -1643,17 +1741,19 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host, then convert to float
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
-
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-
-                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                        CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
+                        float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                        ggml_fp16_to_fp32_row(tmp, d, d_ne);
+                    } else {
+                        // FIXME: convert dst to fp32 on device
+                    }
                 }
             }
         }
     }
 
-    if (src0->backend != GGML_BACKEND_GPU) {
+    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
     ggml_cl_pool_free(d_Y, y_size);
@@ -1698,7 +1798,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
     cl_mem d_Q;
-    if (src0->backend == GGML_BACKEND_CPU) {
+    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
         d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
     }
 
@@ -1717,10 +1817,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 // copy src0 to device if necessary
-                if (src0->backend == GGML_BACKEND_CPU) {
+                if (src0->backend == GGML_BACKEND_TYPE_CPU) {
                     events.emplace_back();
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-                } else if (src0->backend == GGML_BACKEND_GPU) {
+                } else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
                     d_Q = (cl_mem) src0->extra;
                 } else {
                     GGML_ASSERT(false);
@@ -1729,7 +1829,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 if (!mul_mat_vec) {
                     // convert src0 to fp32 on device
                     const size_t global = x_ne / global_denom;
-                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                    const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
                     CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
@@ -1743,7 +1843,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
                         // compute
                         const size_t global = ne01 * local;
-                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                         const cl_int ncols = ne00;
                         events.emplace_back();
                         CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
@@ -1795,13 +1895,13 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     }
     ggml_cl_pool_free(d_Y, y_size);
     ggml_cl_pool_free(d_D, d_size);
-    if (src0->backend == GGML_BACKEND_CPU) {
+    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
         ggml_cl_pool_free(d_Q, q_size);
     }
 }
 
 
-bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst) {
     const int64_t ne10 = src1->ne[0];
 
     const int64_t ne0 = dst->ne[0];
@@ -1811,7 +1911,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
         return true;
     }
 
@@ -1893,5 +1993,304 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->extra = dst;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
 }
+
+// ggml-backend
+
+// buffer
+
+struct ggml_backend_opencl_buffer_context {
+    ~ggml_backend_opencl_buffer_context() {
+        if (buffer) {
+            clReleaseMemObject(buffer);
+        }
+        for (auto * sub_buffer : sub_buffers) {
+            clReleaseMemObject(sub_buffer);
+        }
+    }
+
+    cl_mem buffer;
+    std::vector<cl_mem> sub_buffers;
+};
+
+static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
+
+static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return cl_ptr_base;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        tensor->extra = tensor->view_src->extra;
+    } else {
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
+        cl_int err;
+        cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        ctx->sub_buffers.push_back(sub_buffer);
+        tensor->extra = sub_buffer;
+    }
+    tensor->backend = GGML_BACKEND_TYPE_GPU;
+}
+
+static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    cl_mem tensor_buffer = (cl_mem) tensor->extra;
+    CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+}
+
+static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    for (auto * sub_buffer : ctx->sub_buffers) {
+        clReleaseMemObject(sub_buffer);
+    }
+    ctx->sub_buffers.clear();
+}
+
+static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
+    /* .get_name        = */ ggml_backend_opencl_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_opencl_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_opencl_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_opencl_buffer_init_tensor,
+    /* .set_tensor      = */ ggml_backend_opencl_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_opencl_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_opencl_buffer_clear,
+    /* .reset           = */ ggml_backend_opencl_buffer_reset,
+};
+
+// buffer type
+
+static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
+    ggml_cl_init();
+
+    cl_int err;
+    cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
+    if (err != CL_SUCCESS) {
+        fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
+        return nullptr;
+    }
+
+    ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
+
+    return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
+    // FIXME: not thread safe, device may not be initialized yet
+    static cl_uint alignment = -1;
+    if (alignment == (cl_uint)-1) {
+        ggml_cl_init();
+        clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
+    }
+    return alignment;
+
+    GGML_UNUSED(buffer_type);
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
+    static size_t max_size = -1;
+    if (max_size == (size_t)-1) {
+        ggml_cl_init();
+        clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_size, NULL);
+    }
+    return max_size;
+}
+
+static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
+    //return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
+    return ggml_backend_is_cpu(backend);
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_opencl_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_opencl_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ NULL,
+    /* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
+    /* .is_host          = */ NULL,
+};
+
+
+ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
+    static ggml_backend_buffer_type buffer_type = {
+        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
+        /* .context = */ nullptr,
+    };
+
+    return &buffer_type;
+}
+
+#if 0
+// host buffer type
+
+static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CL_Host";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CL_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_cl_host_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr = ggml_cl_host_malloc(size);
+
+    if (ptr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_opencl_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_opencl_buffer_type_host;
+}
+
+// backend
+
+static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
+    return "OpenCL";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_opencl_free(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_opencl_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
+    for (int i = 0; i < graph->n_nodes; ++i) {
+        ggml_tensor * node = graph->nodes[i];
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
+                break;
+            case GGML_OP_MUL:
+                ggml_cl_mul(node->src[0], node->src[1], node);
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
+        case GGML_OP_MUL:
+            // return ggml_can_repeat_rows(op->src[1], op->src[0]);
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+static ggml_backend_i opencl_backend_i = {
+    /* .get_name                = */ ggml_backend_opencl_name,
+    /* .free                    = */ ggml_backend_opencl_free,
+    /* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_from_async   = */ NULL,
+    /* .cpy_tensor_to_async     = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
+    /* .supports_op             = */ ggml_backend_opencl_supports_op,
+};
+
+ggml_backend_t ggml_backend_opencl_init() {
+    ggml_backend_t backend = new ggml_backend {
+        /* .interface = */ opencl_backend_i,
+        /* .context   = */ nullptr
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_opencl(ggml_backend_t backend) {
+    return backend && backend->iface.get_name == ggml_backend_opencl_name;
+}
+#endif

ggml-opencl.h

@@ -1,24 +1,35 @@
 #pragma once
 
 #include "ggml.h"
+#include "ggml-backend.h"
 
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-void ggml_cl_init(void);
+GGML_API void ggml_cl_init(void);
 
-void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
+GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
 
-void * ggml_cl_host_malloc(size_t size);
-void   ggml_cl_host_free(void * ptr);
+// GGML_API void * ggml_cl_host_malloc(size_t size);
+// GGML_API void   ggml_cl_host_free(void * ptr);
 
-void ggml_cl_free_data(const struct ggml_tensor* tensor);
+GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
 
-void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
+
+// backend API
+
+// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
+
+// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }

ggml-quants.h

@@ -1,224 +1,130 @@
 #pragma once
 
-#include "ggml-impl.h"
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.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
+#ifdef __cplusplus
+extern "C" {
 #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_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int k);
+void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int k);
+void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int k);
+void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int k);
+void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int k);
+void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_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_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k);
+void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int k);
+void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int k);
+void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int k);
+void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k);
+void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_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_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
+void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int k);
+void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int k);
+void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int k);
+void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_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);
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+
+void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_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_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_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);
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_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_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
+void iq3xs_init_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif
 
-void ggml_vec_dot_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);