ggml : disable CUDA graphs for non-llama.cpp projects

* scripts : update sync [no ci]

* files : reorganize [no ci]

* sync : llama.cpp

* cmake : link math library

* cmake : build normal ggml library

* files : move headers to include

* objc : fix path to ggml-metal.h

* ci : fix WHISPER_CUDA -> GGML_CUDA

* scripts : sync LICENSE [no ci]

.devops/cublas.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV WHISPER_CUBLAS=1
+ENV GGML_CUDA=1
 
 RUN make
 

.devops/main-cuda.Dockerfile

@@ -14,7 +14,7 @@ ARG CUDA_DOCKER_ARCH=all
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV WHISPER_CUBLAS=1
+ENV GGML_CUDA=1
 
 RUN apt-get update && \
     apt-get install -y build-essential \

.github/workflows/build.yml

@@ -15,10 +15,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -36,7 +36,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         run: |
@@ -53,10 +53,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
-        uses: cross-platform-actions/action@v0.15.0
+        uses: cross-platform-actions/action@v0.24.0
         with:
           operating_system: freebsd
           version: '13.2'
@@ -77,10 +77,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -101,14 +101,17 @@ jobs:
       fail-fast: false
       matrix:
         build: [Debug, Release]
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        #arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        # TODO: arm/v7 disabled due to clang bug
+        #       https://github.com/ggerganov/whisper.cpp/actions/runs/9657764109/job/26637633042?pr=2256#step:4:1990
+        arch: [linux/amd64, linux/arm64, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -133,10 +136,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: docker/setup-qemu-action@v3
 
       - name: Build ${{ matrix.arch }}
         run: |
@@ -165,7 +168,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -189,7 +192,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -197,7 +200,7 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
   ubuntu-22-cmake-sycl-fp16:
@@ -215,7 +218,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -239,7 +242,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -247,7 +250,7 @@ jobs:
           source /opt/intel/oneapi/setvars.sh
           mkdir build
           cd build
-          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DGGML_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
           cmake --build . --config Release -j $(nproc)
 
   windows-msys2:
@@ -262,7 +265,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup ${{ matrix.sys }}
         uses: msys2/setup-msys2@v2
@@ -289,7 +292,7 @@ jobs:
       - name: Build using make w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            make WHISPER_OPENBLAS=1 -j $(nproc)
+            make GGML_OPENBLAS=1 -j $(nproc)
 
       - name: Build using CMake
         shell: msys2 {0}
@@ -305,7 +308,7 @@ jobs:
       - name: Build using CMake w/ OpenBLAS
         shell: msys2 {0}
         run: |
-            cmake -B build -DWHISPER_OPENBLAS=ON
+            cmake -B build -DGGML_OPENBLAS=ON
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows:
@@ -328,10 +331,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1
+        uses: microsoft/setup-msbuild@v2
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -356,14 +359,14 @@ jobs:
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
       - name: Upload dll
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.jnaPath }}_whisper.dll
           path: build/bin/${{ matrix.build }}/whisper.dll
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v1
+        uses: actions/upload-artifact@v4
         with:
           name: whisper-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
@@ -381,21 +384,18 @@ 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: 2.28.5
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1
+        uses: microsoft/setup-msbuild@v2
 
       - name: Fetch OpenBLAS
         if: matrix.blas == 'ON'
@@ -413,26 +413,13 @@ jobs:
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
-      - name: Install OpenCL
-        if: matrix.clblast == 'ON'
-        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
-
-      - name: Fetch CLBlast and set CLBlast_DIR
-        if: matrix.clblast == 'ON'
-        run: |
-          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
-          7z x clblast.zip
-          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
-          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
-
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_OPENBLAS=${{ matrix.blas }}
+          -DGGML_OPENBLAS=${{ matrix.blas }}
           -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
-          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
@@ -447,19 +434,15 @@ 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
+        uses: actions/upload-artifact@v4
         with:
-          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
+          name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   windows-cublas:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     strategy:
       matrix:
@@ -476,14 +459,14 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1
+        uses: microsoft/setup-msbuild@v2
 
       - name: Install CUDA Toolkit
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.11
+        uses: Jimver/cuda-toolkit@v0.2.15
         with:
           cuda: '${{ matrix.cuda-toolkit }}'
 
@@ -498,7 +481,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUBLAS=${{ matrix.cublas }}
+          -DGGML_CUDA=${{ matrix.cublas }}
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build ${{ matrix.cuda-toolkit }}
@@ -519,7 +502,7 @@ jobs:
 
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v1
+        uses: actions/upload-artifact@v4
         with:
           name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
@@ -533,10 +516,10 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup emsdk
-        uses: mymindstorm/setup-emsdk@v12
+        uses: mymindstorm/setup-emsdk@v14
 
       - name: Verify
         run: emcc -v
@@ -555,7 +538,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Configure
         run: |
@@ -573,24 +556,24 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           path: whisper
 
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           repository: ggerganov/ggml
           path: ggml
 
       - name: Install Java
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v4
         with:
           distribution: zulu
-          java-version: 17
+          java-version: 21
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
+        uses: android-actions/setup-android@v3
 
       - name: Build
         run: |
@@ -608,20 +591,19 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: set up JDK 11
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v4
         with:
           java-version: '11'
           distribution: 'temurin'
           cache: gradle
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
+        uses: android-actions/setup-android@v3
         with:
-          api-level: 30
-          build-tools-version: 30.0.3
+          cmdline-tools-version: 9.0
 
       - name: Build
         run: |
@@ -633,15 +615,16 @@ jobs:
     needs: [ 'windows' ]
     runs-on: windows-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
 
       - name: Install Java
-        uses: actions/setup-java@v1
+        uses: actions/setup-java@v4
         with:
-          java-version: 17
+          distribution: zulu
+          java-version: 20
 
       - name: Download Windows lib
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
         with:
           name: win32-x86-64_whisper.dll
           path: bindings/java/build/generated/resources/main/win32-x86-64
@@ -654,7 +637,7 @@ jobs:
           ./gradlew build
 
       - name: Upload jar
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: whispercpp.jar
           path: bindings/java/build/libs/whispercpp-*.jar
@@ -676,7 +659,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Test quantize
         run: |

.gitignore

@@ -10,16 +10,7 @@
 /CMakeSettings.json
 
 build/
-build-coreml/
-build-em/
-build-debug/
-build-release/
-build-rwdi/
-build-static/
-build-cublas/
-build-no-accel/
-build-sanitize-addr/
-build-sanitize-thread/
+build-*/
 
 # SPM
 .build/

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "bindings/ios"]
-	path = bindings/ios
-	url = https://github.com/ggerganov/whisper.spm

CMakeLists.txt

@@ -1,25 +1,31 @@
-cmake_minimum_required (VERSION 3.5)
+cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
+project("whisper.cpp" C CXX)
+project("whisper.cpp" VERSION 1.6.2)
+include(CheckIncludeFileCXX)
 
-# Allow for the creation of solution folders.
-set_property(GLOBAL PROPERTY USE_FOLDERS ON)
-
-project(whisper.cpp VERSION 1.6.0)
 set(SOVERSION 1)
 
+#set(CMAKE_WARN_DEPRECATED YES)
+set(CMAKE_WARN_UNUSED_CLI YES)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(WHISPER_STANDALONE ON)
-    include(GitVars)
-    include(BuildTypes)
+
+    include(git-vars)
 
     # configure project version
-    if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
-        configure_file(${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl ${CMAKE_SOURCE_DIR}/bindings/ios/Makefile @ONLY)
-    endif()
     configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
 else()
     set(WHISPER_STANDALONE OFF)
@@ -29,6 +35,11 @@ if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
     option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
+
+    # TODO: without these, we get the following error:
+    #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
 else()
     if (MINGW)
         set(BUILD_SHARED_LIBS_DEFAULT OFF)
@@ -37,728 +48,145 @@ else()
     endif()
 endif()
 
-# options
+option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 
-if (APPLE)
-    set(WHISPER_METAL_DEFAULT ON)
-else()
-    set(WHISPER_METAL_DEFAULT OFF)
-endif()
+#
+# option list
+#
 
-option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
+# general
+option(WHISPER_CCACHE "whisper: use ccache if available" ON)
 
+# debug
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
 option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
 
-option(WHISPER_SANITIZE_THREAD        "whisper: enable thread sanitizer"    OFF)
-option(WHISPER_SANITIZE_ADDRESS       "whisper: enable address sanitizer"   OFF)
-option(WHISPER_SANITIZE_UNDEFINED     "whisper: enable undefined sanitizer" OFF)
-
-option(WHISPER_BUILD_TESTS            "whisper: build tests"    ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDALONE})
-
-option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
-
-option(WHISPER_NO_AVX                 "whisper: disable AVX"         OFF)
-option(WHISPER_NO_AVX2                "whisper: disable AVX2"        OFF)
-option(WHISPER_NO_AVX512              "whisper: disable AVX512"      ON)
-option(WHISPER_NO_AVX512_VBMI         "whisper: disable AVX512-VBMI" ON)
-option(WHISPER_NO_AVX512_VNNI         "whisper: disable AVX512-VNNI" ON)
-option(WHISPER_NO_FMA                 "whisper: disable FMA"         OFF)
-option(WHISPER_NO_F16C                "whisper: disable F16c"        OFF)
-
-option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
-
-if (APPLE)
-    option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
-    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
-    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
-    option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
-    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
-    option(WHISPER_METAL_EMBED_LIBRARY   "whisper: embed Metal library"          OFF)
-else()
-    option(WHISPER_BLAS                  "whisper: use BLAS libraries"                        OFF)
-    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor"                       Generic)
-    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"                           OFF)
-    option(WHISPER_OPENBLAS_INTERFACE64  "whisper: use OpenBLAS w/ 64-bit interface"          OFF)
-    option(WHISPER_CUDA                  "whisper: support for CUDA"                          OFF)
-    option(WHISPER_CUBLAS                "whisper: support for CUDA (deprecated)"             OFF)
-    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
-    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               OFF)
-    option(WHISPER_MKL                   "whisper: use Intel Math Kernel Library (MKL)"       OFF)
-    option(WHISPER_SYCL                  "whisper: use SYCL"                                  OFF)
-    option(WHISPER_SYCL_F16              "whisper: use 16 bit floats for sycl calculations"   OFF)
-endif()
-
-option(WHISPER_PERF "whisper: enable perf timings" OFF)
+# build
+option(WHISPER_FATAL_WARNINGS "whisper: enable -Werror flag" OFF)
 
 # sanitizers
+option(WHISPER_SANITIZE_THREAD    "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS   "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED "whisper: enable undefined sanitizer" OFF)
 
-if (NOT MSVC)
-    if (WHISPER_SANITIZE_THREAD)
-        set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
-    endif()
+# extra artifacts
+option(WHISPER_BUILD_TESTS    "whisper: build tests"          ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES "whisper: build examples"       ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_SERVER   "whisper: build server example" ${WHISPER_STANDALONE})
 
-    if (WHISPER_SANITIZE_ADDRESS)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=address -fno-omit-frame-pointer")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
-    endif()
+# 3rd party libs
+option(WHISPER_CURL "whisper: use libcurl to download model from an URL" OFF)
+option(WHISPER_SDL2 "whisper: support for libSDL2" OFF)
 
-    if (WHISPER_SANITIZE_UNDEFINED)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=undefined")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
-    endif()
-endif()
-
-#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
-#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
-
-# dependencies
-
-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
-    if (NOT WHISPER_NO_ACCELERATE)
-        find_library(ACCELERATE_FRAMEWORK Accelerate)
-
-        if (ACCELERATE_FRAMEWORK)
-            message(STATUS "Accelerate framework found")
-
-            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64)
-        else()
-            message(FATAL_ERROR "Accelerate framework not found")
-        endif()
-    endif()
-
-    if (WHISPER_METAL)
-        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
-        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
-        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
-
-        if (METAL_FRAMEWORK)
-            message(STATUS "Metal framework found")
-
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
-                ${FOUNDATION_LIBRARY}
-                ${METAL_FRAMEWORK}
-                ${METALKIT_FRAMEWORK}
-                )
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
-
-            if (WHISPER_METAL_NDEBUG)
-                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
-            endif()
-        else()
-            message(FATAL_ERROR "Metal framework not found")
-        endif()
-
-        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
-
-        # copy ggml-common.h and ggml-metal.metal to bin directory
-        configure_file(ggml-common.h    bin/ggml-common.h    COPYONLY)
-        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
-
-        if (WHISPER_METAL_EMBED_LIBRARY)
-            enable_language(ASM)
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)
-
-            set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-            set(COMMON_HEADER   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-
-            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-            set(EMBED_METALLIB_SOURCE "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-combined.metal")
-
-            add_custom_command(
-                OUTPUT ${EMBED_METALLIB_SOURCE}
-                COMMAND sed -e "/^#include \\\"ggml-common.h\\\"/r ${COMMON_HEADER}" -e "/^#include \\\"ggml-common.h\\\"/d" ${METALLIB_SOURCE} > ${EMBED_METALLIB_SOURCE}
-                DEPENDS ${METALLIB_SOURCE} ${COMMON_HEADER}
-                COMMENT "Generating combined Metal library for embedding"
-            )
-
-            add_custom_command(
-                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".incbin \\\"${EMBED_METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-                DEPENDS ${EMBED_METALLIB_SOURCE}
-                COMMENT "Generate assembly for embedded Metal library"
-            )
-
-            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-        endif()
-    endif()
-
-    if (WHISPER_COREML)
-        find_library(FOUNDATION_FRAMEWORK Foundation)
-        find_library(COREML_FRAMEWORK CoreML)
-
-        if (COREML_FRAMEWORK)
-            message(STATUS "CoreML framework found")
-
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
-        else()
-            message(FATAL_ERROR "CoreML framework not found")
-        endif()
-
-        if (WHISPER_COREML_ALLOW_FALLBACK)
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
-        endif()
-    endif()
-endif()
-
-if (WHISPER_OPENBLAS)
-    set(WHISPER_BLAS_VENDOR "OpenBLAS")
-    set(WHISPER_BLAS ON)
-    # BLA_PKGCONFIG_BLAS is supported since CMake 3.25.
-    # FindBLAS.cmake pkg-config logic seems incomplete, because when
-    # BLA_SIZEOF_INTEGER is 8, then it should search for blas64 instead of blas.
-    # blas.pc/blas64.pc are not always provided, so let's be more specific
-    # and go with openblas.pc/openblas64.pc if WHISPER_OPENBLAS is on.
-    if (WHISPER_OPENBLAS_INTERFACE64)
-        set(WHISPER_BLAS_LIB "openblas64")
-    else ()
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
-    set(BLA_PKGCONFIG_BLAS ${WHISPER_BLAS_LIB})
-    # OpenBLAS prebuilt libraries for Windows do not have "64" suffix in filename.
-    # (But .pc file has "64" suffix in filename for USE_64BITINT=1 Windows build.)
-    if (MSVC)
-        set(WHISPER_BLAS_LIB "openblas")
-    endif ()
-endif()
-
-if (WHISPER_BLAS)
-    if (NOT "$ENV{OPENBLAS_PATH}" STREQUAL "")
-        if (WHISPER_STATIC)
-            set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_STATIC_LIBRARY_PREFIX})
-            set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_STATIC_LIBRARY_SUFFIX})
-        else ()
-            if (CMAKE_IMPORT_LIBRARY_SUFFIX)
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_IMPORT_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_IMPORT_LIBRARY_SUFFIX})
-            else ()
-                set(WHISPER_BLAS_LIB_PREFIX ${CMAKE_SHARED_LIBRARY_PREFIX})
-                set(WHISPER_BLAS_LIB_SUFFIX ${CMAKE_SHARED_LIBRARY_SUFFIX})
-            endif ()
-        endif ()
-        # OpenBLAS prebuilt libraries hardcode "lib" prefix in filename even on Windows
-        if (WHISPER_OPENBLAS)
-            set(WHISPER_BLAS_LIB_PREFIX "lib")
-        endif ()
-        message(STATUS "BLAS compatible library path provided")
-        set(BLAS_LIBRARIES "$ENV{OPENBLAS_PATH}/lib/${WHISPER_BLAS_LIB_PREFIX}${WHISPER_BLAS_LIB}${WHISPER_BLAS_LIB_SUFFIX}")
-        message(STATUS "Libraries ${BLAS_LIBRARIES}")
-        set(BLAS_INCLUDE_DIRS "$ENV{OPENBLAS_PATH}/include")
-        message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
-        if (NOT EXISTS "${BLAS_LIBRARIES}")
-            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH misdefined.")
-        endif ()
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-        include_directories(${BLAS_INCLUDE_DIRS})
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
-    else ()
-        if (WHISPER_STATIC)
-            # FindBLAS.cmake pkg-config logic seems incomplete, because when
-            # BLA_STATIC is on, then it should use pkg_check_modules_static
-            # instead of pkg_check_modules.
-            # Some manual variable overriding may be necessary if you don't
-            # achieve desired results.
-            set(BLA_STATIC 1)
-        endif ()
-        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
-        if (WHISPER_OPENBLAS_INTERFACE64)
-            set(BLA_SIZEOF_INTEGER 8)
-        else ()
-            set(BLA_SIZEOF_INTEGER 4)
-        endif()
-        set(BLA_PREFER_PKGCONFIG 1)
-        find_package(BLAS)
-
-        if(BLAS_FOUND)
-            message(STATUS "BLAS compatible library found")
-            message(STATUS "Libraries ${BLAS_LIBRARIES}")
-            if (NOT DEFINED BLAS_INCLUDE_DIRS)
-                if (PKGC_BLAS_FOUND)
-                    set(BLAS_INCLUDE_DIRS "${PKGC_BLAS_INCLUDE_DIRS}")
-                else ()
-                    find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas)
-                endif()
-            endif()
-            message(STATUS "Include dirs ${BLAS_INCLUDE_DIRS}")
-            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-            include_directories(${BLAS_INCLUDE_DIRS})
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
-        else()
-            message(FATAL_ERROR "BLAS library was not found")
-        endif()
-    endif ()
-endif ()
-
-if (WHISPER_MKL)
-    find_package(MKL CONFIG REQUIRED PATHS $ENV{MKLROOT})
-    message(STATUS "Imported oneMKL targets: ${MKL_IMPORTED_TARGETS}")
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_BLAS_USE_MKL)
-endif()
-
-if (WHISPER_CUBLAS)
-    message(WARNING "WHISPER_CUBLAS is deprecated and will be removed in the future.\nUse WHISPER_CUDA instead")
-    set(WHISPER_CUDA ON)
-endif()
-
-if (WHISPER_CUDA)
-    cmake_minimum_required(VERSION 3.17)
-
-    find_package(CUDAToolkit)
-
-    if (CUDAToolkit_FOUND)
-        message(STATUS "cuBLAS found")
-
-        enable_language(CUDA)
-
-        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.h)
-        list(APPEND GGML_SOURCES_CUDA  ggml-cuda.cu)
-
-        add_compile_definitions(GGML_USE_CUDA)
-
-        if (WHISPER_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            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()
-endif()
-
-
-if (WHISPER_HIPBLAS)
-    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
-    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
-    endif()
-    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-    endif()
-
-    find_package(hip)
-    find_package(hipblas)
-    find_package(rocblas)
-
-    if (${hipblas_FOUND} AND ${hip_FOUND})
-        message(STATUS "HIP and hipBLAS found")
-        set(GGML_HEADERS_ROCM "ggml-cuda.h")
-
-        file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
-
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-        if (WHISPER_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
-    else()
-        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
-    endif()
-endif()
-
-if (WHISPER_CLBLAST)
-    find_package(CLBlast)
-    if (CLBlast_FOUND)
-        message(STATUS "CLBlast found")
-
-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
-
-        add_compile_definitions(GGML_USE_CLBLAST)
-
-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
-    else()
-        message(FATAL_ERROR "CLBlast not found")
-    endif()
-endif()
-
-if( WHISPER_OPENVINO )
-    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
-endif()
-
-if (WHISPER_SYCL)
-    if ( NOT DEFINED ENV{ONEAPI_ROOT})
-        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
-    endif()
-    #todo: AOT
-
-    find_package(IntelSYCL REQUIRED)
-    if (WHISPER_SYCL_F16)
-        add_compile_definitions(GGML_SYCL_F16)
-    endif()
-    add_compile_definitions(GGML_USE_SYCL)
-
-    add_compile_options(-I./) #include DPCT
-    add_compile_options(-I/${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 ${WHISPER_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-endif()
-# compiler flags
-
-if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
-    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
-    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
-endif ()
-
-if (WHISPER_ALL_WARNINGS)
-    if (NOT MSVC)
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} \
-            -Wall                           \
-            -Wextra                         \
-            -Wpedantic                      \
-            -Wshadow                        \
-            -Wcast-qual                     \
-            -Wstrict-prototypes             \
-            -Wpointer-arith                 \
-            -Wno-unused-function            \
-        ")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
-            -Wall                           \
-            -Wextra                         \
-            -Wpedantic                      \
-            -Wcast-qual                     \
-        ")
-    else()
-        # todo : msvc
-    endif()
-endif()
-
-if (NOT MSVC)
-    # TODO: temporary disabled until we figure out ggml-metal.m
-    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
-    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
-endif()
-
-message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
-
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
-    message(STATUS "ARM detected")
-elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-    message(STATUS "PowerPC detected")
-else()
-    message(STATUS "x86 detected")
-    if (MSVC)
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
-        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
-        if(NOT WHISPER_NO_AVX512)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX512")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX512")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX512")
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (NOT WHISPER_NO_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
-            endif()
-            if (NOT WHISPER_NO_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
-            endif()
-        elseif(NOT WHISPER_NO_AVX2)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-        elseif(NOT WHISPER_NO_AVX)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
-            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
-        endif()
-    else()
-        if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
-        else()
-            if(NOT WHISPER_NO_AVX)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
-            endif()
-            if(NOT WHISPER_NO_AVX2)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
-            endif()
-            if(NOT WHISPER_NO_AVX512)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -mavx512vl -mavx512dq -mavx512bw")
-                if(NOT WHISPER_NO_AVX512_VBMI)
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vbmi")
-                endif()
-                if(NOT WHISPER_NO_AVX512_VNNI)
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512vnni")
-                endif()
-            endif()
-            if(NOT WHISPER_NO_FMA)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
-            endif()
-            if(NOT WHISPER_NO_F16C)
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
-            endif()
-        endif()
-    endif()
-endif()
-
-#
-# POSIX conformance
-#
-
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-add_compile_definitions(_XOPEN_SOURCE=600)
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    remove_definitions(-D_XOPEN_SOURCE=600)
-    add_compile_definitions(_XOPEN_SOURCE=700)
-endif()
-
-# Data types, macros and functions related to controlling CPU affinity
-# are available on Linux through GNU extensions in libc
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_compile_definitions(_GNU_SOURCE)
+    option(WHISPER_FFMPEG "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
 endif()
 
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
-    add_compile_definitions(_DARWIN_C_SOURCE)
-endif()
+option(WHISPER_COREML                "whisper: enable Core ML framework"  OFF)
+option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
+option(WHISPER_OPENVINO              "whisper: support for OpenVINO"      OFF)
 
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
-    add_compile_definitions(__BSD_VISIBLE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
-    add_compile_definitions(_NETBSD_SOURCE)
-endif()
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    add_compile_definitions(_BSD_SOURCE)
-endif()
+# Required for relocatable CMake package
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
-if (WHISPER_PERF)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
-endif()
+# override ggml options
+set(GGML_CCACHE             ${WHISPER_CCACHE})
+set(GGML_SANITIZE_THREAD    ${WHISPER_SANITIZE_THREAD})
+set(GGML_SANITIZE_ADDRESS   ${WHISPER_SANITIZE_ADDRESS})
+set(GGML_SANITIZE_UNDEFINED ${WHISPER_SANITIZE_UNDEFINED})
+set(GGML_ALL_WARNINGS       ${WHISPER_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS     ${WHISPER_FATAL_WARNINGS})
 
-#
-# whisper.coreml - Core ML support
-#
-
-if (WHISPER_COREML)
-    set(TARGET whisper.coreml)
-
-    add_library(${TARGET}
-        coreml/whisper-encoder.h
-        coreml/whisper-encoder.mm
-        coreml/whisper-encoder-impl.h
-        coreml/whisper-encoder-impl.m
-        )
-
-    include(DefaultTargetOptions)
-
-    target_include_directories(${TARGET} PUBLIC
-        .
-        )
-
-    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
-
-    set_target_properties(${TARGET} PROPERTIES
-        COMPILE_FLAGS "-fobjc-arc"
-        )
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
-endif()
-
-if (WHISPER_OPENVINO)
-    set(TARGET whisper.openvino)
-
-    add_library(${TARGET} OBJECT
-        openvino/whisper-openvino-encoder.h
-        openvino/whisper-openvino-encoder.cpp
-        )
-
-    target_include_directories(${TARGET} PUBLIC
-        .
-        )
-
-    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
-
-    target_link_libraries(${TARGET} PRIVATE openvino::runtime)
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
-endif()
-
-#
-# whisper - this is the main library of the project
-#
-
-set(TARGET whisper)
-
-add_library(${TARGET}
-    ggml.h
-    ggml.c
-    ggml-alloc.h
-    ggml-alloc.c
-    ggml-backend.h
-    ggml-backend.c
-    ggml-quants.h
-    ggml-quants.c
-    ${GGML_SOURCES_METAL}
-    ${GGML_SOURCES_CUDA}
-    ${GGML_SOURCES_OPENCL}
-    ${GGML_SOURCES_SYCL}        ${GGML_HEADERS_SYCL}
-    ${GGML_SOURCES_ROCM}        ${GGML_HEADERS_ROCM}
-    whisper.h
-    whisper.cpp
-    )
-
-include_directories (
-    .
-)
-# Set the version numbers
-set_target_properties(whisper PROPERTIES
-    VERSION ${PROJECT_VERSION}
-    SOVERSION ${SOVERSION}
-)
-
-include(DefaultTargetOptions)
-
-target_include_directories(${TARGET} PUBLIC
-    .
-    )
-
-if (WHISPER_COREML)
-    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
-endif()
-
-if (WHISPER_OPENVINO)
-    target_link_libraries(${TARGET} PRIVATE whisper.openvino)
-endif()
-
-if (WHISPER_MKL)
-    target_link_libraries(${TARGET} PUBLIC MKL::MKL)
-endif()
-
-if (MSVC)
-    target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
-
-    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
-else()
-    target_link_libraries(${TARGET} PRIVATE m ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
-endif()
-
-if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_link_libraries(${TARGET} PUBLIC
-        ${CMAKE_DL_LIBS}
-        )
-
-    target_compile_definitions(${TARGET} PUBLIC
-        WHISPER_SHARED
-        GGML_SHARED
-        )
-
-    target_compile_definitions(${TARGET} PRIVATE
-        WHISPER_BUILD
-        GGML_BUILD
-        )
-
-    if (WHISPER_METAL)
-        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
-        #       but for some reason it does not work here like it does in llama.cpp
-        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+# transition helpers
+function (whisper_option_depr TYPE OLD NEW)
+    if (${OLD})
+        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
+        set(${NEW} ON)
     endif()
-endif()
+endfunction()
 
-if (GGML_SOURCES_CUDA)
-    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    # Only configure gmml CUDA architectures is not globally set
-    if (NOT DEFINED GGML_CUDA_ARCHITECTURES)
-        # Not overriden by user, so set defaults
-        set(GGML_CUDA_ARCHITECTURES 52 61 70)
-    endif()
-    message(STATUS "GGML Configuring CUDA architectures ${GGML_CUDA_ARCHITECTURES}")
-    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES ${GGML_CUDA_ARCHITECTURES})
-    set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-endif()
+whisper_option_depr(FATAL_ERROR WHISPER_CUBLAS              GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_CUDA                GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_KOMPUTE             GGML_KOMPUTE)
+whisper_option_depr(WARNING     WHISPER_METAL               GGML_METAL)
+whisper_option_depr(WARNING     WHISPER_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
+whisper_option_depr(WARNING     WHISPER_NATIVE              GGML_NATIVE)
+whisper_option_depr(WARNING     WHISPER_OPENMP              GGML_OPENMP)
+whisper_option_depr(WARNING     WHISPER_RPC                 GGML_RPC)
+whisper_option_depr(WARNING     WHISPER_SYCL                GGML_SYCL)
+whisper_option_depr(WARNING     WHISPER_SYCL_F16            GGML_SYCL_F16)
 
-if (EMSCRIPTEN)
-    set_target_properties(${TARGET} PROPERTIES COMPILE_FLAGS "-msimd128")
-endif()
+#
+# build the library
+#
 
-target_compile_definitions(${TARGET} PUBLIC
-    ${WHISPER_EXTRA_FLAGS}
-    )
+add_subdirectory(ggml)
+add_subdirectory(src)
 
-set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+#
+# install
+#
 
 include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
 
-install(TARGETS ${TARGET}
-    LIBRARY  DESTINATION lib
-    ARCHIVE  DESTINATION lib/static
-    RUNTIME  DESTINATION bin
-    RESOURCE DESTINATION bin
-    PUBLIC_HEADER DESTINATION include
-    )
+set(WHISPER_BUILD_NUMBER        ${BUILD_NUMBER})
+set(WHISPER_BUILD_COMMIT        ${BUILD_COMMIT})
+set(WHISPER_INSTALL_VERSION     ${CMAKE_PROJECT_VERSION})
 
-#
-# bindings
-#
+set(WHISPER_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(WHISPER_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(WHISPER_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
-add_subdirectory(bindings)
+get_directory_property(WHISPER_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+
+set_target_properties(whisper PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/whisper.h)
+install(TARGETS whisper LIBRARY PUBLIC_HEADER)
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/whisper-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper
+    PATH_VARS
+    WHISPER_INCLUDE_INSTALL_DIR
+    WHISPER_LIB_INSTALL_DIR
+    WHISPER_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+    ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+    VERSION ${WHISPER_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper)
+
+install(
+    FILES convert-hf-to-gguf.py
+    PERMISSIONS
+        OWNER_READ
+        OWNER_WRITE
+        OWNER_EXECUTE
+        GROUP_READ
+        GROUP_EXECUTE
+        WORLD_READ
+        WORLD_EXECUTE
+    DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+configure_file(cmake/whisper.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        DESTINATION lib/pkgconfig)
 
 #
 # programs, examples and tests
 #
 
 if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    enable_testing()
-    add_subdirectory(tests)
+    #include(CTest)
+    #add_subdirectory(tests)
 endif ()
 
 if (WHISPER_BUILD_EXAMPLES)

Package.swift

@@ -27,17 +27,15 @@ let package = Package(
                "samples",
                "tests",
                "CMakeLists.txt",
-               "ggml-cuda.cu",
-               "ggml-cuda.h",
                "Makefile"
             ],
             sources: [
-                "ggml.c",
-                "whisper.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m"
+                "ggml/src/ggml.c",
+                "src/whisper.cpp",
+                "ggml/src/ggml-alloc.c",
+                "ggml/src/ggml-backend.c",
+                "ggml/src/ggml-quants.c",
+                "ggml/src/ggml-metal.m"
             ],
             resources: [.process("ggml-metal.metal")],
             publicHeadersPath: "spm-headers",

README.md

@@ -4,9 +4,10 @@
 
 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.6.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.6.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -19,7 +20,6 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - Zero memory allocations at runtime
 - Support for CPU-only inference
 - [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
-- [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
 - [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
@@ -418,31 +418,9 @@ Now build `whisper.cpp` with CUDA support:
 
 ```
 make clean
-WHISPER_CUDA=1 make -j
+GGML_CUDA=1 make -j
 ```
 
-## OpenCL GPU support via CLBlast
-
-For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APUs or low end devices for up to ~2x speedup.
-
-First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
-
-Now build `whisper.cpp` with CLBlast support:
-
-```
-Makefile:
-cd whisper.cpp
-make clean
-WHISPER_CLBLAST=1 make -j
-
-CMake:
-cd whisper.cpp
-cmake -B build -DWHISPER_CLBLAST=ON
-cmake --build build -j --config Release
-```
-
-Run all the examples as usual.
-
 ## BLAS CPU support via OpenBLAS
 
 Encoder processing can be accelerated on the CPU via OpenBLAS.
@@ -452,7 +430,7 @@ Now build `whisper.cpp` with OpenBLAS support:
 
 ```
 make clean
-WHISPER_OPENBLAS=1 make -j
+GGML_OPENBLAS=1 make -j
 ```
 
 ## BLAS CPU support via Intel MKL
@@ -502,6 +480,16 @@ docker run -it --rm \
   whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
 ```
 
+## Installing with Conan
+
+You can install pre-built binaries for whisper.cpp or build it from source using [Conan](https://conan.io/). Use the following command:
+
+```
+conan install --requires="whisper-cpp/[*]" --build=missing
+```
+
+For detailed instructions on how to use Conan, please refer to the [Conan documentation](https://docs.conan.io/2/).
+
 ## Limitations
 
 - Inference only
@@ -710,7 +698,7 @@ The [main](examples/main) example provides support for output of karaoke-style m
 currently pronounced word is highlighted. Use the `-wts` argument and run the generated bash script.
 This requires to have `ffmpeg` installed.
 
-Here are a few *"typical"* examples:
+Here are a few _"typical"_ examples:
 
 ```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts

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

@@ -68,10 +68,6 @@ func (flags *Flags) GetOut() string {
 	return strings.ToLower(flags.Lookup("out").Value.String())
 }
 
-func (flags *Flags) IsSpeedup() bool {
-	return flags.Lookup("speedup").Value.String() == "true"
-}
-
 func (flags *Flags) IsTokens() bool {
 	return flags.Lookup("tokens").Value.String() == "true"
 }
@@ -111,10 +107,6 @@ func (flags *Flags) SetParams(context whisper.Context) error {
 		fmt.Fprintf(flags.Output(), "Setting duration to %v\n", duration)
 		context.SetDuration(duration)
 	}
-	if flags.IsSpeedup() {
-		fmt.Fprintf(flags.Output(), "Setting speedup to true\n")
-		context.SetSpeedup(true)
-	}
 	if threads := flags.GetThreads(); threads != 0 {
 		fmt.Fprintf(flags.Output(), "Setting threads to %d\n", threads)
 		context.SetThreads(threads)
@@ -146,7 +138,6 @@ func registerFlags(flag *Flags) {
 	flag.Duration("offset", 0, "Time offset")
 	flag.Duration("duration", 0, "Duration of audio to process")
 	flag.Uint("threads", 0, "Number of threads to use")
-	flag.Bool("speedup", false, "Enable speedup")
 	flag.Uint("max-len", 0, "Maximum segment length in characters")
 	flag.Uint("max-tokens", 0, "Maximum tokens per segment")
 	flag.Float64("word-thold", 0, "Maximum segment score")

bindings/go/params.go

@@ -47,10 +47,6 @@ func (p *Params) SetPrintTimestamps(v bool) {
 	p.print_timestamps = toBool(v)
 }
 
-func (p *Params) SetSpeedup(v bool) {
-	p.speed_up = toBool(v)
-}
-
 // Set language id
 func (p *Params) SetLanguage(lang int) error {
 	if lang == -1 {
@@ -177,9 +173,6 @@ func (p *Params) String() string {
 	if p.token_timestamps {
 		str += " token_timestamps"
 	}
-	if p.speed_up {
-		str += " speed_up"
-	}
 
 	return str + ">"
 }

bindings/go/pkg/whisper/context.go

@@ -76,11 +76,6 @@ func (context *context) SetTranslate(v bool) {
 	context.params.SetTranslate(v)
 }
 
-// Set speedup flag
-func (context *context) SetSpeedup(v bool) {
-	context.params.SetSpeedup(v)
-}
-
 func (context *context) SetSplitOnWord(v bool) {
 	context.params.SetSplitOnWord(v)
 }

bindings/go/pkg/whisper/interface.go

@@ -41,7 +41,6 @@ type Context interface {
 	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

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

@@ -20,7 +20,7 @@ public interface WhisperCppJnaLibrary extends Library {
      * @return Whisper context on success, null on failure
      */
     Pointer whisper_init_from_file(String path_model);
-    
+
     /**
      * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
      * Because this function allocates memory for the params, the caller must call either:
@@ -304,14 +304,6 @@ public interface WhisperCppJnaLibrary extends Library {
     /** Language id associated with the provided state */
     int whisper_full_lang_id_from_state(Pointer state);
 
-    /**
-     * Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
-     * The resulting spectrogram is stored inside the default state of the provided whisper context.
-     * @return 0 on success
-     */
-    int whisper_pcm_to_mel_phase_vocoder(Pointer ctx, final float[] samples, int n_samples, int n_threads);
-
-    int whisper_pcm_to_mel_phase_vocoder_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);
 
     /** Get the start time of the specified segment. */
     long whisper_full_get_segment_t0(Pointer ctx, int i_segment);

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

@@ -129,14 +129,6 @@ public class WhisperFullParams extends Structure {
     /** Maximum tokens per segment (0, default = no limit) */
     public int max_tokens;
 
-    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
-    public CBool speed_up;
-
-    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
-    public void speedUp(boolean enable) {
-        speed_up = enable ? CBool.TRUE : CBool.FALSE;
-    }
-
     /** Overwrite the audio context size (0 = use default). */
     public int audio_ctx;
 
@@ -321,7 +313,7 @@ public class WhisperFullParams extends Structure {
         return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
                 "no_context", "single_segment", "no_timestamps",
                 "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
-                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "audio_ctx",
                 "tdrz_enable", "suppress_regex", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
                 "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
                 "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.6.0",
+  "version": "1.6.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/Rakefile

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

bindings/ruby/ext/extconf.rb

@@ -1,6 +1,7 @@
 require 'mkmf'
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper-mel.hpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")

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

@@ -12,31 +12,63 @@ extern "C" {
     // Backend buffer
     //
 
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
+        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+        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 {
-        struct ggml_backend_buffer_i iface;
-
-        ggml_backend_t                backend;
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
         ggml_backend_buffer_context_t context;
-
         size_t size;
+        enum ggml_backend_buffer_usage usage;
     };
 
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
+    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
     //
@@ -44,44 +76,66 @@ 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_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
 
-        // get buffer alignment
-        size_t (*get_alignment)(ggml_backend_t backend);
+        // (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);
 
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
-        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
+        // (optional) complete all pending operations
+        void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        // compute graph with a plan (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
-        bool (*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);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
+        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
     };
 
     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_CALL ggml_backend_init_fn)(const char * params, void * user_data);
+
+    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
 #ifdef  __cplusplus
 }
 #endif

bindings/ruby/ext/ggml-backend.h

@@ -7,69 +7,123 @@
 extern "C" {
 #endif
 
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
     //
     // Backend buffer
     //
 
-    struct ggml_backend_buffer;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    // buffer type
+    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
+    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
-    // backend buffer functions
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    // 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
     //
 
-    struct ggml_backend;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
-
+    GGML_API 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_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    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 size_t ggml_backend_get_alignment(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_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API 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 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);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
 
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
+
     //
     // CPU backend
     //
 
     GGML_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_API 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(ggml_backend_t backend_cpu, 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_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
 
     //
     // Backend scheduler
@@ -83,53 +137,96 @@ 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 assigned
+        // 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);
+
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+
+    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
+
+    // Reset all assignments and allocators - must be called before changing the node backends
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+
+    // Set a callback to be called for each resulting node during graph compute
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
 
-    // Allocate a graph on the backend scheduler
-    GGML_API void ggml_backend_sched_graph_compute(
-            ggml_backend_sched_t sched,
-            struct ggml_cgraph * graph);
 
 #ifdef  __cplusplus
 }

bindings/ruby/ext/ggml-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__))
@@ -34,16 +37,17 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
 #endif
-
-#undef MIN
-#undef MAX
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
 
 // 16-bit float
 // on Arm, we use __fp16
@@ -56,14 +60,30 @@ 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
@@ -217,8 +237,7 @@ 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));
@@ -226,19 +245,23 @@ 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
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
-// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // return index, asserts if table is full

bindings/ruby/ext/ggml-quants.h

@@ -1,224 +1,133 @@
 #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, int64_t k);
+void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
 
-void quantize_row_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, int64_t k);
+void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
 
-void quantize_row_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, int64_t k);
+void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
 
-void quantize_row_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, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 // 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, int64_t k);
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
-void dequantize_row_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, int64_t k);
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 // 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_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
+void iq3xs_init_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif
 
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

bindings/ruby/ext/ruby_whisper.cpp

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

bindings/ruby/tests/test_whisper.rb

@@ -117,13 +117,6 @@ class TestWhisper < Test::Unit::TestCase
     assert !@params.split_on_word
   end
 
-  def test_speed_up
-    @params.speed_up = true
-    assert @params.speed_up
-    @params.speed_up = false
-    assert !@params.speed_up
-  end
-
   def test_whisper
     @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
     params  = Whisper::Params.new

bindings/ruby/whispercpp.gemspec

@@ -0,0 +1,28 @@
+Gem::Specification.new do |s|
+  s.name    = "whispercpp"
+  s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
+  s.version = '1.3.0'
+  s.date    = '2024-05-14'
+  s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
+  s.email   = 'todd.fisher@gmail.com'
+  s.extra_rdoc_files = ['LICENSE', 'README.md']
+  
+  s.files = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"]
+
+  #### Load-time details
+  s.require_paths = ['lib','ext']
+  s.summary = %q{Ruby whisper.cpp bindings}
+  s.test_files = ["tests/test_whisper.rb"]
+  
+  s.extensions << 'ext/extconf.rb'
+  
+
+  #### Documentation and testing.
+  s.homepage = 'https://github.com/ggerganov/whisper.cpp'
+  s.rdoc_options = ['--main', '../../README.md']
+
+  
+    s.platform = Gem::Platform::RUBY
+  
+  s.licenses = ['MIT']
+end

cmake/BuildTypes.cmake

@@ -1,54 +0,0 @@
-# Add new build types
-
-# ReleaseGG - Release with enabled asserts
-
-SET(CMAKE_CXX_FLAGS_RELEASEGG
-    "-O3"
-    CACHE STRING "Flags used by the c++ compiler during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_C_FLAGS_RELEASEGG
-    "-O3"
-    CACHE STRING "Flags used by the compiler during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_EXE_LINKER_FLAGS_RELEASEGG
-    ""
-    CACHE STRING "Flags used for linking binaries during release builds with enabled asserts."
-    FORCE )
-SET(CMAKE_SHARED_LINKER_FLAGS_RELEASEGG
-    ""
-    CACHE STRING "Flags used by the shared libraries linker during release builds with enabled asserts."
-    FORCE )
-MARK_AS_ADVANCED(
-    CMAKE_CXX_FLAGS_RELEASEGG
-    CMAKE_C_FLAGS_RELEASEGG
-    CMAKE_EXE_LINKER_FLAGS_RELEASEGG
-    CMAKE_SHARED_LINKER_FLAGS_RELEASEGG )
-
-# RelWithDebInfoGG - RelWithDebInfo with enabled asserts
-
-SET(CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
-    "-O2 -g"
-    CACHE STRING "Flags used by the c++ compiler during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_C_FLAGS_RELWITHDEBINFOGG
-    "-O2 -g"
-    CACHE STRING "Flags used by the compiler during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
-    ""
-    CACHE STRING "Flags used for linking binaries during release builds with debug symbols and enabled asserts."
-    FORCE )
-SET(CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG
-    ""
-    CACHE STRING "Flags used by the shared libraries linker during release builds with debug symbols and enabled asserts."
-    FORCE )
-MARK_AS_ADVANCED(
-    CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
-    CMAKE_C_FLAGS_RELWITHDEBINFOGG
-    CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
-    CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG )
-
-if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
-    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
-    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo" "ReleaseGG" "RelWithDebInfoGG")
-endif()

cmake/FindFFmpeg.cmake

@@ -0,0 +1,163 @@
+# From
+# https://github.com/snikulov/cmake-modules/blob/master/FindFFmpeg.cmake
+#
+# vim: ts=2 sw=2
+# - Try to find the required ffmpeg components(default: AVFORMAT, AVUTIL, AVCODEC)
+#
+# Once done this will define
+#  FFMPEG_FOUND         - System has the all required components.
+#  FFMPEG_INCLUDE_DIRS  - Include directory necessary for using the required components headers.
+#  FFMPEG_LIBRARIES     - Link these to use the required ffmpeg components.
+#  FFMPEG_DEFINITIONS   - Compiler switches required for using the required ffmpeg components.
+#
+# For each of the components it will additionally set.
+#   - AVCODEC
+#   - AVDEVICE
+#   - AVFORMAT
+#   - AVFILTER
+#   - AVUTIL
+#   - POSTPROC
+#   - SWSCALE
+# the following variables will be defined
+#  <component>_FOUND        - System has <component>
+#  <component>_INCLUDE_DIRS - Include directory necessary for using the <component> headers
+#  <component>_LIBRARIES    - Link these to use <component>
+#  <component>_DEFINITIONS  - Compiler switches required for using <component>
+#  <component>_VERSION      - The components version
+#
+# Copyright (c) 2006, Matthias Kretz, <kretz@kde.org>
+# Copyright (c) 2008, Alexander Neundorf, <neundorf@kde.org>
+# Copyright (c) 2011, Michael Jansen, <kde@michael-jansen.biz>
+#
+# Redistribution and use is allowed according to the terms of the BSD license.
+# For details see the accompanying COPYING-CMAKE-SCRIPTS file.
+
+include(FindPackageHandleStandardArgs)
+
+# The default components were taken from a survey over other FindFFMPEG.cmake files
+if (NOT FFmpeg_FIND_COMPONENTS)
+  set(FFmpeg_FIND_COMPONENTS AVFORMAT AVCODEC AVUTIL SWRESAMPLE) 
+endif()
+
+#
+### Macro: set_component_found
+#
+# Marks the given component as found if both *_LIBRARIES AND *_INCLUDE_DIRS is present.
+#
+macro(set_component_found _component )
+  if (${_component}_LIBRARIES AND ${_component}_INCLUDE_DIRS)
+    message(DEBUG "  - ${_component} found.")
+    set(${_component}_FOUND TRUE)
+  else ()
+  message(DEBUG "  - ${_component} not found.")
+  endif ()
+endmacro()
+
+#
+### Macro: find_component
+#
+# Checks for the given component by invoking pkgconfig and then looking up the libraries and
+# include directories.
+#
+macro(find_component _component _pkgconfig _library _header)
+
+  if (NOT WIN32)
+     # use pkg-config to get the directories and then use these values
+     # in the FIND_PATH() and FIND_LIBRARY() calls
+     find_package(PkgConfig)
+     if (PKG_CONFIG_FOUND)
+       pkg_check_modules(PC_${_component} ${_pkgconfig})
+       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDEDIR}")
+       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDE_DIRS}")
+       message(STATUS "${PC_${_component}_CFLAGS}")
+     endif ()
+  endif (NOT WIN32)
+
+
+  find_path(${_component}_INCLUDE_DIRS ${_header}
+    HINTS
+      ${PC_${_component}_INCLUDEDIR}
+      ${PC_${_component}_INCLUDE_DIRS}
+    PATH_SUFFIXES
+      ffmpeg
+  )
+
+  # CMake's default is to search first for shared libraries and then for static libraries.
+  # Todo later: add option to prefer static libs over dynamic:
+  find_library(${_component}_LIBRARIES NAMES ${_library} lib${_library}.a  
+      HINTS
+      ${PC_${_component}_LIBDIR}
+      ${PC_${_component}_LIBRARY_DIRS}
+  )
+
+  set(${_component}_DEFINITIONS  ${PC_${_component}_CFLAGS_OTHER} CACHE STRING "The ${_component} CFLAGS.")
+  set(${_component}_VERSION      ${PC_${_component}_VERSION}      CACHE STRING "The ${_component} version number.")
+
+  set_component_found(${_component})
+
+  mark_as_advanced(
+    ${_component}_INCLUDE_DIRS
+    ${_component}_LIBRARIES
+    ${_component}_DEFINITIONS
+    ${_component}_VERSION)
+
+endmacro()
+
+
+# Check for cached results. If there are skip the costly part.
+if (NOT FFMPEG_LIBRARIES)
+
+  # Check for all possible component.
+  find_component(AVCODEC    libavcodec    avcodec  libavcodec/avcodec.h)
+  find_component(AVFORMAT   libavformat   avformat libavformat/avformat.h)
+  find_component(AVDEVICE   libavdevice   avdevice libavdevice/avdevice.h)
+  #find_component(AVRESAMPLE libavresample avresample libavresample/avresample.h) # old name for swresample
+  find_component(AVUTIL     libavutil     avutil   libavutil/avutil.h)
+  find_component(AVFILTER   libavfilter   avfilter libavfilter/avfilter.h)
+  find_component(SWSCALE    libswscale    swscale  libswscale/swscale.h)
+  find_component(POSTPROC   libpostproc   postproc libpostproc/postprocess.h)
+  find_component(SWRESAMPLE libswresample swresample libswresample/swresample.h)
+
+  # Check if the required components were found and add their stuff to the FFMPEG_* vars.
+  foreach (_component ${FFmpeg_FIND_COMPONENTS})
+    if (${_component}_FOUND)
+      # message(STATUS "Required component ${_component} present.")
+      set(FFMPEG_LIBRARIES   ${FFMPEG_LIBRARIES}   ${${_component}_LIBRARIES})
+      set(FFMPEG_DEFINITIONS ${FFMPEG_DEFINITIONS} ${${_component}_DEFINITIONS})
+      list(APPEND FFMPEG_INCLUDE_DIRS ${${_component}_INCLUDE_DIRS})
+    else ()
+      # message(STATUS "Required component ${_component} missing.")
+    endif ()
+  endforeach ()
+
+  # Build the include path with duplicates removed.
+  if (FFMPEG_INCLUDE_DIRS)
+    list(REMOVE_DUPLICATES FFMPEG_INCLUDE_DIRS)
+  endif ()
+
+  # cache the vars.
+  set(FFMPEG_INCLUDE_DIRS ${FFMPEG_INCLUDE_DIRS} CACHE STRING "The FFmpeg include directories." FORCE)
+  set(FFMPEG_LIBRARIES    ${FFMPEG_LIBRARIES}    CACHE STRING "The FFmpeg libraries." FORCE)
+  set(FFMPEG_DEFINITIONS  ${FFMPEG_DEFINITIONS}  CACHE STRING "The FFmpeg cflags." FORCE)
+
+  mark_as_advanced(FFMPEG_INCLUDE_DIRS
+                   FFMPEG_LIBRARIES
+                   FFMPEG_DEFINITIONS)
+
+endif ()
+
+# Now set the noncached _FOUND vars for the components.
+# whisper.cpp does not need SWSCALE
+foreach (_component AVCODEC AVDEVICE AVFORMAT AVRESAMPLE AVUTIL POSTPROCESS)
+  set_component_found(${_component})
+endforeach ()
+
+# Compile the list of required vars
+set(_FFmpeg_REQUIRED_VARS FFMPEG_LIBRARIES FFMPEG_INCLUDE_DIRS)
+foreach (_component ${FFmpeg_FIND_COMPONENTS})
+  list(APPEND _FFmpeg_REQUIRED_VARS ${_component}_LIBRARIES ${_component}_INCLUDE_DIRS)
+endforeach ()
+
+# Give a nice error message if some of the required vars are missing.
+find_package_handle_standard_args(FFmpeg DEFAULT_MSG ${_FFmpeg_REQUIRED_VARS})
+

cmake/build-info.cmake

@@ -0,0 +1,58 @@
+set(BUILD_NUMBER 0)
+set(BUILD_COMMIT "unknown")
+set(BUILD_COMPILER "unknown")
+set(BUILD_TARGET "unknown")
+
+# Look for git
+find_package(Git)
+if(NOT Git_FOUND)
+    find_program(GIT_EXECUTABLE NAMES git git.exe)
+    if(GIT_EXECUTABLE)
+        set(Git_FOUND TRUE)
+        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
+    else()
+        message(WARNING "Git not found. Build info will not be accurate.")
+    endif()
+endif()
+
+# Get the commit count and hash
+if(Git_FOUND)
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE HEAD
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_COMMIT ${HEAD})
+    endif()
+    execute_process(
+        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE COUNT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
+    )
+    if (RES EQUAL 0)
+        set(BUILD_NUMBER ${COUNT})
+    endif()
+endif()
+
+if(MSVC)
+    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+else()
+    execute_process(
+        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_COMPILER ${OUT})
+    execute_process(
+        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
+        OUTPUT_VARIABLE OUT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    set(BUILD_TARGET ${OUT})
+endif()

cmake/whisper-config.cmake.in

@@ -0,0 +1,65 @@
+set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
+set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
+set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
+set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
+
+set(GGML_BLAS       @GGML_BLAS@)
+set(GGML_CUDA       @GGML_CUDA@)
+set(GGML_METAL      @GGML_METAL@)
+set(GGML_HIPBLAS    @GGML_HIPBLAS@)
+set(GGML_ACCELERATE @GGML_ACCELERATE@)
+
+@PACKAGE_INIT@
+
+set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
+set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
+set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
+
+# Ensure transient dependencies satisfied
+
+find_package(Threads REQUIRED)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+endif()
+
+if (GGML_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (GGML_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (GGML_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+find_library(llama_LIBRARY llama
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
+set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
+
+add_library(llama UNKNOWN IMPORTED)
+
+set_target_properties(llama
+    PROPERTIES
+        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
+        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
+        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+        IMPORTED_LOCATION "${llama_LIBRARY}"
+        INTERFACE_COMPILE_FEATURES cxx_std_11
+        POSITION_INDEPENDENT_CODE ON )
+
+check_required_components(Llama)

cmake/whisper.pc.in

@@ -0,0 +1,10 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=${exec_prefix}/lib
+includedir=${prefix}/include
+
+Name: whisper
+Description: Port of OpenAI's Whisper model in C/C++
+Version: @PROJECT_VERSION@
+Libs: -L${libdir} -lwhisper
+Cflags: -I${includedir}

examples/CMakeLists.txt

@@ -11,7 +11,7 @@ if (WHISPER_SDL2)
     string(STRIP "${SDL2_LIBRARIES}" SDL2_LIBRARIES)
 
     message(STATUS "SDL2_INCLUDE_DIRS = ${SDL2_INCLUDE_DIRS}")
-    message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
+    message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
 endif()
 
 if (WHISPER_CLBLAST)
@@ -22,6 +22,35 @@ endif()
 
 set(TARGET common)
 
+unset(COMMON_EXTRA_LIBS)
+
+if (WHISPER_FFMPEG)
+    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
+    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
+    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
+    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
+    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
+    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
+    find_package(FFmpeg REQUIRED)
+
+    if (NOT ${FFMPEG_FOUND})
+        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
+    endif()
+
+    message(STATUS "Found ffmpeg libs:       ${FFMPEG_LIBRARIES}")
+    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
+    message(STATUS "ffmpeg definitions:      ${FFMPEG_DEFINITIONS}")
+    message(STATUS "Found avformat ${AVFORMAT_VERSION}")
+
+    include_directories(${FFMPEG_INCLUDE_DIRS})
+    add_compile_definitions(WHISPER_FFMPEG)
+
+    list(APPEND COMMON_EXTRA_LIBS ${FFMPEG_LIBRARIES})
+
+    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
+endif()
+
+
 add_library(${TARGET} STATIC
     common.h
     common.cpp
@@ -29,11 +58,12 @@ add_library(${TARGET} STATIC
     common-ggml.cpp
     grammar-parser.h
     grammar-parser.cpp
+    ${COMMON_SOURCES_FFMPEG}
     )
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper)
+target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
 
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
@@ -50,8 +80,8 @@ if (WHISPER_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
+    target_include_directories(${TARGET} PUBLIC  ${SDL2_INCLUDE_DIRS})
+    target_link_libraries     (${TARGET} PRIVATE ${SDL2_LIBRARIES})
 
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
     set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
@@ -103,7 +133,7 @@ if (WHISPER_SDL2)
     set_target_properties(talk-llama PROPERTIES FOLDER "examples")
     add_subdirectory(lsp)
     set_target_properties(lsp PROPERTIES FOLDER "examples")
-    if (LLAMA_SYCL)
+    if (GGML_SYCL)
         add_subdirectory(sycl)
         set_target_properties(sycl PROPERTIES FOLDER "examples")
     endif()

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

@@ -12,6 +12,7 @@ const whisperParamsMock = {
   model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
   use_gpu: true,
+  flash_attn: false,
   no_prints: true,
   comma_in_time: false,
   translate: true,

examples/addon.node/addon.cpp

@@ -25,7 +25,6 @@ struct whisper_params {
     float entropy_thold = 2.4f;
     float logprob_thold = -1.0f;
 
-    bool speed_up       = false;
     bool translate      = false;
     bool diarize        = false;
     bool output_txt     = false;
@@ -39,6 +38,7 @@ struct whisper_params {
     bool no_timestamps  = false;
     bool no_prints      = false;
     bool use_gpu        = true;
+    bool flash_attn     = false;
     bool comma_in_time  = true;
 
     std::string language = "en";
@@ -146,6 +146,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     struct whisper_context_params cparams = whisper_context_default_params();
     cparams.use_gpu = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
     struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     if (ctx == nullptr) {
@@ -230,8 +231,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
             wparams.audio_ctx        = params.audio_ctx;
 
-            wparams.speed_up         = params.speed_up;
-
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
@@ -326,6 +325,7 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   std::string model = whisper_params.Get("model").As<Napi::String>();
   std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
   bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
+  bool flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
   bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
   bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
   int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
@@ -346,6 +346,7 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   params.model = model;
   params.fname_inp.emplace_back(input);
   params.use_gpu = use_gpu;
+  params.flash_attn = flash_attn;
   params.no_prints = no_prints;
   params.no_timestamps = no_timestamps;
   params.audio_ctx = audio_ctx;

examples/addon.node/index.js

@@ -12,6 +12,7 @@ const whisperParams = {
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
   use_gpu: true,
+  flash_attn: false,
   no_prints: true,
   comma_in_time: false,
   translate: true,

examples/bench/bench.cpp

@@ -18,7 +18,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -58,7 +58,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-int whisper_bench_full(const whisper_params & params) {
+static int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
     struct whisper_context_params cparams = whisper_context_default_params();

examples/command/command.cpp

@@ -38,7 +38,6 @@ struct whisper_params {
 
     grammar_parser::parse_state grammar_parsed;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
@@ -60,7 +59,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -76,7 +75,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
@@ -115,7 +113,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
@@ -133,7 +130,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(
+static std::string transcribe(
                  whisper_context * ctx,
             const whisper_params & params,
         const std::vector<float> & pcmf32,
@@ -165,7 +162,6 @@ std::string transcribe(
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx = params.audio_ctx;
-    wparams.speed_up  = params.speed_up;
 
     wparams.temperature     = 0.4f;
     wparams.temperature_inc = 1.0f;
@@ -220,7 +216,7 @@ std::string transcribe(
     return result;
 }
 
-std::vector<std::string> read_allowed_commands(const std::string & fname) {
+static std::vector<std::string> read_allowed_commands(const std::string & fname) {
     std::vector<std::string> allowed_commands;
 
     std::ifstream ifs(fname);
@@ -242,7 +238,7 @@ std::vector<std::string> read_allowed_commands(const std::string & fname) {
     return allowed_commands;
 }
 
-std::vector<std::string> get_words(const std::string &txt) {
+static std::vector<std::string> get_words(const std::string &txt) {
     std::vector<std::string> words;
 
     std::istringstream iss(txt);
@@ -256,7 +252,7 @@ std::vector<std::string> get_words(const std::string &txt) {
 
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: guided mode\n", __func__);
 
@@ -371,7 +367,6 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
             wparams.n_threads        = params.n_threads;
 
             wparams.audio_ctx        = params.audio_ctx;
-            wparams.speed_up         = params.speed_up;
 
             wparams.prompt_tokens    = k_tokens.data();
             wparams.prompt_n_tokens  = k_tokens.size();
@@ -468,7 +463,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
 
 // always-prompt mode
 // transcribe the voice into text after valid prompt
-int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running = true;
     bool ask_prompt = true;
 
@@ -548,7 +543,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
 
 // general-purpose mode
 // freely transcribe the voice into text
-int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running  = true;
     bool have_prompt = false;
     bool ask_prompt  = true;

examples/common-sdl.cpp

@@ -219,7 +219,7 @@ bool sdl_poll_events() {
             case SDL_QUIT:
                 {
                     return false;
-                } break;
+                }
             default:
                 break;
         }

examples/common.cpp

@@ -24,8 +24,13 @@
 #include <io.h>
 #endif
 
+#ifdef WHISPER_FFMPEG
+// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
+extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
+#endif
+
 // Function to check if the next argument exists
-std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
         return argv[++i];
     } else {
@@ -341,7 +346,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
-std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
     std::vector<gpt_vocab::id> output;
     std::stringstream ss(input);
     std::string token;
@@ -353,7 +358,7 @@ std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, ch
     return output;
 }
 
-std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
     if (fpath_test.empty()){
         fprintf(stderr, "%s : No test file found.\n", __func__);
         return std::map<std::string, std::vector<gpt_vocab::id>>();
@@ -637,7 +642,7 @@ bool is_wav_buffer(const std::string buf) {
 
 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
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin or ffmpeg decoding output
 
     if (fname == "-") {
         {
@@ -670,8 +675,19 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
         }
     }
     else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+#if defined(WHISPER_FFMPEG)
+        if (ffmpeg_decode_audio(fname, wav_data) != 0) {
+            fprintf(stderr, "error: failed to ffmpeg decode '%s' \n", fname.c_str());
+            return false;
+        }
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to read wav data as wav \n");
+            return false;
+        }
+#else
         fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
         return false;
+#endif
     }
 
     if (wav.channels != 1 && wav.channels != 2) {

examples/common.h

@@ -21,7 +21,7 @@ struct gpt_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
     int32_t n_predict    = 200;  // new tokens to predict
     int32_t n_parallel   = 1;    // number of parallel streams
-    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_batch      = 32;   // batch size for prompt processing
     int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
     int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
 
@@ -185,7 +185,7 @@ private:
     // It is assumed that PCM data is normalized to a range from -1 to 1
     bool write_audio(const float * data, size_t length) {
         for (size_t i = 0; i < length; ++i) {
-            const int16_t intSample = data[i] * 32767;
+            const int16_t intSample = int16_t(data[i] * 32767);
             file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
             dataSize += sizeof(int16_t);
         }

examples/ffmpeg-transcode.cpp

@@ -0,0 +1,350 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+/*
+ * transcode.c - convert audio file to WAVE
+ *
+ * Copyright (C) 2019		Andrew Clayton <andrew@digital-domain.net>
+ * Copyright (C) 2024       William Tambellini <william.tambellini@gmail.com>
+ */
+
+// Just for conveninent C++ API
+#include <vector>
+#include <string>
+
+// C
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <sys/mman.h>
+
+extern "C" {
+#include <libavutil/opt.h>
+#include <libavcodec/avcodec.h>
+#include <libavformat/avformat.h>
+#include <libswresample/swresample.h>
+}
+
+typedef uint64_t u64;
+typedef int64_t  s64;
+typedef uint32_t u32;
+typedef int32_t  s32;
+typedef uint16_t u16;
+typedef int16_t  s16;
+typedef uint8_t   u8;
+typedef int8_t    s8;
+
+#define WAVE_SAMPLE_RATE	16000
+#define AVIO_CTX_BUF_SZ		 4096
+
+static const char* ffmpegLog = getenv("FFMPEG_LOG");
+// Todo: add __FILE__ __LINE__
+#define LOG(...) \
+  do { if (ffmpegLog) fprintf(stderr, __VA_ARGS__); } while(0) // C99
+
+/*
+ * WAVE file header based on definition from
+ * https://gist.github.com/Jon-Schneider/8b7c53d27a7a13346a643dac9c19d34f
+ *
+ * We must ensure this structure doesn't have any holes or
+ * padding so we can just map it straight to the WAVE data.
+ */
+struct wave_hdr {
+	/* RIFF Header: "RIFF" */
+	char riff_header[4];
+	/* size of audio data + sizeof(struct wave_hdr) - 8 */
+	int wav_size;
+	/* "WAVE" */
+	char wav_header[4];
+
+	/* Format Header */
+	/* "fmt " (includes trailing space) */
+	char fmt_header[4];
+	/* Should be 16 for PCM */
+	int fmt_chunk_size;
+	/* Should be 1 for PCM. 3 for IEEE Float */
+	s16 audio_format;
+	s16 num_channels;
+	int sample_rate;
+	/*
+	 * Number of bytes per second
+	 * sample_rate * num_channels * bit_depth/8
+	 */
+	int byte_rate;
+	/* num_channels * bytes per sample */
+	s16 sample_alignment;
+	/* bits per sample */
+	s16 bit_depth;
+
+	/* Data Header */
+	/* "data" */
+	char data_header[4];
+	/*
+	 * size of audio
+	 * number of samples * num_channels * bit_depth/8
+	 */
+	int data_bytes;
+} __attribute__((__packed__));
+
+struct audio_buffer {
+	u8 *ptr;
+	int size; /* size left in the buffer */
+};
+
+static void set_wave_hdr(wave_hdr& wh, size_t size) {
+    memcpy(&wh.riff_header, "RIFF", 4);
+    wh.wav_size = size + sizeof(struct wave_hdr) - 8;
+    memcpy(&wh.wav_header, "WAVE", 4);
+    memcpy(&wh.fmt_header, "fmt ", 4);
+    wh.fmt_chunk_size = 16;
+    wh.audio_format = 1;
+    wh.num_channels = 1;
+    wh.sample_rate = WAVE_SAMPLE_RATE;
+    wh.sample_alignment = 2;
+    wh.bit_depth = 16;
+    wh.byte_rate = wh.sample_rate * wh.sample_alignment;
+    memcpy(&wh.data_header, "data", 4);
+    wh.data_bytes = size;
+}
+
+static void write_wave_hdr(int fd, size_t size) {
+	struct wave_hdr wh;
+    set_wave_hdr(wh, size);
+	write(fd, &wh, sizeof(struct wave_hdr));
+}
+
+static int map_file(int fd, u8 **ptr, size_t *size)
+{
+	struct stat sb;
+
+	fstat(fd, &sb);
+	*size = sb.st_size;
+
+    *ptr = (u8*)mmap(NULL, *size, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0);
+	if (*ptr == MAP_FAILED) {
+		perror("mmap");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int read_packet(void *opaque, u8 *buf, int buf_size)
+{
+    struct audio_buffer *audio_buf = (audio_buffer*)opaque;
+
+	buf_size = FFMIN(buf_size, audio_buf->size);
+
+	/* copy internal buffer data to buf */
+	memcpy(buf, audio_buf->ptr, buf_size);
+	audio_buf->ptr += buf_size;
+	audio_buf->size -= buf_size;
+
+	return buf_size;
+}
+
+static void convert_frame(struct SwrContext *swr, AVCodecContext *codec,
+			  AVFrame *frame, s16 **data, int *size, bool flush)
+{
+	int nr_samples;
+	s64 delay;
+	u8 *buffer;
+
+	delay = swr_get_delay(swr, codec->sample_rate);
+	nr_samples = av_rescale_rnd(delay + frame->nb_samples,
+				    WAVE_SAMPLE_RATE, codec->sample_rate,
+				    AV_ROUND_UP);
+	av_samples_alloc(&buffer, NULL, 1, nr_samples, AV_SAMPLE_FMT_S16, 0);
+
+	/*
+	 * !flush is used to check if we are flushing any remaining
+	 * conversion buffers...
+	 */
+	nr_samples = swr_convert(swr, &buffer, nr_samples,
+				 !flush ? (const u8 **)frame->data : NULL,
+				 !flush ? frame->nb_samples : 0);
+
+    *data = (s16*)realloc(*data, (*size + nr_samples) * sizeof(s16));
+	memcpy(*data + *size, buffer, nr_samples * sizeof(s16));
+	*size += nr_samples;
+	av_freep(&buffer);
+}
+
+static bool is_audio_stream(const AVStream *stream)
+{
+	if (stream->codecpar->codec_type == AVMEDIA_TYPE_AUDIO)
+		return true;
+
+	return false;
+}
+
+// Return non zero on error, 0 on success
+// audio_buffer: input memory
+// data: decoded output audio data (wav file)
+// size: size of output data
+static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
+{
+    LOG("decode_audio: input size: %d\n", audio_buf->size);
+	AVFormatContext *fmt_ctx;
+	AVIOContext *avio_ctx;
+	AVStream *stream;
+	AVCodecContext *codec;
+	AVPacket packet;
+	AVFrame *frame;
+	struct SwrContext *swr;
+	u8 *avio_ctx_buffer;
+	unsigned int i;
+	int stream_index = -1;
+	int err;
+    const size_t errbuffsize = 1024;
+    char errbuff[errbuffsize];
+
+    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
+
+    fmt_ctx = avformat_alloc_context();
+    avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
+    LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);
+    avio_ctx = avio_alloc_context(avio_ctx_buffer, AVIO_CTX_BUF_SZ, 0, audio_buf, &read_packet, NULL, NULL);
+	fmt_ctx->pb = avio_ctx;
+
+    // open the input stream and read header
+	err = avformat_open_input(&fmt_ctx, NULL, NULL, NULL);
+	if (err) {
+        LOG("Could not read audio buffer: %d: %s\n", err, av_make_error_string(errbuff, errbuffsize, err));
+        return err;
+	}
+
+	err = avformat_find_stream_info(fmt_ctx, NULL);
+	if (err < 0) {
+        LOG("Could not retrieve stream info from audio buffer: %d\n", err);
+        return err;
+	}
+
+	for (i = 0; i < fmt_ctx->nb_streams; i++) {
+		if (is_audio_stream(fmt_ctx->streams[i])) {
+			stream_index = i;
+			break;
+		}
+	}
+
+	if (stream_index == -1) {
+        LOG("Could not retrieve audio stream from buffer\n");
+		return -1;
+	}
+
+	stream = fmt_ctx->streams[stream_index];
+	codec = avcodec_alloc_context3(
+			avcodec_find_decoder(stream->codecpar->codec_id));
+	avcodec_parameters_to_context(codec, stream->codecpar);
+	err = avcodec_open2(codec, avcodec_find_decoder(codec->codec_id),
+							NULL);
+	if (err) {
+        LOG("Failed to open decoder for stream #%d in audio buffer\n", stream_index);
+        return err;
+	}
+
+	/* prepare resampler */
+	swr = swr_alloc();
+
+	av_opt_set_int(swr, "in_channel_count", codec->channels, 0);
+	av_opt_set_int(swr, "out_channel_count", 1, 0);
+	av_opt_set_int(swr, "in_channel_layout", codec->channel_layout, 0);
+	av_opt_set_int(swr, "out_channel_layout", AV_CH_LAYOUT_MONO, 0);
+	av_opt_set_int(swr, "in_sample_rate", codec->sample_rate, 0);
+	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
+	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
+	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
+
+	swr_init(swr);
+	if (!swr_is_initialized(swr)) {
+        LOG("Resampler has not been properly initialized\n");
+		return -1;
+	}
+
+	av_init_packet(&packet);
+	frame = av_frame_alloc();
+	if (!frame) {
+        LOG("Error allocating the frame\n");
+		return -1;
+	}
+
+	/* iterate through frames */
+	*data = NULL;
+	*size = 0;
+	while (av_read_frame(fmt_ctx, &packet) >= 0) {
+		avcodec_send_packet(codec, &packet);
+
+		err = avcodec_receive_frame(codec, frame);
+		if (err == AVERROR(EAGAIN))
+			continue;
+
+		convert_frame(swr, codec, frame, data, size, false);
+	}
+	/* Flush any remaining conversion buffers... */
+	convert_frame(swr, codec, frame, data, size, true);
+
+	av_frame_free(&frame);
+	swr_free(&swr);
+    //avio_context_free(); // todo?
+	avcodec_close(codec);
+	avformat_close_input(&fmt_ctx);
+	avformat_free_context(fmt_ctx);
+
+	if (avio_ctx) {
+		av_freep(&avio_ctx->buffer);
+		av_freep(&avio_ctx);
+	}
+
+	return 0;
+}
+
+// in mem decoding/conversion/resampling:
+// ifname: input file path
+// owav_data: in mem wav file. Can be forwarded as it to whisper/drwav
+// return 0 on success
+int ffmpeg_decode_audio(const std::string &ifname, std::vector<uint8_t>& owav_data) {
+    LOG("ffmpeg_decode_audio: %s\n", ifname.c_str());
+    int ifd = open(ifname.c_str(), O_RDONLY);
+    if (ifd == -1) {
+        fprintf(stderr, "Couldn't open input file %s\n", ifname.c_str());
+        return -1;
+    }
+    u8 *ibuf = NULL;
+    size_t ibuf_size;
+    int err = map_file(ifd, &ibuf, &ibuf_size);
+    if (err) {
+        LOG("Couldn't map input file %s\n", ifname.c_str());
+        return err;
+    }
+    LOG("Mapped input file: %x size: %d\n", ibuf, ibuf_size);
+    struct audio_buffer inaudio_buf;
+    inaudio_buf.ptr = ibuf;
+    inaudio_buf.size = ibuf_size;
+
+    s16 *odata=NULL;
+    int osize=0;
+
+    err = decode_audio(&inaudio_buf, &odata, &osize);
+    LOG("decode_audio returned %d \n", err);
+    if (err != 0) {
+        LOG("decode_audio failed\n");
+        return err;
+    }
+    LOG("decode_audio output size: %d\n", osize);
+
+    wave_hdr wh;
+    const size_t outdatasize = osize * sizeof(s16);
+    set_wave_hdr(wh, outdatasize);
+    owav_data.resize(sizeof(wave_hdr) + outdatasize);
+    // header:
+    memcpy(owav_data.data(), &wh, sizeof(wave_hdr));
+    // the data:
+    memcpy(owav_data.data() + sizeof(wave_hdr), odata, osize* sizeof(s16));
+
+    return 0;
+}

examples/grammar-parser.cpp

@@ -9,7 +9,7 @@
 namespace grammar_parser {
     // NOTE: assumes valid utf8 (but checks for overrun)
     // copied from whisper.cpp
-    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
         static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
         uint8_t  first_byte = static_cast<uint8_t>(*src);
         uint8_t  highbits   = first_byte >> 4;
@@ -24,19 +24,19 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
         return result.first->second;
     }
 
-    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
         state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
         return next_id;
     }
 
-    void add_rule(
+    static void add_rule(
             parse_state & state,
             uint32_t      rule_id,
             const std::vector<whisper_grammar_element> & rule) {
@@ -46,11 +46,11 @@ namespace grammar_parser {
         state.rules[rule_id] = rule;
     }
 
-    bool is_word_char(char c) {
+    static bool is_word_char(char c) {
         return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
     }
 
-    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
         const char * pos   = src;
         const char * end   = src + size;
         uint32_t     value = 0;
@@ -73,7 +73,7 @@ namespace grammar_parser {
         return std::make_pair(value, pos);
     }
 
-    const char * parse_space(const char * src, bool newline_ok) {
+    static const char * parse_space(const char * src, bool newline_ok) {
         const char * pos = src;
         while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                 (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@@ -88,7 +88,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_name(const char * src) {
+    static const char * parse_name(const char * src) {
         const char * pos = src;
         while (is_word_char(*pos)) {
             pos++;
@@ -99,7 +99,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    std::pair<uint32_t, const char *> parse_char(const char * src) {
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
         if (*src == '\\') {
             switch (src[1]) {
                 case 'x': return parse_hex(src + 2, 2);
@@ -122,14 +122,14 @@ namespace grammar_parser {
         throw std::runtime_error("unexpected end of input");
     }
 
-    const char * parse_alternates(
+    static const char * parse_alternates(
             parse_state       & state,
             const char        * src,
             const std::string & rule_name,
             uint32_t            rule_id,
             bool                is_nested);
 
-    const char * parse_sequence(
+    static const char * parse_sequence(
             parse_state                        & state,
             const char                         * src,
             const std::string                  & rule_name,
@@ -229,7 +229,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_alternates(
+    static const char * parse_alternates(
             parse_state       & state,
             const char        * src,
             const std::string & rule_name,
@@ -247,7 +247,7 @@ namespace grammar_parser {
         return pos;
     }
 
-    const char * parse_rule(parse_state & state, const char * src) {
+    static const char * parse_rule(parse_state & state, const char * src) {
         const char * name_end = parse_name(src);
         const char * pos      = parse_space(name_end, false);
         size_t       name_len = name_end - src;
@@ -285,7 +285,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_grammar_char(FILE * file, uint32_t c) {
+    static void print_grammar_char(FILE * file, uint32_t c) {
         if (0x20 <= c && c <= 0x7f) {
             fprintf(file, "%c", static_cast<char>(c));
         } else {
@@ -294,7 +294,7 @@ namespace grammar_parser {
         }
     }
 
-    bool is_char_element(whisper_grammar_element elem) {
+    static bool is_char_element(whisper_grammar_element elem) {
         switch (elem.type) {
             case WHISPER_GRETYPE_CHAR:           return true;
             case WHISPER_GRETYPE_CHAR_NOT:       return true;
@@ -304,7 +304,7 @@ namespace grammar_parser {
         }
     }
 
-    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+    static void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
         for (auto elem : rule) {
             switch (elem.type) {
                 case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
@@ -334,7 +334,7 @@ namespace grammar_parser {
         fprintf(file, "\n");
     }
 
-    void print_rule(
+    static void print_rule(
             FILE     * file,
             uint32_t   rule_id,
             const std::vector<whisper_grammar_element> & rule,
@@ -413,7 +413,7 @@ namespace grammar_parser {
         }
     }
 
-    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const {
         std::vector<const whisper_grammar_element *> ret;
         for (const auto & rule : rules) {
             ret.push_back(rule.data());

examples/lsp/lsp.cpp

@@ -26,7 +26,6 @@ struct whisper_params {
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
@@ -54,7 +53,7 @@ struct commandset {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -70,7 +69,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
@@ -102,7 +100,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
@@ -112,7 +109,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "\n");
 }
-uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
+static uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
     using namespace std::chrono;
     uint64_t time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
     uint64_t start_time = time_now;
@@ -156,7 +153,7 @@ uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params &
     return time_now;
 }
 
-json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
+static json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
     std::vector<whisper_token> prompt_tokens;
     std::vector<float> pcmf32;
     uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 10000U, pcmf32);
@@ -184,7 +181,6 @@ json unguided_transcription(struct whisper_context * ctx, audio_async &audio, js
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
     wparams.suppress_non_speech_tokens = true;
     // run the transformer and a single decoding pass
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
@@ -203,7 +199,7 @@ json unguided_transcription(struct whisper_context * ctx, audio_async &audio, js
 
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
+static json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
     struct commandset cs = commandset_list[jparams.value("commandset_index", commandset_list.size()-1)];
     std::vector<float> pcmf32;
     uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 2000U, pcmf32);
@@ -223,7 +219,6 @@ json guided_transcription(struct whisper_context * ctx, audio_async &audio, cons
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
 
     // TODO: Do some time testing. Does an overly long prompt slow down processing?
     // Set up command sets/precompute prompts
@@ -290,7 +285,7 @@ json guided_transcription(struct whisper_context * ctx, audio_async &audio, cons
     }
 }
 
-json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
+static json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
     // TODO: check for token collision
     struct commandset cs;
 
@@ -330,7 +325,8 @@ json register_commandset(struct whisper_context * ctx, json jparams, std::vector
     commandset_list.push_back(cs);
     return json{{"index",index}};
 }
-json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
+
+static json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
     // whisper_state has the pertinent offsets, but there also seem to be a large
     // number of scratch buffers that would prevent rewinding context in a manner similar to llama
     // I'll give this a another pass once everything else is implemented,
@@ -340,7 +336,8 @@ json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*para
             {"message", "Seeking is not yet supported."}
     };
 }
-json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
+
+static json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
     // See: https://www.jsonrpc.org/specification
     json id = body.at("id");
     try {
@@ -380,7 +377,7 @@ json parse_job(const json &body, struct whisper_context * ctx, audio_async &audi
     }
 }
 
-void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+static void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
     std::deque<json> jobqueue;
     std::vector<struct commandset> commandset_list;
     while (true) {

examples/main/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

examples/main/main.cpp

@@ -17,7 +17,7 @@
 #endif
 
 // helper function to replace substrings
-void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     for (size_t pos = 0; ; pos += replace.length()) {
         pos = s.find(search, pos);
         if (pos == std::string::npos) break;
@@ -47,7 +47,6 @@ struct whisper_params {
     float temperature     = 0.0f;
     float temperature_inc = 0.2f;
 
-    bool speed_up        = false;
     bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
@@ -95,17 +94,17 @@ struct whisper_params {
     grammar_parser::parse_state grammar_parsed;
 };
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+static void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-char* whisper_param_turn_lowercase(char* in){
+static char * whisper_param_turn_lowercase(char * in){
     int string_len = strlen(in);
-    for(int i = 0; i < string_len; i++){
+    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) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -138,7 +137,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
         else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
         else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
-        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@@ -184,7 +182,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
     fprintf(stderr, "\n");
@@ -206,7 +204,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
     fprintf(stderr, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
     fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
-    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@@ -251,7 +248,7 @@ struct whisper_print_user_data {
     int progress_prev;
 };
 
-std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+static std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
 
@@ -283,7 +280,8 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
 
     return speaker;
 }
-void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+
+static void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
     int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
     int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
     if (progress >= *progress_prev + progress_step) {
@@ -292,7 +290,7 @@ void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct wh
     }
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+static 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;
 
@@ -361,7 +359,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
-bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -388,7 +386,7 @@ bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -420,7 +418,7 @@ bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -449,7 +447,7 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-char *escape_double_quotes_and_backslashes(const char *str) {
+static char * escape_double_quotes_and_backslashes(const char * str) {
     if (str == NULL) {
         return NULL;
     }
@@ -462,7 +460,7 @@ char *escape_double_quotes_and_backslashes(const char *str) {
         }
     }
 
-    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    char * escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
     if (escaped == NULL) {
         return NULL;
     }
@@ -481,7 +479,7 @@ char *escape_double_quotes_and_backslashes(const char *str) {
 }
 
 // double quote should be escaped by another double quote. (rfc4180)
-char *escape_double_quotes_in_csv(const char *str) {
+static char * escape_double_quotes_in_csv(const char * str) {
     if (str == NULL) {
         return NULL;
     }
@@ -512,7 +510,7 @@ char *escape_double_quotes_in_csv(const char *str) {
     return escaped;
 }
 
-bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -547,7 +545,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
-bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+static bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
     std::ofstream fout(fname);
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
@@ -566,7 +564,7 @@ bool output_score(struct whisper_context * ctx, const char * fname, const whispe
     return true;
 }
 
-bool output_json(
+static bool output_json(
              struct whisper_context * ctx,
                          const char * fname,
                const whisper_params & params,
@@ -737,7 +735,7 @@ bool output_json(
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
+static bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@@ -862,7 +860,7 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
     return true;
 }
 
-bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+static bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -903,7 +901,7 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
 }
 
 
-void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
 
 int main(int argc, char ** argv) {
     whisper_params params;
@@ -1106,7 +1104,6 @@ int main(int argc, char ** argv) {
             wparams.split_on_word    = params.split_on_word;
             wparams.audio_ctx        = params.audio_ctx;
 
-            wparams.speed_up         = params.speed_up;
             wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]

examples/quantize/quantize.cpp

@@ -36,7 +36,7 @@ struct whisper_filters {
 };
 
 // quantize a model
-bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+static bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
     gpt_vocab vocab;
 
     printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());

examples/server/server.cpp

@@ -61,7 +61,6 @@ struct whisper_params {
     float temperature     =  0.00f;
     float temperature_inc =  0.20f;
 
-    bool speed_up        = false;
     bool debug_mode      = false;
     bool translate       = false;
     bool detect_language = false;
@@ -112,7 +111,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@@ -159,7 +157,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
         else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@@ -768,7 +765,6 @@ int main(int argc, char ** argv) {
             wparams.split_on_word    = params.split_on_word;
             wparams.audio_ctx        = params.audio_ctx;
 
-            wparams.speed_up         = params.speed_up;
             wparams.debug_mode       = params.debug_mode;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
@@ -947,7 +943,7 @@ int main(int argc, char ** argv) {
                             "application/json");
         }
 
-        // reset params to thier defaults
+        // reset params to their defaults
         params = default_params;
     });
     svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){

examples/stream/stream.cpp

@@ -27,7 +27,6 @@ struct whisper_params {
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool no_fallback   = false;
     bool print_special = false;
@@ -45,7 +44,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -62,7 +61,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
         else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
@@ -100,7 +98,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
     fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
@@ -314,7 +311,6 @@ int main(int argc, char ** argv) {
             wparams.n_threads        = params.n_threads;
 
             wparams.audio_ctx        = params.audio_ctx;
-            wparams.speed_up         = params.speed_up;
 
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 

examples/talk-llama/llama.h

@@ -81,9 +81,12 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
         LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
         LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
+        LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
+        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
+        LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
+        LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
+        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
+        LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -95,7 +98,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_GLM  =  4,
     };
 
-    enum llama_token_type {
+    enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
         LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
         LLAMA_TOKEN_TYPE_NORMAL       = 1,
         LLAMA_TOKEN_TYPE_UNKNOWN      = 2,
@@ -105,6 +108,20 @@ extern "C" {
         LLAMA_TOKEN_TYPE_BYTE         = 6,
     };
 
+    enum llama_token_attr {
+        LLAMA_TOKEN_ATTR_UNDEFINED    = 0,
+        LLAMA_TOKEN_ATTR_UNKNOWN      = 1 << 0,
+        LLAMA_TOKEN_ATTR_UNUSED       = 1 << 1,
+        LLAMA_TOKEN_ATTR_NORMAL       = 1 << 2,
+        LLAMA_TOKEN_ATTR_CONTROL      = 1 << 3,  // SPECIAL?
+        LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4,
+        LLAMA_TOKEN_ATTR_BYTE         = 1 << 5,
+        LLAMA_TOKEN_ATTR_NORMALIZED   = 1 << 6,
+        LLAMA_TOKEN_ATTR_LSTRIP       = 1 << 7,
+        LLAMA_TOKEN_ATTR_RSTRIP       = 1 << 8,
+        LLAMA_TOKEN_ATTR_SINGLE_WORD  = 1 << 9,
+    };
+
     // model file types
     enum llama_ftype {
         LLAMA_FTYPE_ALL_F32              = 0,
@@ -157,6 +174,7 @@ extern "C" {
         LLAMA_POOLING_TYPE_NONE = 0,
         LLAMA_POOLING_TYPE_MEAN = 1,
         LLAMA_POOLING_TYPE_CLS  = 2,
+        LLAMA_POOLING_TYPE_LAST = 3,
     };
 
     enum llama_split_mode {
@@ -242,6 +260,9 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
+        // comma separated list of RPC servers to use for offloading
+        const char * rpc_servers;
+
         // 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.
@@ -260,6 +281,8 @@ extern "C" {
         bool check_tensors; // validate model tensor data
     };
 
+    // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
+    //       https://github.com/ggerganov/llama.cpp/pull/7544
     struct llama_context_params {
         uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
@@ -271,7 +294,6 @@ extern "C" {
 
         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
@@ -286,14 +308,14 @@ extern "C" {
         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
+        enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
+        enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention
+        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -344,6 +366,9 @@ extern "C" {
         // modifies a preceding LLAMA_GRETYPE_CHAR or
         // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
         LLAMA_GRETYPE_CHAR_ALT       = 6,
+
+        // any character (.)
+        LLAMA_GRETYPE_CHAR_ANY       = 7,
     };
 
     typedef struct llama_grammar_element {
@@ -417,8 +442,8 @@ extern "C" {
 
     LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
 
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model   * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model   * model);
+    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model * model);
+    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model * model);
 
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@@ -755,6 +780,16 @@ 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);
 
+    // Get the number of threads used for generation of a single token.
+    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
+
+    // Get the number of threads used for prompt and batch processing (multiple token).
+    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
+
+    // Set whether the model is in embeddings mode or not
+    // If true, embeddings will be returned but logits will not
+    LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
+
     // 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);
@@ -808,11 +843,14 @@ extern "C" {
 
     LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
 
-    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);
+    LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token);
 
     // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
     LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
 
+    // Identify if Token Id is a control token or a render-able token
+    LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
+
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
@@ -1026,49 +1064,9 @@ extern "C" {
                      llama_token   token);
 
     //
-    // Beam search
+    // Model split
     //
 
-    struct llama_beam_view {
-        const llama_token * tokens;
-
-        size_t n_tokens;
-        float  p;        // Cumulative beam probability (renormalized relative to all beams)
-        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
-    };
-
-    // Passed to beam_search_callback function.
-    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
-    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
-    // These pointers are valid only during the synchronous callback, so should not be saved.
-    struct llama_beams_state {
-        struct llama_beam_view * beam_views;
-
-        size_t n_beams;               // Number of elements in beam_views[].
-        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
-        bool   last_call;             // True iff this is the last callback invocation.
-    };
-
-    // Type of pointer to the beam_search_callback function.
-    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
-    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
-    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state);
-
-    /// @details Deterministically returns entire sentence constructed by a beam search.
-    /// @param ctx Pointer to the llama_context.
-    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
-    /// @param callback_data A pointer that is simply passed back to callback.
-    /// @param n_beams Number of beams to use.
-    /// @param n_past Number of tokens already evaluated.
-    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
-    LLAMA_API void llama_beam_search(
-                   struct llama_context * ctx,
-        llama_beam_search_callback_fn_t   callback,
-                                   void * callback_data,
-                                 size_t   n_beams,
-                                int32_t   n_past,
-                                int32_t   n_predict);
-
     /// @details Build a split GGUF final path for this chunk.
     ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
     //  Returns the split_path length.

examples/talk-llama/talk-llama.cpp

@@ -16,7 +16,7 @@
 #include <regex>
 #include <sstream>
 
-std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+static std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
     auto * model = llama_get_model(ctx);
 
     // upper limit for the number of tokens
@@ -33,7 +33,7 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
     return result;
 }
 
-std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
     const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
     if (n_tokens < 0) {
@@ -59,7 +59,6 @@ struct whisper_params {
     float vad_thold  = 0.6f;
     float freq_thold = 100.0f;
 
-    bool speed_up       = false;
     bool translate      = false;
     bool print_special  = false;
     bool print_energy   = false;
@@ -84,7 +83,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -100,7 +99,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ngl" || arg == "--n-gpu-layers")   { params.n_gpu_layers   = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")      { params.vad_thold      = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")     { params.freq_thold     = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")       { params.speed_up       = true; }
         else if (arg == "-tr"  || arg == "--translate")      { params.translate      = true; }
         else if (arg == "-ps"  || arg == "--print-special")  { params.print_special  = true; }
         else if (arg == "-pe"  || arg == "--print-energy")   { params.print_energy   = true; }
@@ -149,7 +147,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ngl N,   --n-gpu-layers N [%-7d] number of layers to store in VRAM\n",           params.n_gpu_layers);
     fprintf(stderr, "  -vth N,   --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,      --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,      --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
@@ -171,7 +168,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(
+static std::string transcribe(
         whisper_context * ctx,
         const whisper_params & params,
         const std::vector<float> & pcmf32,
@@ -205,7 +202,6 @@ std::string transcribe(
     wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
@@ -239,7 +235,7 @@ std::string transcribe(
     return result;
 }
 
-std::vector<std::string> get_words(const std::string &txt) {
+static std::vector<std::string> get_words(const std::string &txt) {
     std::vector<std::string> words;
 
     std::istringstream iss(txt);

examples/talk-llama/unicode-data.h

@@ -1,17 +1,20 @@
 #pragma once
 
 #include <cstdint>
-#include <map>
-#include <utility>
 #include <vector>
+#include <unordered_map>
+#include <unordered_set>
 
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol;
-extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control;
-extern const std::multimap<uint32_t, uint32_t>          unicode_map_nfd;
-extern const std::map<char32_t, char32_t>               unicode_map_lowercase;
+struct range_nfd {
+    uint32_t first;
+    uint32_t last;
+    uint32_t nfd;
+};
+
+static const uint32_t MAX_CODEPOINTS = 0x110000;
+
+extern const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags;
+extern const std::unordered_set<uint32_t> unicode_set_whitespace;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase;
+extern const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase;
+extern const std::vector<range_nfd> unicode_ranges_nfd;

examples/talk-llama/unicode.cpp

@@ -1,4 +1,4 @@
-#include "unicode.h"
+#include "unicode.h"
 #include "unicode-data.h"
 
 #include <cassert>
@@ -109,57 +109,49 @@ static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset)
 //    return result;
 //}
 
-static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
-    std::unordered_map<uint32_t, int> cpt_types;
-    for (auto p : unicode_ranges_number) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_NUMBER;
+static std::vector<codepoint_flags> unicode_cpt_flags_array() {
+    std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+
+    assert (unicode_ranges_flags.front().first == 0);
+    assert (unicode_ranges_flags.back().first == MAX_CODEPOINTS);
+    for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
+        const auto range_ini = unicode_ranges_flags[i-1];  // codepoint_ini, flags
+        const auto range_end = unicode_ranges_flags[i];    // codepoint_end, flags
+        for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
+            cpt_flags[cpt] = range_ini.second;
         }
     }
-    for (auto p : unicode_ranges_letter) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_LETTER;
-        }
+
+    for (auto cpt : unicode_set_whitespace) {
+        cpt_flags[cpt].is_whitespace = true;
     }
-    for (auto p : unicode_ranges_separator) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_SEPARATOR;
-        }
+
+    for (auto p : unicode_map_lowercase) {
+        cpt_flags[p.second].is_lowercase = true;
     }
-    for (auto p : unicode_ranges_accent_mark) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
-        }
+
+    for (auto p : unicode_map_uppercase) {
+        cpt_flags[p.second].is_uppercase = true;
     }
-    for (auto p : unicode_ranges_punctuation) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_PUNCTUATION;
-        }
+
+    for (auto &range : unicode_ranges_nfd) {  // start, last, nfd
+        cpt_flags[range.nfd].is_nfd = true;
     }
-    for  (auto p : unicode_ranges_symbol) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_SYMBOL;
-        }
-    }
-    for (auto p : unicode_ranges_control) {
-        for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_CONTROL;
-        }
-    }
-    return cpt_types;
+
+    return cpt_flags;
 }
 
 static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
     std::unordered_map<uint8_t, std::string> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
         assert(0 <= ch && ch < 256);
         map[ch] = unicode_cpt_to_utf8(ch);
     }
@@ -175,15 +167,15 @@ static std::unordered_map<uint8_t, std::string> unicode_byte_to_utf8_map() {
 
 static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
     std::unordered_map<std::string, uint8_t> map;
-    for (int ch = u'!'; ch <= u'~'; ++ch) {
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
-    for (int ch = u'¡'; ch <= u'¬'; ++ch) {
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
-    for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
         assert(0 <= ch && ch < 256);
         map[unicode_cpt_to_utf8(ch)] = ch;
     }
@@ -234,12 +226,14 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
         assert(offset_end <= cpts.size());
         start = offset_end;
 
-        auto _get_cpt = [&] (const size_t pos) -> char32_t {
-            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_cpt_type = [&] (const size_t pos) -> int {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
         };
 
         size_t _prev_end = offset_ini;
@@ -260,18 +254,18 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
         };
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
-            const char32_t cpt = _get_cpt(pos);
-            const int cpt_type = _get_cpt_type(pos);
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
 
             // regex: 's|'t|'re|'ve|'m|'ll|'d
             if (cpt == '\'' && pos+1 < offset_end) {
-                char32_t cpt_next = _get_cpt(pos+1);
+                uint32_t cpt_next = _get_cpt(pos+1);
                 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
                     pos += _add_token(pos+2);
                     continue;
                 }
                 if (pos+2 < offset_end) {
-                    char32_t cpt_next_next = _get_cpt(pos+2);
+                    uint32_t cpt_next_next = _get_cpt(pos+2);
                     if ((cpt_next == 'r' && cpt_next_next == 'e') ||
                         (cpt_next == 'v' && cpt_next_next == 'e') ||
                         (cpt_next == 'l' && cpt_next_next == 'l')) {
@@ -281,44 +275,42 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
                 }
             }
 
-            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
-            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
             // regex: <space>?\p{L}+
-            if (cpt2_type == CODEPOINT_TYPE_LETTER) {
+            if (flags2.is_letter) {
                 pos += (cpt == ' ');
-                while (cpt2_type == CODEPOINT_TYPE_LETTER) {
-                    cpt2_type = _get_cpt_type(++pos);
+                while (flags2.is_letter) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
             // regex: <space>?\p{N}+
-            if (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+            if (flags2.is_number) {
                 pos += (cpt == ' ');
-                while (cpt2_type == CODEPOINT_TYPE_NUMBER) {
-                    cpt2_type = _get_cpt_type(++pos);
+                while (flags2.is_number) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
             // regex: <space>?[^\s\p{L}\p{N}]+
-            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
                 pos += (cpt == ' ');
-                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
-                    cpt2_type = _get_cpt_type(++pos);
-                    cpt2 = _get_cpt(pos);
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
+                    flags2 = _get_flags(++pos);
                 }
                 _add_token(pos);
                 continue;
             }
 
             size_t num_whitespaces = 0;
-            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
                 num_whitespaces++;
             }
 
             // regex: \s+(?!\S)
-            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
                 pos += num_whitespaces - 1;
                 _add_token(pos);
                 continue;
@@ -353,12 +345,14 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
         assert(offset_end <= cpts.size());
         start = offset_end;
 
-        auto _get_cpt = [&] (const size_t pos) -> char32_t {
-            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_cpt_type = [&] (const size_t pos) -> int {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : undef;
         };
 
         size_t _prev_end = offset_ini;
@@ -379,18 +373,18 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
         };
 
         for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
-            const char32_t cpt = _get_cpt(pos);
-            const int cpt_type = _get_cpt_type(pos);
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
 
             // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
             if (cpt == '\'' && pos+1 < offset_end) {
-                char32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
+                uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
                 if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
                     pos += _add_token(pos+2);
                     continue;
                 }
                 if (pos+2 < offset_end) {
-                    char32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
+                    uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
                     if ((cpt_next == 'r' && cpt_next_next == 'e') ||
                         (cpt_next == 'v' && cpt_next_next == 'e') ||
                         (cpt_next == 'l' && cpt_next_next == 'l')) {
@@ -401,10 +395,10 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: [^\r\n\p{L}\p{N}]?\p{L}+  //####FIXME: the first \p{L} is correct?
-            if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) {
-                if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) {  // one or more letters
+            if (!(cpt == '\r' || cpt == '\n' || /*flags.is_letter |*/ flags.is_number)) {
+                if (flags.is_letter || _get_flags(pos+1).is_letter) {  // one or more letters
                     pos++;
-                    while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) {
+                    while (_get_flags(pos).is_letter) {
                         pos++;
                     }
                     _add_token(pos);
@@ -413,9 +407,9 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: \p{N}{1,3}
-            if (cpt_type == CODEPOINT_TYPE_NUMBER) {
+            if (flags.is_number) {
                 size_t ini = pos;
-                while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) {
+                while (_get_flags(pos).is_number) {
                     if (++pos - ini >= 3 ) {
                         _add_token(pos);
                         ini = pos;
@@ -426,14 +420,13 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
-            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
-            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
-            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
                 pos += (cpt == ' ');
-                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
-                    cpt2_type = _get_cpt_type(++pos);
-                    cpt2 = _get_cpt(pos);
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number || flags2.is_undefined)) {
+                    flags2 = _get_flags(++pos);
                 }
+                uint32_t cpt2 = _get_cpt(pos);
                 while (cpt2 == '\r' || cpt2 == '\n') {
                     cpt2 = _get_cpt(++pos);
                 }
@@ -443,8 +436,8 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
             size_t num_whitespaces = 0;
             size_t last_end_r_or_n = 0;
-            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
-                char32_t cpt2 = _get_cpt(pos+num_whitespaces);
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
+                uint32_t cpt2 = _get_cpt(pos+num_whitespaces);
                 if (cpt2 == '\r' || cpt2 == '\n') {
                     last_end_r_or_n = pos + num_whitespaces + 1;
                 }
@@ -459,7 +452,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             }
 
             // regex: \s+(?!\S)
-            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
                 pos += num_whitespaces - 1;
                 _add_token(pos);
                 continue;
@@ -589,21 +582,21 @@ std::string unicode_cpt_to_utf8(uint32_t cp) {
 }
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts) {
-    std::vector<uint32_t> result;
-    result.reserve(cpts.size());
+    auto comp = [] (const uint32_t cpt, const range_nfd & range) {
+        return cpt < range.first;
+    };
+    std::vector<uint32_t> result(cpts.size());
     for (size_t i = 0; i < cpts.size(); ++i) {
-        auto it = unicode_map_nfd.find(cpts[i]);
-        if (it == unicode_map_nfd.end()) {
-            result.push_back(cpts[i]);
-        } else {
-            result.push_back(it->second);
-        }
+        const uint32_t cpt = cpts[i];
+        auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1;
+        result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
     }
     return result;
 }
 
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     std::vector<uint32_t> result;
+    result.reserve(utf8.size());
     size_t offset = 0;
     while (offset < utf8.size()) {
         result.push_back(unicode_cpt_from_utf8(utf8, offset));
@@ -611,31 +604,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     return result;
 }
 
-int unicode_cpt_type(uint32_t cp) {
-    static std::unordered_map<uint32_t, int> cpt_types = unicode_cpt_type_map();
-    const auto it = cpt_types.find(cp);
-    return it == cpt_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
+codepoint_flags unicode_cpt_flags(const uint32_t cp) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    static const auto cpt_flags = unicode_cpt_flags_array();
+    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
 }
 
-int unicode_cpt_type(const std::string & utf8) {
-    if (utf8.length() == 0) {
-        return CODEPOINT_TYPE_UNIDENTIFIED;
+codepoint_flags unicode_cpt_flags(const std::string & utf8) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    if (utf8.empty()) {
+        return undef;  // undefined
     }
     size_t offset = 0;
-    return unicode_cpt_type(unicode_cpt_from_utf8(utf8, offset));
-}
-
-bool unicode_cpt_is_whitespace(uint32_t cp) {
-    static const std::unordered_set<uint32_t> is_whitespace = [] {
-        std::unordered_set<uint32_t> is_whitespace;
-        for (auto p : unicode_ranges_whitespace) {
-            for (auto i = p.first; i <= p.second; ++i) {
-                is_whitespace.insert(i);
-            }
-        }
-        return is_whitespace;
-    }();
-    return (bool)is_whitespace.count(cp);
+    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
 }
 
 std::string unicode_byte_to_utf8(uint8_t byte) {
@@ -648,7 +629,7 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) {
     return map.at(utf8);
 }
 
-char32_t unicode_tolower(char32_t cp) {
+uint32_t unicode_tolower(uint32_t cp) {
     auto it = unicode_map_lowercase.find(cp);
     return it == unicode_map_lowercase.end() ? cp : it->second;
 }
@@ -656,21 +637,21 @@ char32_t unicode_tolower(char32_t cp) {
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {
-        { "\\p{N}", CODEPOINT_TYPE_NUMBER },
-        { "\\p{L}", CODEPOINT_TYPE_LETTER },
-        { "\\p{P}", CODEPOINT_TYPE_PUNCTUATION },
+        { "\\p{N}", codepoint_flags::NUMBER },
+        { "\\p{L}", codepoint_flags::LETTER },
+        { "\\p{P}", codepoint_flags::PUNCTUATION },
     };
 
     static const std::map<int, int> k_ucat_cpt = {
-        { CODEPOINT_TYPE_NUMBER,        0xD1 },
-        { CODEPOINT_TYPE_LETTER,        0xD2 },
-        { CODEPOINT_TYPE_PUNCTUATION,   0xD3 },
+        { codepoint_flags::NUMBER,        0xD1 },
+        { codepoint_flags::LETTER,        0xD2 },
+        { codepoint_flags::PUNCTUATION,   0xD3 },
     };
 
     static const std::map<int, std::string> k_ucat_map = {
-        { CODEPOINT_TYPE_NUMBER,        "\x30-\x39" }, // 0-9
-        { CODEPOINT_TYPE_LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
-        { CODEPOINT_TYPE_PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
+        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
     };
 
     // compute collapsed codepoints only if needed by at least one regex
@@ -701,10 +682,14 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const int cpt_type = unicode_cpt_type(cpts[i]);
+            const auto flags = unicode_cpt_flags(cpts[i]);
 
-            if (k_ucat_cpt.find(cpt_type) != k_ucat_cpt.end()) {
-                text_collapsed[i] = k_ucat_cpt.at(cpt_type);
+            if (flags.is_whitespace) {
+                //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
+                //text_collapsed[i] = (char) 0x85;  // <Next Line> as whitespace fallback
+                text_collapsed[i] = (char) 0x0B;    // <vertical tab> as whitespace fallback
+            } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
+                text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());
             } else {
                 text_collapsed[i] = (char) 0xD0; // fallback
             }
@@ -788,9 +773,16 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
             } else {
                 // no unicode category used, we can use std::wregex directly
-                const std::wstring wtext       = unicode_wstring_from_utf8(text);
                 const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
 
+                // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
+                std::wstring wtext(cpts.begin(), cpts.end());
+                for (size_t i = 0; i < wtext.size(); ++i) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                        wtext[i] = 0x0B;
+                    }
+                }
+
                 //printf("text: %s\n", text.c_str());
                 //printf("regex_expr: %s\n", regex_expr.c_str());
                 bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);

examples/talk-llama/unicode.h

@@ -4,28 +4,60 @@
 #include <string>
 #include <vector>
 
-#define CODEPOINT_TYPE_UNIDENTIFIED 0
-#define CODEPOINT_TYPE_NUMBER       1
-#define CODEPOINT_TYPE_LETTER       2
-#define CODEPOINT_TYPE_SEPARATOR    3
-#define CODEPOINT_TYPE_ACCENT_MARK  4
-#define CODEPOINT_TYPE_PUNCTUATION  5
-#define CODEPOINT_TYPE_SYMBOL       6
-#define CODEPOINT_TYPE_CONTROL      7
+struct codepoint_flags {
+    enum {
+        UNDEFINED       = 0x0001,
+        NUMBER          = 0x0002,  // regex: \p{N}
+        LETTER          = 0x0004,  // regex: \p{L}
+        SEPARATOR       = 0x0008,  // regex: \p{Z}
+        ACCENT_MARK     = 0x0010,  // regex: \p{M}
+        PUNCTUATION     = 0x0020,  // regex: \p{P}
+        SYMBOL          = 0x0040,  // regex: \p{S}
+        CONTROL         = 0x0080,  // regex: \p{C}
+        MASK_CATEGORIES = 0x00FF,
+    };
+
+    // codepoint type
+    uint16_t is_undefined   : 1;
+    uint16_t is_number      : 1;  // regex: \p{N}
+    uint16_t is_letter      : 1;  // regex: \p{L}
+    uint16_t is_separator   : 1;  // regex: \p{Z}
+    uint16_t is_accent_mark : 1;  // regex: \p{M}
+    uint16_t is_punctuation : 1;  // regex: \p{P}
+    uint16_t is_symbol      : 1;  // regex: \p{S}
+    uint16_t is_control     : 1;  // regex: \p{C}
+    // helper flags
+    uint16_t is_whitespace  : 1;  // regex: \s
+    uint16_t is_lowercase   : 1;
+    uint16_t is_uppercase   : 1;
+    uint16_t is_nfd         : 1;
+
+    // decode from uint16
+    inline codepoint_flags(const uint16_t flags=0) {
+        *reinterpret_cast<uint16_t*>(this) = flags;
+    }
+
+    inline uint16_t as_uint() const {
+        return *reinterpret_cast<const uint16_t*>(this);
+    }
+
+    inline uint16_t category_flag() const {
+        return this->as_uint() & MASK_CATEGORIES;
+    }
+};
+
 
 std::string unicode_cpt_to_utf8(uint32_t cp);
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-int unicode_cpt_type(uint32_t cp);
-int unicode_cpt_type(const std::string & utf8);
-
-bool unicode_cpt_is_whitespace(uint32_t cp);
+codepoint_flags unicode_cpt_flags(const uint32_t cp);
+codepoint_flags unicode_cpt_flags(const std::string & utf8);
 
 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);
 
-char32_t unicode_tolower(char32_t cp);
+uint32_t unicode_tolower(uint32_t cp);
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

examples/talk/gpt-2.cpp

@@ -72,7 +72,7 @@ struct gpt2_model {
 };
 
 // load the model's weights from a file
-bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
+static bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
     printf("%s: loading model from '%s'\n", __func__, fname.c_str());
 
     auto fin = std::ifstream(fname, std::ios::binary);
@@ -380,7 +380,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 //   - embd_w:    the predicted logits for the next token
 //
 // TODO: sync latest version from ggml repo
-bool gpt2_eval(
+static bool gpt2_eval(
         const gpt2_model & model,
         const int n_threads,
         const int n_past,

examples/talk/talk.cpp

@@ -26,7 +26,6 @@ struct whisper_params {
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
@@ -45,7 +44,7 @@ struct whisper_params {
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
 
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -60,7 +59,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
@@ -96,7 +94,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
@@ -112,7 +109,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
+static std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
     prob = 0.0f;
@@ -132,7 +129,6 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";

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

@@ -26,7 +26,6 @@ struct whisper_params {
 
     float grammar_penalty = 100.0f;
 
-    bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
@@ -57,7 +56,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
     fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
     fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
@@ -89,7 +87,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
         else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }

examples/whisper.android.java/app/src/main/jni/whisper/CMakeLists.txt

@@ -5,15 +5,14 @@ project(whisper.cpp)
 set(CMAKE_CXX_STANDARD 11)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)
 
-set(
-        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
-)
+set(SOURCE_FILES
+    ${WHISPER_LIB_DIR}/ggml/src/ggml.c
+    ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
+    ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.c
+    ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
+    ${WHISPER_LIB_DIR}/src/whisper.cpp
+    ${CMAKE_SOURCE_DIR}/jni.c
+    )
 
 find_library(LOG_LIB log)
 
@@ -41,7 +40,6 @@ 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()
 
@@ -54,3 +52,7 @@ elseif (${ANDROID_ABI} STREQUAL "armeabi-v7a")
 endif ()
 
 include_directories(${WHISPER_LIB_DIR})
+include_directories(${WHISPER_LIB_DIR}/src)
+include_directories(${WHISPER_LIB_DIR}/include)
+include_directories(${WHISPER_LIB_DIR}/ggml/include)
+include_directories(${WHISPER_LIB_DIR}/ggml/src)

examples/whisper.android/README.md

@@ -12,47 +12,3 @@ 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/src/main/jni/whisper/CMakeLists.txt

@@ -10,7 +10,7 @@ option(GGML_HOME       "whisper: Path to external GGML source" OFF)
 
 set(
     SOURCE_FILES
-    ${WHISPER_LIB_DIR}/whisper.cpp
+    ${WHISPER_LIB_DIR}/src/whisper.cpp
     ${CMAKE_SOURCE_DIR}/jni.c
     )
 
@@ -18,10 +18,10 @@ 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}/ggml/src/ggml.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
         )
 endif()
 
@@ -75,3 +75,7 @@ endif ()
 build_library("whisper") # Default target
 
 include_directories(${WHISPER_LIB_DIR})
+include_directories(${WHISPER_LIB_DIR}/src)
+include_directories(${WHISPER_LIB_DIR}/include)
+include_directories(${WHISPER_LIB_DIR}/ggml/include)
+include_directories(${WHISPER_LIB_DIR}/ggml/src)

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -9,7 +9,6 @@
 /* Begin PBXBuildFile section */
 		1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 184447182AB211A2007D6BFE /* ggml-alloc.c */; };
 		1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 1844471B2AB21655007D6BFE /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-framework Foundation -framework Metal -framework MetalKit -fno-objc-arc"; }; };
-		184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7A29052BDF00BD2A04 /* AppDelegate.m */; };
 		18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7D29052BDF00BD2A04 /* SceneDelegate.m */; };
 		18627C8129052BDF00BD2A04 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8029052BDF00BD2A04 /* ViewController.m */; };
@@ -20,6 +19,8 @@
 		18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK -DGGML_USE_METAL"; }; };
 		18627C9629052C5800BD2A04 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9529052C5800BD2A04 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL"; }; };
 		18627C9B29052CFF00BD2A04 /* ggml-base.en.bin in Resources */ = {isa = PBXBuildFile; fileRef = 18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */; };
+		18A276062C2A98A5001C8D37 /* ggml-metal.metal in Copy Files */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
+		18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18ABE15A2AF556340044A204 /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.c */; };
 		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
@@ -29,23 +30,24 @@
 /* End PBXBuildFile section */
 
 /* Begin PBXCopyFilesBuildPhase section */
-		184447202AB21B25007D6BFE /* CopyFiles */ = {
+		184447202AB21B25007D6BFE /* Copy Files */ = {
 			isa = PBXCopyFilesBuildPhase;
 			buildActionMask = 2147483647;
 			dstPath = "";
 			dstSubfolderSpec = 7;
 			files = (
-				184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */,
+				18A276062C2A98A5001C8D37 /* ggml-metal.metal in Copy Files */,
 			);
+			name = "Copy Files";
 			runOnlyForDeploymentPostprocessing = 0;
 		};
 /* End PBXCopyFilesBuildPhase section */
 
 /* Begin PBXFileReference section */
-		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml-alloc.c"; sourceTree = "<group>"; };
-		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml-alloc.h"; sourceTree = "<group>"; };
-		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml-metal.m"; sourceTree = "<group>"; };
-		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml-metal.metal"; sourceTree = "<group>"; };
+		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml/src/ggml-alloc.c"; sourceTree = "<group>"; };
+		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml/include/ggml-alloc.h"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@@ -58,17 +60,19 @@
 		18627C8829052BE000BD2A04 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
 		18627C8A29052BE000BD2A04 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
 		18627C8B29052BE000BD2A04 /* main.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = main.m; sourceTree = "<group>"; };
-		18627C9229052C2B00BD2A04 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../whisper.h; sourceTree = "<group>"; };
-		18627C9329052C4900BD2A04 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../whisper.cpp; sourceTree = "<group>"; };
-		18627C9529052C5800BD2A04 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml.c; sourceTree = "<group>"; };
-		18627C9729052C6600BD2A04 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml.h; sourceTree = "<group>"; };
+		18627C9229052C2B00BD2A04 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../include/whisper.h; sourceTree = "<group>"; };
+		18627C9329052C4900BD2A04 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../src/whisper.cpp; sourceTree = "<group>"; };
+		18627C9529052C5800BD2A04 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml/src/ggml.c; sourceTree = "<group>"; };
+		18627C9729052C6600BD2A04 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml/include/ggml.h; sourceTree = "<group>"; };
 		18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; name = "ggml-base.en.bin"; path = "../../../models/ggml-base.en.bin"; sourceTree = "<group>"; };
-		18ABE1542AF556340044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../../ggml-quants.h"; sourceTree = "<group>"; };
-		18ABE1552AF556340044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../../ggml-backend.h"; sourceTree = "<group>"; };
-		18ABE1562AF556340044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend-impl.h"; path = "../../../ggml-backend-impl.h"; sourceTree = "<group>"; };
-		18ABE1572AF556340044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../../ggml-backend.c"; sourceTree = "<group>"; };
-		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml-impl.h"; sourceTree = "<group>"; };
-		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml-quants.c"; sourceTree = "<group>"; };
+		18A275FE2C2A94DE001C8D37 /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../../ggml/include/ggml-metal.h"; sourceTree = "<group>"; };
+		18A275FF2C2A9563001C8D37 /* ggml-common.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-common.h"; path = "../../../ggml/src/ggml-common.h"; sourceTree = "<group>"; };
+		18ABE1542AF556340044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../../ggml/src/ggml-quants.h"; sourceTree = "<group>"; };
+		18ABE1552AF556340044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../../ggml/include/ggml-backend.h"; sourceTree = "<group>"; };
+		18ABE1562AF556340044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend-impl.h"; path = "../../../ggml/src/ggml-backend-impl.h"; sourceTree = "<group>"; };
+		18ABE1572AF556340044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../../ggml/src/ggml-backend.c"; sourceTree = "<group>"; };
+		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
+		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@@ -108,6 +112,8 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				18A275FF2C2A9563001C8D37 /* ggml-common.h */,
+				18A275FE2C2A94DE001C8D37 /* ggml-metal.h */,
 				18ABE1562AF556340044A204 /* ggml-backend-impl.h */,
 				18ABE1572AF556340044A204 /* ggml-backend.c */,
 				18ABE1552AF556340044A204 /* ggml-backend.h */,
@@ -151,7 +157,7 @@
 				7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */,
 			);
 			name = coreml;
-			path = ../../../coreml;
+			path = ../../../src/coreml;
 			sourceTree = "<group>";
 		};
 /* End PBXGroup section */
@@ -164,7 +170,7 @@
 				18627C7229052BDF00BD2A04 /* Sources */,
 				18627C7329052BDF00BD2A04 /* Frameworks */,
 				18627C7429052BDF00BD2A04 /* Resources */,
-				184447202AB21B25007D6BFE /* CopyFiles */,
+				184447202AB21B25007D6BFE /* Copy Files */,
 			);
 			buildRules = (
 			);
@@ -182,7 +188,7 @@
 			isa = PBXProject;
 			attributes = {
 				BuildIndependentTargetsInParallel = 1;
-				LastUpgradeCheck = 1400;
+				LastUpgradeCheck = 1540;
 				TargetAttributes = {
 					18627C7529052BDF00BD2A04 = {
 						CreatedOnToolsVersion = 14.0.1;
@@ -212,6 +218,7 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */,
 				18627C8929052BE000BD2A04 /* LaunchScreen.storyboard in Resources */,
 				7FE3424F2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc in Resources */,
 				18627C8629052BE000BD2A04 /* Assets.xcassets in Resources */,
@@ -301,6 +308,7 @@
 				DEBUG_INFORMATION_FORMAT = dwarf;
 				ENABLE_STRICT_OBJC_MSGSEND = YES;
 				ENABLE_TESTABILITY = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GCC_C_LANGUAGE_STANDARD = gnu11;
 				GCC_DYNAMIC_NO_PIC = NO;
 				GCC_NO_COMMON_BLOCKS = YES;
@@ -359,6 +367,7 @@
 				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
 				ENABLE_NS_ASSERTIONS = NO;
 				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GCC_C_LANGUAGE_STANDARD = gnu11;
 				GCC_NO_COMMON_BLOCKS = YES;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@@ -400,6 +409,7 @@
 					"@executable_path/Frameworks",
 				);
 				MARKETING_VERSION = 1.0;
+				MTL_HEADER_SEARCH_PATHS = "";
 				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SWIFT_EMIT_LOC_STRINGS = YES;
@@ -428,6 +438,7 @@
 					"@executable_path/Frameworks",
 				);
 				MARKETING_VERSION = 1.0;
+				MTL_HEADER_SEARCH_PATHS = "";
 				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SWIFT_EMIT_LOC_STRINGS = YES;

examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift

@@ -15,7 +15,7 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
     private var audioPlayer: AVAudioPlayer?
     
     private var modelUrl: URL? {
-        Bundle.main.url(forResource: "ggml-tiny.en", withExtension: "bin", subdirectory: "models")
+        Bundle.main.url(forResource: "ggml-base.en", withExtension: "bin", subdirectory: "models")
     }
     
     private var sampleUrl: URL? {

ggml-cuda/concat.cu

@@ -1,49 +0,0 @@
-#include "concat.cuh"
-
-static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    // operation
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    if (blockIdx.z < ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
-    concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
-}
-
-void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-        concat_f32_cuda(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4), dst_d + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], stream);
-    }
-}

ggml-cuda/fattn-common.cuh

@@ -1,47 +0,0 @@
-#define FATTN_KQ_STRIDE       256
-#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
-#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
-
-template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_combine_results(
-        const float  * __restrict__ VKQ_parts,
-        const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
-
-    const int tid = threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float2 meta[parallel_blocks];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
-    }
-
-    __syncthreads();
-
-    float kqmax = meta[0].x;
-#pragma unroll
-    for (int l = 1; l < parallel_blocks; ++l) {
-        kqmax = max(kqmax, meta[l].x);
-    }
-
-    float VKQ_numerator   = 0.0f;
-    float VKQ_denominator = 0.0f;
-#pragma unroll
-    for (int l = 0; l < parallel_blocks; ++l) {
-        const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
-        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
-        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
-
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
-        VKQ_denominator += KQ_max_scale * meta[l].y;
-    }
-
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
-}

ggml-cuda/fattn-vec-f16.cu

@@ -1,430 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-#include "fattn-vec-f16.cuh"
-
-template<int D, int ncols, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_vec_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-#if FP16_AVAILABLE
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
-
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
-
-    const int stride_KV  = nb11 / sizeof(half);
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    half  slopeh = __float2half(1.0f);
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slopeh = __float2half(powf(base, exph));
-    }
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ half KQ[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -HALF_MAX_HALF;
-    }
-    half2 * KQ2 = (half2 *) KQ;
-
-    half kqmax[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -HALF_MAX_HALF;
-    }
-    half kqsum[ncols] = {0.0f};
-
-    __shared__ half kqmax_shared[ncols][WARP_SIZE];
-    __shared__ half kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-    __syncthreads();
-
-    // Convert Q to half2 and store in registers:
-    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-            const int i = i0 + threadIdx.x;
-
-            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
-            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
-        }
-    }
-
-    half2 VKQ[ncols] = {{0.0f, 0.0f}};
-
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
-        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
-        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
-        half kqmax_new = kqmax[0];
-        half kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-            half2 sum2[ncols] = {{0.0f, 0.0f}};
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
-                const int k_KQ = k_KQ_0 + threadIdx.x;
-
-                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-#pragma unroll
-                for (int j = 0; j < ncols; ++j) {
-                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
-                }
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                sum2[j] = warp_reduce_sum(sum2[j]);
-                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
-                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
-
-                if (ncols == 1) {
-                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
-                } else {
-                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
-                }
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum;
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
-
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= __half2half2(KQ_max_scale);
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < D; k0 += 2) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
-                break;
-            }
-
-            half2 V_k;
-            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
-            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
-            }
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum(kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
-
-        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
-    }
-
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
-    }
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
-void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
-
-    const int32_t precision = KQV->op_params[2];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
-    constexpr int cols_per_block = 1;
-    constexpr int parallel_blocks = 4;
-    switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 256:
-            launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-}
-
-void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
-
-    const int32_t precision = KQV->op_params[2];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
-
-    if (Q->ne[1] == 1) {
-        constexpr int cols_per_block = 1;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block = 8;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 8;
-    constexpr int parallel_blocks = 1;
-    switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-}

ggml-cuda/fattn-vec-f16.cuh

@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
-
-void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-vec-f32.cu

@@ -1,384 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-#include "fattn-vec-f32.cuh"
-
-template<int D, int ncols, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_vec_ext_f32(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
-
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
-
-    const int stride_KV  = nb11 / sizeof(half);
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    float slope = 1.0f;
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slope = powf(base, exph);
-    }
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float KQ[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -FLT_MAX/2.0f;
-    }
-
-    float kqmax[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -FLT_MAX/2.0f;
-    }
-    float kqsum[ncols] = {0.0f};
-
-    __shared__ float kqmax_shared[ncols][WARP_SIZE];
-    __shared__ float kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-    __syncthreads();
-
-    // Convert Q to half2 and store in registers:
-    float2 Q_h2[ncols][D/(2*WARP_SIZE)];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-            const int i = i0 + threadIdx.x;
-
-            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
-            Q_h2[j][i0/WARP_SIZE].x *= scale;
-            Q_h2[j][i0/WARP_SIZE].y *= scale;
-        }
-    }
-
-    float VKQ[ncols] = {0.0f};
-
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        float kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-            float sum[ncols] = {0.0f};
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
-                const int k_KQ = k_KQ_0 + threadIdx.x;
-
-                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-#pragma unroll
-                for (int j = 0; j < ncols; ++j) {
-                    sum[j] +=  __low2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].x;
-                    sum[j] += __high2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].y;
-                }
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                sum[j] = warp_reduce_sum(sum[j]);
-                sum[j] += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
-
-                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum[j]);
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum[j];
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            float kqmax_new_j = kqmax_new_arr[j];
-
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const float val = expf(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= KQ_max_scale;
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k = 0; k < D; ++k) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
-                break;
-            }
-
-            const float V_ki = __half2float(V_h[(k_VKQ_0 + k)*stride_KV + tid]);
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_ki*KQ[j*D + k];
-            }
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum(kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
-
-        float dst_val = VKQ[j_VKQ];
-        if (parallel_blocks == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
-    }
-
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
-    }
-}
-
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
-void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    const ggml_tensor * K = dst->src[1];
-    const ggml_tensor * V = dst->src[2];
-
-    const ggml_tensor * mask = dst->src[3];
-
-    ggml_tensor * KQV = dst;
-
-    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
-
-    if (Q->ne[1] == 1) {
-        constexpr int cols_per_block = 1;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block = 8;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 8;
-    constexpr int parallel_blocks = 1;
-    switch (Q->ne[0]) {
-        case 64:
-            launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        case 128:
-            launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-}

ggml-cuda/fattn-vec-f32.cuh

@@ -1,3 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/mmq.cuh

@@ -1,9 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_op_mul_mat_q(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream);
-
-bool ggml_cuda_supports_mmq(enum ggml_type type);

ggml-cuda/quantize.cu

@@ -1,45 +0,0 @@
-#include "quantize.cuh"
-
-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx_padded) {
-    const int64_t ix = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (ix >= kx_padded) {
-        return;
-    }
-
-    const int64_t iy = (int64_t)blockDim.y*blockIdx.y + threadIdx.y;
-
-    const int64_t i_padded = (int64_t)iy*kx_padded + ix;
-
-    block_q8_1 * y = (block_q8_1 *) vy;
-
-    const int64_t ib = i_padded / QK8_1; // block index
-    const int64_t iqs = i_padded % QK8_1; // quant index
-
-    const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
-    float amax = fabsf(xi);
-    float sum = xi;
-
-    amax = warp_reduce_max(amax);
-    sum = warp_reduce_sum(sum);
-
-    const float d = amax / 127;
-    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
-
-    y[ib].qs[iqs] = q;
-
-    if (iqs > 0) {
-        return;
-    }
-
-    reinterpret_cast<half&>(y[ib].ds.x) = d;
-    reinterpret_cast<half&>(y[ib].ds.y) = sum;
-}
-
-void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream) {
-    const int64_t block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, ky, 1);
-    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
-}
-

ggml-cuda/quantize.cuh

@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_QUANTIZE_BLOCK_SIZE 256
-
-void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream);

ggml-cuda/rope.cu

@@ -1,308 +0,0 @@
-#include "rope.cuh"
-
-struct rope_corr_dims {
-    float v[4];
-};
-
-static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
-// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
-static __device__ void rope_yarn(
-    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
-    float * cos_theta, float * sin_theta
-) {
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
-    }
-    *cos_theta = cosf(theta) * mscale;
-    *sin_theta = sinf(theta) * mscale;
-}
-
-// rope == RoPE == rotary positional embedding
-template<typename T, bool has_pos>
-static __global__ void rope(
-    const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims
-) {
-    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i = row*ncols + col;
-    const int i2 = row/p_delta_rows;
-
-    const int p = has_pos ? pos[i2] : 0;
-    const float theta_base = p*powf(freq_base, -float(col)/ncols);
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + 1];
-
-    dst[i + 0] = x0*cos_theta - x1*sin_theta;
-    dst[i + 1] = x0*sin_theta + x1*cos_theta;
-}
-
-template<typename T, bool has_pos>
-static __global__ void rope_neox(
-    const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
-) {
-    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.x*blockIdx.x + threadIdx.x;
-    const int ib = col / n_dims;
-    const int ic = col % n_dims;
-
-    if (ib > 0) {
-        const int i = row*ncols + ib*n_dims + ic;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
-        return;
-    }
-
-    const int i  = row*ncols + ib*n_dims + ic/2;
-    const int i2 = row/p_delta_rows;
-
-    float cur_rot = inv_ndims * ic - ib;
-
-    const int p = has_pos ? pos[i2] : 0;
-    const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + n_dims/2];
-
-    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
-}
-
-static __global__ void rope_glm_f32(
-    const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
-    int n_ctx
-) {
-    const int col = blockDim.x*blockIdx.x + threadIdx.x;
-    const int half_n_dims = ncols/4;
-
-    if (col >= half_n_dims) {
-        return;
-    }
-
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i = row*ncols + col;
-    const int i2 = row/p_delta_rows;
-
-    const float col_theta_scale = powf(freq_base, -2.0f*col/ncols);
-     // FIXME: this is likely wrong
-    const int p = pos != nullptr ? pos[i2] : 0;
-
-    const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale;
-    const float sin_theta = sinf(theta);
-    const float cos_theta = cosf(theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + half_n_dims];
-
-    dst[i + 0]           = x0*cos_theta - x1*sin_theta;
-    dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta;
-
-    const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale;
-    const float sin_block_theta = sinf(block_theta);
-    const float cos_block_theta = cosf(block_theta);
-
-    const float x2 = x[i + half_n_dims * 2];
-    const float x3 = x[i + half_n_dims * 3];
-
-    dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta;
-    dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
-}
-
-
-template<typename T>
-static void rope_cuda(
-    const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 2 == 0);
-    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
-    const dim3 block_nums(nrows, num_blocks_x, 1);
-    if (pos == nullptr) {
-        rope<T, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
-        );
-    } else {
-        rope<T, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
-        );
-    }
-}
-
-template<typename T>
-static void rope_neox_cuda(
-    const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 2 == 0);
-    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
-    const dim3 block_nums(nrows, num_blocks_x, 1);
-
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
-    const float inv_ndims = -1.0f / n_dims;
-
-    if (pos == nullptr) {
-        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
-        );
-    } else {
-        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
-        );
-    }
-}
-
-static void rope_glm_f32_cuda(
-    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, int n_ctx, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 4 == 0);
-    const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
-    const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
-    const dim3 block_nums(num_blocks_x, nrows, 1);
-    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx);
-}
-
-static void rope_cuda_f16(
-    const half * x, half * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
-
-    rope_cuda<half>(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
-}
-
-static void rope_cuda_f32(
-    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
-
-    rope_cuda<float>(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
-}
-
-static void rope_neox_cuda_f16(
-    const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
-
-    rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
-}
-
-static void rope_neox_cuda_f32(
-    const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
-) {
-
-    rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
-}
-
-void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
-    GGML_ASSERT(src0->type == dst->type);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne2 = dst->ne[2];
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past      = ((int32_t *) dst->op_params)[0];
-    const int n_dims      = ((int32_t *) dst->op_params)[1];
-    const int mode        = ((int32_t *) dst->op_params)[2];
-    const int n_ctx       = ((int32_t *) dst->op_params)[3];
-    const int n_orig_ctx  = ((int32_t *) dst->op_params)[4];
-
-    // RoPE alteration for extended context
-    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-
-    const int32_t * pos = nullptr;
-    if ((mode & 1) == 0) {
-        GGML_ASSERT(src1->type == GGML_TYPE_I32);
-        GGML_ASSERT(src1->ne[0] == ne2);
-        pos = (const int32_t *) src1_d;
-    }
-
-    const bool is_neox = mode & 2;
-    const bool is_glm  = mode & 4;
-
-    rope_corr_dims corr_dims;
-    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
-
-    // compute
-    if (is_glm) {
-        GGML_ASSERT(false);
-        rope_glm_f32_cuda(src0_d, dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, stream);
-    } else if (is_neox) {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_neox_cuda_f32(
-                (const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
-            );
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_neox_cuda_f16(
-                (const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
-            );
-        } else {
-            GGML_ASSERT(false);
-        }
-    } else {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_cuda_f32(
-                (const float *)src0_d, (float *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
-            );
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_cuda_f16(
-                (const half *)src0_d, (half *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
-            );
-        } else {
-            GGML_ASSERT(false);
-        }
-    }
-}

ggml-cuda/upscale.cu

@@ -1,48 +0,0 @@
-#include "upscale.cuh"
-
-static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) {
-    // blockIdx.z: idx of ne02*ne03
-    // blockIdx.y: idx of ne01*scale_factor， aka ne1
-    // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE
-    // ne00xne01: ne00 * ne01
-    int ne0 = ne00 * scale_factor;
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    // operation
-    int i00 = nidx / scale_factor;
-    int i01 = blockIdx.y / scale_factor;
-    int offset_src =
-        i00 +
-        i01 * ne00 +
-        blockIdx.z * ne00xne01;
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    dst[offset_dst] = x[offset_src];
-}
-
-static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03,
-                             const int scale_factor, cudaStream_t stream) {
-    int ne0 = (ne00 * scale_factor);
-    int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03);
-    upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
-}
-
-void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-
-    const int scale_factor = dst->op_params[0];
-
-    upscale_f32_cuda(src0_d, dst_d, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, stream);
-}

ggml/.gitignore

@@ -0,0 +1 @@
+src/ggml-metal-embed.metal

ggml/CMakeLists.txt

@@ -0,0 +1,238 @@
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+project("ggml" C CXX)
+include(CheckIncludeFileCXX)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(GGML_STANDALONE ON)
+
+    set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+    # configure project version
+    # TODO
+else()
+    set(GGML_STANDALONE OFF)
+endif()
+
+if (EMSCRIPTEN)
+    set(BUILD_SHARED_LIBS_DEFAULT OFF)
+
+    option(GGML_WASM_SINGLE_FILE "ggml: embed WASM inside the generated ggml.js" ON)
+else()
+    if (MINGW)
+        set(BUILD_SHARED_LIBS_DEFAULT OFF)
+    else()
+        set(BUILD_SHARED_LIBS_DEFAULT ON)
+    endif()
+endif()
+
+option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+
+#
+# option list
+#
+
+# TODO: mark all options as advanced when not GGML_STANDALONE
+
+if (APPLE)
+    set(GGML_METAL_DEFAULT ON)
+    set(GGML_BLAS_DEFAULT ON)
+    set(GGML_BLAS_VENDOR_DEFAULT "Apple")
+else()
+    set(GGML_METAL_DEFAULT OFF)
+    set(GGML_BLAS_DEFAULT OFF)
+    set(GGML_BLAS_VENDOR_DEFAULT "Generic")
+endif()
+
+# general
+option(GGML_STATIC "ggml: static link libraries"         OFF)
+option(GGML_NATIVE "ggml: enable -march=native flag"     ON)
+option(GGML_LTO    "ggml: enable link time optimization" OFF)
+option(GGML_CCACHE "ggml: use ccache if available"       ON)
+
+# debug
+option(GGML_ALL_WARNINGS           "ggml: enable all compiler warnings"                   ON)
+option(GGML_ALL_WARNINGS_3RD_PARTY "ggml: enable all compiler warnings in 3rd party libs" OFF)
+option(GGML_GPROF                  "ggml: enable gprof"                                   OFF)
+
+# build
+option(GGML_FATAL_WARNINGS    "ggml: enable -Werror flag"    OFF)
+
+# sanitizers
+option(GGML_SANITIZE_THREAD    "ggml: enable thread sanitizer"    OFF)
+option(GGML_SANITIZE_ADDRESS   "ggml: enable address sanitizer"   OFF)
+option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
+
+# instruction set specific
+if (GGML_NATIVE)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
+
+option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512      "ggml: enable AVX512"           OFF)
+option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
+option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+if (NOT MSVC)
+    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+endif()
+option(GGML_LASX        "ggml: enable lasx"             ON)
+option(GGML_LSX         "ggml: enable lsx"              ON)
+option(GGML_SVE         "ggml: enable SVE"              OFF)
+
+if (WIN32)
+    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+endif()
+
+# ggml core
+set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+
+# 3rd party libs / backends
+option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
+option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
+set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
+                                            "ggml: BLAS library vendor")
+option(GGML_LLAMAFILE                       "ggml: use ggml SGEMM"                            OFF)
+
+option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
+set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
+set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
+option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
+set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
+                                            "ggml: iters./thread per block for Q2_K/Q6_K")
+set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                            "ggml: max. batch size for using peer access")
+option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
+option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
+option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
+
+option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
+option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
+option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
+option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
+option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
+option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
+option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
+option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
+option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
+option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
+set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
+                                            "ggml: metal minimum macOS version")
+set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
+option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_RPC                             "ggml: use RPC"                                   OFF)
+option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
+option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
+set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
+                                            "ggml: sycl target device")
+
+# extra artifacts
+option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
+option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
+
+#
+# dependencies
+#
+
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+if (GGML_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+
+find_package(Threads REQUIRED)
+
+#
+# build the library
+#
+
+add_subdirectory(src)
+
+#
+# tests and examples
+#
+
+if (GGML_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif ()
+
+if (GGML_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif ()
+
+#
+# install
+#
+
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+set(GGML_PUBLIC_HEADERS
+    include/ggml.h
+    include/ggml-alloc.h
+    include/ggml-backend.h
+    "${GGML_HEADERS_CUDA}"
+    "${GGML_HEADERS_METAL}"
+    "${GGML_HEADERS_EXTRA}")
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+#if (GGML_METAL)
+#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
+#endif()
+install(TARGETS ggml PUBLIC_HEADER)
+
+if (BUILD_SHARED_LIBS)
+    install(TARGETS ggml LIBRARY)
+endif()
+
+if (GGML_METAL)
+    install(
+        FILES src/ggml-metal.metal
+        PERMISSIONS
+            OWNER_READ
+            OWNER_WRITE
+            GROUP_READ
+            WORLD_READ
+        DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+    if (NOT GGML_METAL_EMBED_LIBRARY)
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+    endif()
+endif()
+
+if (GGML_STANDALONE)
+    configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        @ONLY)
+
+    install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        DESTINATION share/pkgconfig)
+endif()

ggml/cmake/FindSIMD.cmake

@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(GGML_AVX OFF)
+else()
+    set(GGML_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(GGML_AVX2 OFF)
+else()
+    set(GGML_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(GGML_AVX512 OFF)
+else()
+    set(GGML_AVX512 ON)
+endif()

ggml/ggml_vk_generate_shaders.py

@@ -0,0 +1,220 @@
+#!/usr/bin/env python
+
+import logging
+import argparse
+import asyncio
+import os
+from tempfile import gettempdir
+
+logger = logging.getLogger("ggml-vk-generate-shaders")
+
+GLSLC = "glslc"
+
+type_names = [
+    "f32",
+    "f16",
+    "q4_0",
+    "q4_1",
+    "q5_0",
+    "q5_1",
+    "q8_0",
+    "q2_k",
+    "q3_k",
+    "q4_k",
+    "q5_k",
+    "q6_k",
+]
+
+ASYNCIO_CONCURRENCY = 64
+
+input_dir = "vulkan-shaders"
+output_dir = gettempdir()
+
+lock = asyncio.Lock()
+shader_fnames = []
+
+
+async def string_to_spv(name, in_fname, defines, fp16=True):
+    name = f"{name}{'_fp32' if not fp16 else ''}"
+    out_fname = os.path.join(output_dir, f"{name}.spv")
+
+    in_path = os.path.join(input_dir, in_fname)
+
+    cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname]
+
+    cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
+
+    proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
+
+    stdout, stderr = await proc.communicate()
+
+    stdout = stdout.decode()
+    error = stderr.decode()
+
+    if proc.returncode:
+        cmd = " ".join(cmd)
+        logger.error(f"cannot compile {name}\n\n{cmd}\n\n{error}")
+        return
+
+    async with lock:
+        shader_fnames.append((name, out_fname))
+
+
+def matmul_shaders(tasks, fp16, matmul_id):
+    if fp16:
+        load_vec = "8"
+        aligned_b_type_f32 = "mat2x4"
+        aligned_b_type_f16 = "f16mat2x4"
+    else:
+        load_vec = "4"
+        aligned_b_type_f32 = "vec4"
+        aligned_b_type_f16 = "f16vec4"
+
+    base_dict = {"FLOAT_TYPE": "float" if not fp16 else "float16_t"}
+    shader_name = "matmul"
+
+    if matmul_id:
+        base_dict["MUL_MAT_ID"] = "1"
+        shader_name = "matmul_id"
+
+    if fp16:
+        base_dict["FLOAT16"] = "1"
+
+    # Shaders with f16 B_TYPE
+    tasks.append(string_to_spv(f"{shader_name}_f32_f16", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
+    tasks.append(string_to_spv(f"{shader_name}_f32_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
+
+    tasks.append(string_to_spv(f"{shader_name}_f16", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
+    tasks.append(string_to_spv(f"{shader_name}_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
+
+    for tname in type_names:
+        data_a_key = f"DATA_A_{tname.upper()}"
+        load_vec_a = load_vec if tname in ("f32", "f16") else "2"
+        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32", "mul_mm.comp", base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32_aligned", "mul_mm.comp", base_dict | {data_a_key: "2", "LOAD_VEC_A": load_vec_a, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f32, "D_TYPE": "float"}, fp16))
+
+
+async def main():
+    logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
+
+    tasks = []
+
+    for fp16 in (False, True):
+        # MUL_MAT
+        matmul_shaders(tasks, fp16, False)
+        # MUL_MAT_ID
+        matmul_shaders(tasks, fp16, True)
+
+    for tname in type_names:
+        base_dict = {"FLOAT_TYPE": "float"}
+
+        # mul mat vec
+        data_a_key = f"DATA_A_{tname.upper()}"
+        shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp"
+
+        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f32_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
+        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f16_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float16_t", "D_TYPE": "float"}))
+
+        tasks.append(string_to_spv(f"mul_mat_vec_id_{tname}_f32", shader, base_dict | {"MUL_MAT_ID": "1", data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
+
+        # Dequant shaders
+        if tname != "f16":
+            tasks.append(string_to_spv(f"dequant_{tname}", f"dequant_{tname}.comp", base_dict | {data_a_key: "1", "D_TYPE": "float16_t"}))
+
+        # get_rows
+        if not tname.endswith("_k"):
+            shader = "get_rows.comp" if tname in ("f32", "f16") else "get_rows_quant.comp"
+
+            if tname == "f16":
+                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
+            else:
+                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t"}))
+            tasks.append(string_to_spv(f"get_rows_{tname}_f32", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
+
+    # Norms
+    tasks.append(string_to_spv("norm_f32", "norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rms_norm_f32", "rms_norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("cpy_f32_f32", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("cpy_f32_f16", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
+    tasks.append(string_to_spv("cpy_f16_f16", "copy.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
+
+    tasks.append(string_to_spv("add_f32", "add.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}))
+
+    tasks.append(string_to_spv("mul_f32", "mul.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("div_f32", "div.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("scale_f32", "scale.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("sqr_f32", "square.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("clamp_f32", "clamp.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("gelu_f32", "gelu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("silu_f32", "silu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("relu_f32", "relu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("soft_max_f32", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("soft_max_f32_f16", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("rope_norm_f32", "rope_norm.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rope_norm_f16", "rope_norm.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
+
+    tasks.append(string_to_spv("rope_neox_f32", "rope_neox.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rope_neox_f16", "rope_neox.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
+
+    tasks.append(string_to_spv("argsort_f32", "argsort.comp", {"A_TYPE": "float"}))
+
+    tasks.append(string_to_spv("sum_rows_f32", "sum_rows.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    # Helper to decorate tasks with semaphore acquisition.
+    async def withSemaphore(sem, task):
+        async with sem:
+            return await task
+
+    # Run tasks concurrently guarded by a concurrency limit.
+    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
+    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
+
+    with open("ggml-vulkan-shaders.hpp", "w") as f:
+        f.write("#include <cstdint>\n\n")
+        for name, path in sorted(shader_fnames):
+
+            with open(path, "rb") as spv:
+                counter = 0
+                newline_counter = 0
+                f.write(f"unsigned char {name}_data[] = {{\n")
+                for val in spv.read():
+                    f.write(f"0x{val:02x},")
+                    newline_counter += 1
+                    counter += 1
+                    if newline_counter >= 12:
+                        newline_counter = 0
+                        f.write("\n")
+            f.write("\n};\n")
+            f.write(f"const uint64_t {name}_len = {counter};\n\n")
+            os.remove(path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
+
+    parser.add_argument("--glslc", help="Path to glslc")
+    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+
+    if args.glslc:
+        GLSLC = args.glslc
+
+    asyncio.run(main())

ggml/include/ggml-backend.h

@@ -23,7 +23,6 @@ extern "C" {
     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
@@ -74,6 +73,7 @@ extern "C" {
     GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
     GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
@@ -90,7 +90,7 @@ extern "C" {
     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
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event);
 
     //
     // CPU backend
@@ -119,7 +119,7 @@ extern "C" {
 
     GGML_API size_t                     ggml_backend_reg_get_count(void);
     GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is backend_name:params (params is optional)
     GGML_API const char *               ggml_backend_reg_get_name(size_t i);
     GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
     GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
@@ -182,6 +182,9 @@ extern "C" {
     // Initialize backend buffers from a measure graph
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
 
+    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
+
     // 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);
@@ -225,7 +228,7 @@ extern "C" {
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
 
 #ifdef  __cplusplus

ggml/include/ggml-blas.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend);
+
+// number of threads used for conversion to float
+// for openblas and blis, this will also set the number of threads used for blas operations
+GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-cuda.h

@@ -0,0 +1,44 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_CUDA_MAX_DEVICES       16
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
+
+// device buffer
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+
+GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
+#ifdef  __cplusplus
+}
+#endif

ggml/include/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/include/ggml-metal.h

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

ggml/include/ggml-sycl.h

@@ -0,0 +1,42 @@
+//
+//  MIT license
+//  Copyright (C) 2024 Intel Corporation
+//  SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#define GGML_SYCL_NAME "SYCL"
+#define GGML_SYCL_MAX_DEVICES 48
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+// devide buffer
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+
+GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
+GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
+GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
+GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+
+// SYCL doesn't support registering host memory, keep here for reference
+// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-vulkan.h

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