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

@@ -101,7 +101,10 @@ 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
@@ -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:
@@ -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:
@@ -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:
@@ -381,12 +384,9 @@ 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
 
@@ -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: |
@@ -447,15 +434,11 @@ 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@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:
@@ -498,7 +481,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUDA=${{ matrix.cublas }}
+          -DGGML_CUDA=${{ matrix.cublas }}
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build ${{ matrix.cuda-toolkit }}

.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 1.6.2)
 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)
@@ -37,756 +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)
-
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    option(WHISPER_FFMPEG                 "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
-endif()
-
-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()
-
-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)
-    set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS} ${FFMPEG_LIBRARIES})
-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 CUDA::cufft)
-            else ()
-                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static CUDA::cufft_static)
-            endif()
-        else()
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt CUDA::cufft)
-        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}        ${GGML_HEADERS_ROCM}
-    whisper.h
-    whisper.cpp
-    )
-
-if (WHISPER_CUDA)
-    target_sources(${TARGET} PRIVATE whisper-mel-cuda.cu)
-endif()
-
-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

@@ -20,7 +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)
 
@@ -419,31 +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.
@@ -453,7 +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

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/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)
@@ -22,10 +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(${TARGET} STATIC
     common.h
     common.cpp
@@ -38,7 +63,7 @@ add_library(${TARGET} STATIC
 
 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")
@@ -55,8 +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")
@@ -108,7 +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/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

@@ -59,7 +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];
 
@@ -130,7 +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,
@@ -216,7 +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);
@@ -238,7 +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);
@@ -252,7 +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__);
 
@@ -463,7 +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;
 
@@ -543,7 +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

@@ -30,7 +30,7 @@ extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t>
 #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 {
@@ -346,7 +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;
@@ -358,7 +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>>();

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
 

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

@@ -53,7 +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];
 
@@ -109,7 +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;
@@ -153,7 +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);
@@ -199,7 +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);
@@ -285,7 +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;
 
@@ -325,7 +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,
@@ -335,7 +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 {
@@ -375,7 +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/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;
@@ -94,17 +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];
 
@@ -182,7 +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");
@@ -248,7 +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();
 
@@ -280,7 +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) {
@@ -289,7 +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;
 
@@ -358,7 +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);
@@ -385,7 +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);
@@ -417,7 +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);
@@ -446,7 +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;
     }
@@ -459,7 +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;
     }
@@ -478,7 +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;
     }
@@ -509,7 +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);
@@ -544,7 +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);
 
@@ -563,7 +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,
@@ -734,7 +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);
@@ -859,7 +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);
@@ -900,7 +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;

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/stream/stream.cpp

@@ -44,7 +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];
 

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) {
@@ -83,7 +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];
 
@@ -168,7 +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,
@@ -235,7 +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

@@ -44,7 +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];
 
@@ -109,7 +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;

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-opencl.h

@@ -1,36 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-GGML_API void ggml_cl_init(void);
-
-GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
-GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-// GGML_API void * ggml_cl_host_malloc(size_t size);
-// GGML_API void   ggml_cl_host_free(void * ptr);
-
-GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-// backend API
-
-// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
-
-// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
-// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif

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

@@ -38,6 +38,7 @@ GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t *
 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-metal.h

@@ -1,7 +1,7 @@
 // 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, OpenCL, etc.)
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, etc.)
 //
 // How it works?
 //

ggml/include/ggml-sycl.h

@@ -9,13 +9,13 @@
 #include "ggml.h"
 #include "ggml-backend.h"
 
+#define GGML_SYCL_NAME "SYCL"
+#define GGML_SYCL_MAX_DEVICES 48
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
 // backend API
 GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
 
@@ -33,13 +33,6 @@ 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);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
 
 // 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/include/ggml.h

@@ -312,6 +312,12 @@
     GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
     GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
+#define GGML_TENSOR_BINARY_OP_LOCALS01 \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -481,9 +487,7 @@ extern "C" {
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
 
-        GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_ATTN_EXT,
-        GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_SSM_CONV,
         GGML_OP_SSM_SCAN,
@@ -565,7 +569,8 @@ extern "C" {
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type         type;
-        enum ggml_backend_type backend;
+
+        GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
 
         struct ggml_backend_buffer * buffer;
 
@@ -586,11 +591,7 @@ extern "C" {
         struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
-
+        // source tensor and offset for views
         struct ggml_tensor * view_src;
         size_t               view_offs;
 
@@ -600,7 +601,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        // char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -647,11 +648,6 @@ extern "C" {
         struct ggml_hash_set visited_hash_table;
 
         enum ggml_cgraph_eval_order order;
-
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
     };
 
     // scratch buffer
@@ -668,28 +664,6 @@ extern "C" {
         bool   no_alloc;   // don't allocate memory for the tensor data
     };
 
-
-    // compute types
-
-    // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
-    // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
-    enum ggml_task_type {
-        GGML_TASK_TYPE_INIT = 0,
-        GGML_TASK_TYPE_COMPUTE,
-        GGML_TASK_TYPE_FINALIZE,
-    };
-
-    struct ggml_compute_params {
-        enum ggml_task_type type;
-
-        // ith = thread index, nth = number of threads
-        int ith, nth;
-
-        // work buffer for all threads
-        size_t wsize;
-        void * wdata;
-    };
-
     // numa strategies
     enum ggml_numa_strategy {
         GGML_NUMA_STRATEGY_DISABLED   = 0,
@@ -757,7 +731,6 @@ extern "C" {
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
     GGML_API GGML_CALL bool ggml_is_transposed(const struct ggml_tensor * tensor);
-    GGML_API GGML_CALL bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API GGML_CALL bool ggml_is_permuted  (const struct ggml_tensor * tensor);
     GGML_API GGML_CALL bool ggml_is_empty     (const struct ggml_tensor * tensor);
     GGML_API           bool ggml_is_scalar    (const struct ggml_tensor * tensor);
@@ -766,6 +739,11 @@ extern "C" {
     GGML_API           bool ggml_is_3d        (const struct ggml_tensor * tensor);
     GGML_API           int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
+    GGML_API GGML_CALL bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
+    GGML_API GGML_CALL bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
+    GGML_API GGML_CALL bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
+    GGML_API GGML_CALL bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -1008,12 +986,13 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // concat a and b on dim 2
+    // concat a and b along dim
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_concat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   dim);
 
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
@@ -1459,18 +1438,17 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
+    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
     // if mode & 2 == 1, GPT-NeoX style
-    // if mode & 4 == 1, ChatGLM style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_inplace(
@@ -1478,18 +1456,17 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // custom RoPE
-    GGML_API struct ggml_tensor * ggml_rope_custom(
+    GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1498,14 +1475,14 @@ extern "C" {
             float                 beta_slow);
 
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+    GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1513,18 +1490,39 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
-    // compute correction dims for YaRN RoPE scaling
-    GGML_CALL void ggml_rope_yarn_corr_dims(
-        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
-
-    // xPos RoPE, in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            float                 base,
-            bool                  down);
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext instead");
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext_inplace instead");
+
+    // compute correction dims for YaRN RoPE scaling
+    GGML_CALL void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
@@ -1532,18 +1530,16 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
             float                 attn_factor,
             float                 beta_fast,
-            float                 beta_slow,
-            float                 xpos_base,
-            bool                  xpos_down);
+            float                 beta_slow);
 
     // clamp
     // in-place, returns view(a)
@@ -1674,12 +1670,24 @@ extern "C" {
             float                 p1);
 
     // nearest interpolate
+    // multiplies ne0 and ne1 by scale factor
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // nearest interpolate
+    // nearest interpolate to specified dimensions
+    // used in tortoise.cpp
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3);
+
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
             struct ggml_context * ctx,
@@ -1721,13 +1729,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   k);
 
-    GGML_API struct ggml_tensor * ggml_flash_attn(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * q,
-            struct ggml_tensor  * k,
-            struct ggml_tensor  * v,
-            bool                  masked);
-
 #define GGML_KQ_MASK_PAD 32
 
     // q:    [n_embd, n_batch,     n_head,    1]
@@ -1748,6 +1749,7 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1756,14 +1758,6 @@ extern "C" {
            struct ggml_tensor  * d,
            bool                  masked);
 
-    GGML_API struct ggml_tensor * ggml_flash_ff(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b0,
-            struct ggml_tensor  * b1,
-            struct ggml_tensor  * c0,
-            struct ggml_tensor  * c1);
-
     GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
             struct ggml_tensor  * s,
@@ -2377,8 +2371,10 @@ extern "C" {
     GGML_API int ggml_cpu_has_avx512     (void);
     GGML_API int ggml_cpu_has_avx512_vbmi(void);
     GGML_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_API int ggml_cpu_has_avx512_bf16(void);
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
     GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);
@@ -2386,13 +2382,13 @@ extern "C" {
     GGML_API int ggml_cpu_has_wasm_simd  (void);
     GGML_API int ggml_cpu_has_blas       (void);
     GGML_API int ggml_cpu_has_cuda       (void);
-    GGML_API int ggml_cpu_has_clblast    (void);
     GGML_API int ggml_cpu_has_vulkan     (void);
     GGML_API int ggml_cpu_has_kompute    (void);
     GGML_API int ggml_cpu_has_gpublas    (void);
     GGML_API int ggml_cpu_has_sse3       (void);
     GGML_API int ggml_cpu_has_ssse3      (void);
     GGML_API int ggml_cpu_has_sycl       (void);
+    GGML_API int ggml_cpu_has_rpc        (void);
     GGML_API int ggml_cpu_has_vsx        (void);
     GGML_API int ggml_cpu_has_matmul_int8(void);
 

ggml/src/ggml-alloc.c

@@ -339,6 +339,7 @@ struct hash_node {
 };
 
 struct tensor_alloc {
+    int buffer_id;
     size_t offset;
     size_t size_max; // 0 = pre-allocated, unused, or view
 };
@@ -349,7 +350,6 @@ struct leaf_alloc {
 };
 
 struct node_alloc {
-    int buffer_id;
     struct tensor_alloc dst;
     struct tensor_alloc src[GGML_MAX_SRC];
 };
@@ -377,7 +377,7 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
     galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
     GGML_ASSERT(galloc->bufts != NULL);
 
-    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
     GGML_ASSERT(galloc->buffers != NULL);
 
     galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
@@ -386,8 +386,19 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
     for (int i = 0; i < n_bufs; i++) {
         galloc->bufts[i] = bufts[i];
         galloc->buffers[i] = NULL;
-        size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
-        galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+
+        // check if the same buffer type is used multiple times and reuse the same allocator
+        for (int j = 0; j < i; j++) {
+            if (bufts[i] == bufts[j]) {
+                galloc->buf_tallocs[i] = galloc->buf_tallocs[j];
+                break;
+            }
+        }
+
+        if (galloc->buf_tallocs[i] == NULL) {
+            size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
+            galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+        }
     }
     galloc->n_buffers = n_bufs;
 
@@ -405,10 +416,30 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
 
     for (int i = 0; i < galloc->n_buffers; i++) {
         if (galloc->buffers != NULL) {
-            ggml_backend_buffer_free(galloc->buffers[i]);
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buffers[j] == galloc->buffers[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_backend_buffer_free(galloc->buffers[i]);
+            }
         }
         if (galloc->buf_tallocs != NULL) {
-            ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+            }
         }
     }
 
@@ -511,17 +542,18 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
     }
 }
 
-static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
     // graph outputs are never freed
     if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
         AT_PRINTF("not freeing output %s\n", node->name);
         return;
     }
 
-    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
-    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
     size_t offset = hn->offset;
+    int buffer_id = hn->buffer_id;
+    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
     size_t size = ggml_backend_buft_get_alloc_size(buft, node);
     ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
     hn->allocated = false;
@@ -626,11 +658,11 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
                     AT_PRINTF("view_src %s: %d children, %d views\n",
                         view_src->name, view_src_hn->n_children, view_src_hn->n_views);
                     if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
-                        ggml_gallocr_free_node(galloc, view_src, buffer_id);
+                        ggml_gallocr_free_node(galloc, view_src);
                     }
                 }
                 else if (p_hn->allocated) {
-                    ggml_gallocr_free_node(galloc, parent, buffer_id);
+                    ggml_gallocr_free_node(galloc, parent);
                 }
             }
             AT_PRINTF("\n");
@@ -674,22 +706,25 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
-        node_alloc->buffer_id = get_node_buffer_id(node_buffer_ids, i);
         if (node->view_src || node->data) {
+            node_alloc->dst.buffer_id = -1;
             node_alloc->dst.offset = SIZE_MAX;
             node_alloc->dst.size_max = 0;
         } else {
             struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-            node_alloc->dst.offset   = hn->offset;
-            node_alloc->dst.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+            node_alloc->dst.buffer_id = hn->buffer_id;
+            node_alloc->dst.offset    = hn->offset;
+            node_alloc->dst.size_max  = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
         }
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (!src || src->view_src || src->data) {
+                node_alloc->src[j].buffer_id = -1;
                 node_alloc->src[j].offset = SIZE_MAX;
                 node_alloc->src[j].size_max = 0;
             } else {
                 struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
+                node_alloc->src[j].buffer_id = hn->buffer_id;
                 node_alloc->src[j].offset   = hn->offset;
                 node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
             }
@@ -706,9 +741,11 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
         galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
         if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.buffer_id = -1;
             galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
             galloc->leaf_allocs[i].leaf.size_max = 0;
         } else {
+            galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
             galloc->leaf_allocs[i].leaf.offset = hn->offset;
             galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
         }
@@ -716,6 +753,14 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 
     // reallocate buffers if needed
     for (int i = 0; i < galloc->n_buffers; i++) {
+        // if the buffer type is used multiple times, we reuse the same buffer
+        for (int j = 0; j < i; j++) {
+            if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                galloc->buffers[i] = galloc->buffers[j];
+                break;
+            }
+        }
+
         size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
         size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
 
@@ -724,6 +769,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 #ifndef NDEBUG
             fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+
             ggml_backend_buffer_free(galloc->buffers[i]);
             galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
             if (galloc->buffers[i] == NULL) {
@@ -740,7 +786,8 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
+    int buffer_id = tensor_alloc->buffer_id;
     assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
 
     if (tensor->view_src != NULL) {
@@ -750,7 +797,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
+            ggml_backend_view_init(tensor);
         }
     } else {
         if (tensor->data == NULL) {
@@ -768,8 +815,8 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
     }
 }
 
-static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * nalloc, struct tensor_alloc * talloc) {
-    ggml_backend_buffer_type_t buft = galloc->bufts[nalloc->buffer_id];
+static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
+    ggml_backend_buffer_type_t buft = talloc->buffer_id != -1 ? galloc->bufts[talloc->buffer_id] : NULL;
     size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(buft, node);
     return talloc->size_max >= node_size;
 }
@@ -793,7 +840,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
 
-        if (!ggml_gallocr_node_needs_realloc(galloc, node, node_alloc, &node_alloc->dst)) {
+        if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
 #ifndef NDEBUG
             fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
 #endif
@@ -805,7 +852,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
             if (src == NULL) {
                 continue;
             }
-            if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
+            if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
 #ifndef NDEBUG
                 fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
 #endif
@@ -846,7 +893,7 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
+        ggml_gallocr_init_tensor(galloc, leaf, &leaf_alloc->leaf);
     }
     // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
@@ -857,9 +904,9 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
             if (src == NULL) {
                 continue;
             }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
+            ggml_gallocr_init_tensor(galloc, src, &node_alloc->src[j]);
         }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
+        ggml_gallocr_init_tensor(galloc, node, &node_alloc->dst);
     }
 
     return true;
@@ -871,6 +918,15 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
     if (galloc->buffers[buffer_id] == NULL) {
         return 0;
     }
+
+    for (int i = 0; i < buffer_id; i++) {
+        if (galloc->buffers[i] == galloc->buffers[buffer_id]) {
+            // this buffer is the same as a previous one due to the same buffer type being used multiple times
+            // only return the buffer size the first time it appears to avoid double counting
+            return 0;
+        }
+    }
+
     return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
 }
 
@@ -886,7 +942,7 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
         fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
         for (size_t i = 0; i < *n_buffers; i++) {
-            ggml_backend_buffer_free(*buffers[i]);
+            ggml_backend_buffer_free((*buffers)[i]);
         }
         free(*buffers);
         return false;
@@ -899,12 +955,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(&tallocr, t);
             } else if (t->buffer == NULL) {
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         } else {
             if (t->view_src != NULL && t->buffer == NULL) {
                 // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         }
     }

ggml/src/ggml-backend-impl.h

@@ -17,13 +17,15 @@ extern "C" {
 
     struct ggml_backend_buffer_type_i {
         const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
+        // allocate a buffer of this type
         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
+        // tensor alignment
+        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft);
+        // max buffer size that can be allocated
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft);
+        // data size needed to allocate the tensor, including padding
+        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
         // 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);
     };
 
@@ -92,27 +94,37 @@ extern "C" {
         void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan (not used currently)
+        // create a new plan for a graph
         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);
+        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
+        void                      (*GGML_CALL graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
+        // compute the graph with the plan
+        enum ggml_status          (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-        // compute graph with a plan
-        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
+        // check if the backend can compute an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
 
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*GGML_CALL supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+
         // 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
+        // create a new event that can record events on this backend instance
         ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
         void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        // record an event on the backend instance that created it
         void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        // wait for an event on on a different backend instance
         void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        // block until an event is recorded
         void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
     };
 

ggml/src/ggml-backend.c

@@ -44,10 +44,6 @@ GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buf
     return ggml_nbytes(tensor);
 }
 
-bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return buft->iface.supports_backend(buft, backend);
-}
-
 bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
     if (buft->iface.is_host) {
         return buft->iface.is_host(buft);
@@ -151,7 +147,7 @@ void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
 bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
     ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
     if (dst_buf->iface.cpy_tensor) {
-        return src->buffer->iface.cpy_tensor(dst_buf, src, dst);
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
     }
     return false;
 }
@@ -286,6 +282,10 @@ bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor *
     return backend->iface.supports_op(backend, op);
 }
 
+bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return backend->iface.supports_buft(backend, buft);
+}
+
 bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     if (backend->iface.offload_op != NULL) {
         return backend->iface.offload_op(backend, op);
@@ -639,12 +639,6 @@ GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_cpu(backend);
-
-    GGML_UNUSED(buft);
-}
-
 GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return true;
 
@@ -659,7 +653,6 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
             /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
         /* .context = */ NULL,
@@ -715,7 +708,6 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
             /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
         /* .context  = */ NULL,
@@ -836,6 +828,12 @@ GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const
     GGML_UNUSED(backend);
 }
 
+GGML_CALL static bool ggml_backend_cpu_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
 static struct ggml_backend_i cpu_backend_i = {
     /* .get_name                = */ ggml_backend_cpu_name,
     /* .free                    = */ ggml_backend_cpu_free,
@@ -846,9 +844,11 @@ static struct ggml_backend_i cpu_backend_i = {
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
     /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
     /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
     /* .supports_op             = */ ggml_backend_cpu_supports_op,
+    /* .supports_buft           = */ ggml_backend_cpu_supports_buft,
     /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
@@ -1055,6 +1055,9 @@ struct ggml_backend_sched {
     int * node_backend_ids; // [graph_size]
     int * leaf_backend_ids; // [graph_size]
 
+    int * prev_node_backend_ids; // [graph_size]
+    int * prev_leaf_backend_ids; // [graph_size]
+
     // copy of the graph with modified inputs
     struct ggml_cgraph * graph;
 
@@ -1075,6 +1078,8 @@ struct ggml_backend_sched {
     ggml_backend_sched_eval_callback callback_eval;
     void * callback_eval_user_data;
 
+    bool debug;
+
     // align context_buffer to GGML_MEM_ALIGN
 #ifdef _MSC_VER
     __declspec(align(GGML_MEM_ALIGN))
@@ -1097,22 +1102,24 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
     return -1;
 }
 
-static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor) {
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
     ggml_backend_buffer_t buffer = tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
 
-    // find highest prio backend that supports the buffer type
+    // find highest prio backend that supports the buffer type and the op
     for (int i = 0; i < sched->n_backends; i++) {
-        if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) {
+        if (ggml_backend_supports_buft(sched->backends[i], buffer->buft) &&
+            ggml_backend_supports_op(sched->backends[i], op)) {
             return i;
         }
     }
 
-    fprintf(stderr, "%s: error: no backend supports buffer type %s used in tensor %s\n",
-        __func__, ggml_backend_buffer_name(buffer), tensor->name);
-    GGML_ASSERT(false);
+#ifndef NDEBUG
+    fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
+        __func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
+#endif
 
     return -1;
 }
@@ -1131,7 +1138,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     // TODO: use supports_op to check if the backend supports the op
 
     // assign pre-allocated nodes to their backend
-    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor);
+    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
         SET_CAUSE(tensor, "1.dst");
         return cur_backend_id;
@@ -1139,7 +1146,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     // view_src
     if (tensor->view_src != NULL) {
-        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
+        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src, tensor);
         if (cur_backend_id != -1) {
             SET_CAUSE(tensor, "1.vsrc");
             return cur_backend_id;
@@ -1161,11 +1168,11 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
             continue;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src);
+            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
             if (src_backend_id == sched->n_backends - 1) {
                 for (int b = 0; b < src_backend_id; b++) {
-                    if (ggml_backend_offload_op(sched->backends[b], tensor)) {
+                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
                         return b;
                     }
@@ -1223,10 +1230,33 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
     }
 }
 
-//#define DEBUG_PASS1
-//#define DEBUG_PASS2
-//#define DEBUG_PASS3
-//#define DEBUG_PASS4
+static bool ggml_backend_sched_buffer_supported(ggml_backend_sched_t sched, struct ggml_tensor * t, int backend_id) {
+    ggml_backend_buffer_t buf = t->view_src ? t->view_src->buffer : t->buffer;
+    ggml_backend_buffer_type_t buft = NULL;
+
+    if (buf) {
+        // the tensor is already allocated
+        buft = buf->buft;
+    } else {
+        // see if the tensor already has a backend assigned, and use the buffer type of that backend
+        int tensor_backend_id = tensor_backend_id(t);
+        if (tensor_backend_id == -1 && t->view_src) {
+            tensor_backend_id = tensor_backend_id(t->view_src);
+        }
+        if (tensor_backend_id != -1) {
+            buft = sched->bufts[tensor_backend_id];
+        }
+    }
+
+    return buft != NULL && ggml_backend_supports_buft(sched->backends[backend_id], buft);
+}
+
+static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, struct ggml_tensor * node, int cur_backend_id, int * node_backend_id) {
+    if (ggml_backend_supports_op(sched->backends[cur_backend_id], node)) {
+        *node_backend_id = cur_backend_id;
+        SET_CAUSE(node, "2.sup");
+    }
+}
 
 // assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
 static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
@@ -1280,17 +1310,13 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             }
         }
     }
-#ifdef DEBUG_PASS1
-    fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
-#endif
 
     // pass 2: expand current backend assignments
     // assign the same backend to adjacent nodes
     // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
     // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
-
-
-    // pass 2.2 expand gpu down
+    // ops unsupported by the backend being expanded will be left unassigned so that they can be assigned later when the locations of its inputs are known
+    // expand gpu down
     {
         int cur_backend_id = -1;
         for (int i = 0; i < graph->n_nodes; i++) {
@@ -1306,13 +1332,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 } else {
                     cur_backend_id = *node_backend_id;
                 }
-            } else {
-                *node_backend_id = cur_backend_id;
-                SET_CAUSE(node, "2.2");
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
             }
         }
     }
-    // pass 2.1 expand gpu up
+    // expand gpu up
     {
         int cur_backend_id = -1;
         for (int i = graph->n_nodes - 1; i >= 0; i--) {
@@ -1328,13 +1353,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 } else {
                     cur_backend_id = *node_backend_id;
                 }
-            } else {
-                *node_backend_id = cur_backend_id;
-                SET_CAUSE(node, "2.1");
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
             }
         }
     }
-    // pass 2.4 expand rest down
+    // expand rest down
     {
         int cur_backend_id = -1;
         for (int i = 0; i < graph->n_nodes; i++) {
@@ -1345,13 +1369,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             int * node_backend_id = &tensor_backend_id(node);
             if (*node_backend_id != -1) {
                 cur_backend_id = *node_backend_id;
-            } else {
-                *node_backend_id = cur_backend_id;
-                SET_CAUSE(node, "2.4");
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
             }
         }
     }
-    // pass 2.3 expand rest up
+    // expand rest up
     {
         int cur_backend_id = -1;
         for (int i = graph->n_nodes - 1; i >= 0; i--) {
@@ -1362,24 +1385,80 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             int * node_backend_id = &tensor_backend_id(node);
             if (*node_backend_id != -1) {
                 cur_backend_id = *node_backend_id;
-            } else {
-                *node_backend_id = cur_backend_id;
-                SET_CAUSE(node, "2.3");
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
             }
         }
     }
 
-#ifdef DEBUG_PASS2
-    fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
-#endif
+    // pass 3: upgrade nodes to higher prio backends with compatible buffer types
+    // if the tensor is already in the same buffer type (*) as another higher priority backend, we should move it there
+    // however, we also need to verify that the sources are in compatible buffer types
+    // (*) the actual requirement is more relaxed, the buffer type of the backend should be supported by all the users of this tensor further down the graph
+    // however, this is slow to verify, so we have a more strict requirement that the buffer type is the same
+    // this is not uncommon since multiple backends can use host memory, with the same buffer type (eg. BLAS and CPU)
+    // additionally, set remaining unassigned nodes to the backend with the most supported inputs
+    // only nodes that could not be assigned during expansion due to the backend not supporting the op should be unassigned at this point
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        int * node_backend_id = &tensor_backend_id(node);
+        if (*node_backend_id == -1) {
+            // unassigned node: find the backend with the most supported inputs
+            int n_supported_best = -1;
+            for (int b = 0; b < sched->n_backends; b++) {
+                if (ggml_backend_supports_op(sched->backends[b], node)) {
+                    int n_supported = 0;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if ((tensor_backend_id(src) != -1 || tensor_backend_id(src->view_src) != -1) && ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            n_supported++;
+                        }
+                    }
+                    if (n_supported > n_supported_best) {
+                        n_supported_best = n_supported;
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.best");
+                    }
+                }
+            }
+        } else {
+            // assigned node: upgrade to higher prio backend if possible
+            for (int b = 0; b < *node_backend_id; b++) {
+                if (sched->bufts[b] == sched->bufts[*node_backend_id] && ggml_backend_supports_op(sched->backends[b], node)) {
+                    bool supported = true;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if (!ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            supported = false;
+                            break;
+                        }
+                    }
+                    if (supported) {
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.upg");
+                        break;
+                    }
+                }
+            }
+        }
+    }
 
-    // pass 3: assign backends to remaining src from dst and view_src
+    // pass 4: assign backends to remaining src from dst and view_src
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * cur_backend_id = &tensor_backend_id(node);
         if (node->view_src != NULL && *cur_backend_id == -1) {
             *cur_backend_id = tensor_backend_id(node->view_src);
-            SET_CAUSE(node, "3.vsrc");
+            SET_CAUSE(node, "4.vsrc");
         }
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
@@ -1391,17 +1470,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 if (src->view_src != NULL) {
                     // views are always on the same backend as the source
                     *src_backend_id = tensor_backend_id(src->view_src);
-                    SET_CAUSE(src, "3.vsrc");
+                    SET_CAUSE(src, "4.vsrc");
                 } else {
                     *src_backend_id = *cur_backend_id;
-                    SET_CAUSE(src, "3.cur");
+                    SET_CAUSE(src, "4.cur");
                 }
             }
         }
     }
-#ifdef DEBUG_PASS3
-    fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
-#endif
 
     // pass 4: split graph, find tensors that need to be copied
     {
@@ -1448,10 +1524,12 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                         }
                     }
                     // check if the split has too many inputs
+                    // FIXME: count the number of inputs instead of only checking when full
                     if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
                         const size_t id = hash_id(src);
                         int src_backend_id = sched->tensor_backend_id[id];
-                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL) {
+                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
+                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL && !supported) {
                             //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
                             need_new_split = true;
                             break;
@@ -1486,7 +1564,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 const int src_backend_id = tensor_backend_id(src);
                 assert(src_backend_id != -1); // all inputs should be assigned by now
 
-                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1)  {
+                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
                     size_t id = hash_id(src);
                     if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
                         ggml_backend_t backend = sched->backends[src_backend_id];
@@ -1511,7 +1589,8 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     }
                 }
 
-                if (src_backend_id != node_backend_id) {
+                bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
+                if (src_backend_id != cur_backend_id && !supported) {
                     // create a copy of the input in the split's backend
                     const size_t id = hash_id(src);
                     if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
@@ -1537,9 +1616,21 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         split->i_end = graph->n_nodes;
         sched->n_splits = i_split + 1;
     }
-#ifdef DEBUG_PASS4
-    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
-#endif
+
+    if (sched->debug) {
+        ggml_backend_sched_print_assignments(sched, graph);
+    }
+
+    // swap node_backend_ids and leaf_backend_ids and prevs
+    {
+        int * tmp = sched->node_backend_ids;
+        sched->node_backend_ids = sched->prev_node_backend_ids;
+        sched->prev_node_backend_ids = tmp;
+
+        tmp = sched->leaf_backend_ids;
+        sched->leaf_backend_ids = sched->prev_leaf_backend_ids;
+        sched->prev_leaf_backend_ids = tmp;
+    }
 
     // create copies of the graph for each split
     // TODO: avoid this copy
@@ -1613,8 +1704,26 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
+    bool backend_ids_changed = false;
+    for (int i = 0; i < sched->graph->n_nodes; i++) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
+            backend_ids_changed = true;
+            break;
+        }
+    }
+    if (!backend_ids_changed) {
+        for (int i = 0; i < sched->graph->n_leafs; i++) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
+                backend_ids_changed = true;
+                break;
+            }
+        }
+    }
+
     // allocate graph
-    if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
         ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
@@ -1727,6 +1836,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
 
     struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
+    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+
     // initialize hash table
     sched->hash_set          = ggml_hash_set_new(graph_size);
     sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
@@ -1735,6 +1846,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
     sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
     sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
+    sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
     sched->n_backends = n_backends;
 
@@ -1747,7 +1860,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
         sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
-        GGML_ASSERT(ggml_backend_buft_supports_backend(sched->bufts[b], backends[b]));
+        GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
         if (sched->n_copies > 1) {
             for (int c = 0; c < sched->n_copies; c++) {
                 sched->events[b][c] = ggml_backend_event_new(backends[b]);
@@ -1779,6 +1892,8 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     free(sched->tensor_copies);
     free(sched->node_backend_ids);
     free(sched->leaf_backend_ids);
+    free(sched->prev_node_backend_ids);
+    free(sched->prev_leaf_backend_ids);
     free(sched);
 }
 
@@ -1864,6 +1979,15 @@ int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
     return sched->n_copies;
 }
 
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
@@ -1875,6 +1999,7 @@ void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct gg
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     tensor_backend_id(node) = backend_index;
+    SET_CAUSE(node, "usr");
 }
 
 ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
@@ -1887,16 +2012,15 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
 
 // utils
 
-void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);
 
-    tensor->buffer = buffer;
+    tensor->buffer = tensor->view_src->buffer;
     tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    tensor->backend = tensor->view_src->backend;
-    ggml_backend_buffer_init_tensor(buffer, tensor);
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
 }
 
 void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
@@ -1955,7 +2079,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
     struct ggml_tensor * dst = node_copies[id];
     if (dst->view_src != NULL) {
         graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
-        ggml_backend_view_init(dst->view_src->buffer, dst);
+        ggml_backend_view_init(dst);
     }
     else {
         ggml_backend_tensor_copy(src, dst);

ggml/src/ggml-blas.cpp

@@ -0,0 +1,363 @@
+#include "ggml-blas.h"
+#include "ggml-backend-impl.h"
+
+#include <future>
+#include <vector>
+
+#if defined(GGML_USE_ACCELERATE)
+#   include <Accelerate/Accelerate.h>
+#elif defined(GGML_BLAS_USE_MKL)
+#   include <mkl.h>
+#else
+#   include <cblas.h>
+#   ifdef BLIS_ENABLE_CBLAS
+#       include <blis.h>
+#   endif
+#endif
+
+struct ggml_backend_blas_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+#ifndef GGML_USE_OPENMP
+    std::vector<std::future<void>> tasks;
+#endif
+};
+
+// helper function to determine if it is better to use BLAS or not
+// for large matrices, BLAS is faster
+static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1->type == GGML_TYPE_F32 &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+
+        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
+        return true;
+    }
+
+    return false;
+}
+
+static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const int64_t ne_plane      = ne01*ne00;
+    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
+
+    if (ctx->work_size < desired_wsize) {
+        ctx->work_data.reset(new char[desired_wsize]);
+        ctx->work_size = desired_wsize;
+    }
+    void * wdata = ctx->work_data.get();
+
+    // convert src0 to float
+    if (type != GGML_TYPE_F32) {
+        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits.to_float;
+
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
+                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
+
+                const int min_cols_per_thread = 4096;
+                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
+                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
+
+#ifdef GGML_USE_OPENMP
+                #pragma omp parallel for num_threads(n_threads)
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                }
+#else
+                for (int i = 1; i < n_threads; i++) {
+                    const int64_t start =       i*ne01/n_threads;
+                    const int64_t end   = (i + 1)*ne01/n_threads;
+                    if (start < end) {
+                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
+                            for (int64_t i01 = start; i01 < end; i01++) {
+                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                            }
+                        }));
+                    }
+                }
+                {
+                    // reuse the current thread for the first task
+                    const int64_t start = 0;
+                    const int64_t end   = ne01/n_threads;
+                    for (int64_t i01 = start; i01 < end; i01++) {
+                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                    }
+                }
+#endif
+            }
+        }
+
+#ifndef GGML_USE_OPENMP
+        // wait for all tasks to finish
+        for (auto & task : ctx->tasks) {
+            task.get();
+        }
+        ctx->tasks.clear();
+#endif
+    }
+
+#if defined(OPENBLAS_VERSION)
+    openblas_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(BLIS_ENABLE_CBLAS)
+    bli_thread_set_num_threads(ctx->n_threads);
+#endif
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            const int64_t i03 = i13/r3;
+            const int64_t i02 = i12/r2;
+
+            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
+            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
+
+            if (type != GGML_TYPE_F32) {
+                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
+            }
+
+            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+                        ne1, ne01, ne10,
+                        1.0f,   y, ne10,
+                                x, ne00,
+                        0.0f,   d, ne01);
+        }
+    }
+}
+
+static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+    // src0: (k,n)
+    // src1: (k,m)
+    // dst:  (m,n)
+    //
+    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+    // Also expressed as (major,minor)
+    // a: (m,k): so src1 transposed
+    // b: (k,n): so src0
+    // c: (m,n)
+    //
+    // However, if ggml_is_transposed(src1) is true, then
+    // src1->data already contains a transposed version, so sgemm mustn't
+    // transpose it further.
+
+    int n = src0->ne[0];
+    int k = src0->ne[1];
+    int m = src1->ne[0];
+
+    CBLAS_TRANSPOSE transposeA;
+    int lda;
+
+    if (!ggml_is_transposed(src1)) {
+        transposeA = CblasTrans;
+        lda = m;
+    } else {
+        transposeA = CblasNoTrans;
+        lda = k;
+    }
+
+    float * a = (float *) ((char *) src1->data);
+    float * b = (float *) ((char *) src0->data);
+    float * c = (float *) ((char *) dst->data);
+
+    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+    GGML_UNUSED(ctx);
+}
+
+// backend interface
+
+GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+    return "BLAS";
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_backend_blas_mul_mat(ctx, node);
+                break;
+
+            case GGML_OP_OUT_PROD:
+                ggml_backend_blas_out_prod(ctx, node);
+                break;
+
+            case GGML_OP_NONE:
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_PERMUTE:
+            case GGML_OP_TRANSPOSE:
+                break;
+
+            default:
+                fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+                GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
+           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
+                                          op->src[1]->type == GGML_TYPE_F32 &&
+                                          ggml_is_matrix(src0) &&
+                                          ggml_is_matrix(src1) &&
+                                          ggml_is_contiguous(src0) &&
+                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i blas_backend_i = {
+    /* .get_name                = */ ggml_backend_blas_name,
+    /* .free                    = */ ggml_backend_blas_free,
+    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .supports_op             = */ ggml_backend_blas_supports_op,
+    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_blas_guid(void) {
+    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_blas_init(void) {
+    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_blas_guid(),
+        /* .interface = */ blas_backend_i,
+        /* .context   = */ ctx,
+    };
+
+#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
+        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+    }
+#endif
+
+#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#endif
+
+    return backend;
+}
+
+GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
+}
+
+void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
+
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
+    ctx->n_threads = n_threads;
+}

ggml/src/ggml-common.h

@@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
 // QK = number of values after dequantization
 // QK_K = super-block size
 
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
 #define QK_K 256
 #define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
 
 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
 // QR = QK / number of values before dequantization
@@ -128,16 +123,17 @@ typedef sycl::half2 ggml_half2;
 #define QI1_S (QK_K / (4*QR1_S))
 #define QR1_S 8
 
+#define QI1_M (QK_K / (4*QR1_M))
+#define QR1_M 8
+
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2
 
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
-#endif
+
+#define QI3_S (QK_K / (4*QR3_S))
+#define QR3_S 8
 
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
@@ -228,15 +224,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
 // 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_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + 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
@@ -244,20 +231,11 @@ typedef struct {
     ggml_half d;           // super-block scale
 } block_q3_K;
 static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + 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_half 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_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -270,21 +248,11 @@ typedef struct {
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
 static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + 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_half 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_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -298,7 +266,6 @@ typedef struct {
     uint8_t qs[QK_K/2];           // quants, low 4 bits
 } block_q5_K;
 static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + 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
@@ -356,11 +323,7 @@ typedef struct {
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
 // 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
 #define IQ3S_N_SCALE QK_K/64
-#endif
 typedef struct {
     ggml_half d;
     uint8_t qs[QK_K/4];
@@ -381,16 +344,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
 typedef struct {
     uint8_t  qs[QK_K/8];      // grid index, low 8 bits
     uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
     uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
 } block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
 static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
 
 // Used by IQ1_M quants
 typedef union {
@@ -406,9 +362,6 @@ typedef struct {
 } block_iq4_nl;
 static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
 
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
 typedef struct {
     ggml_half d;
     uint16_t scales_h;
@@ -416,7 +369,6 @@ typedef struct {
     uint8_t  qs[QK_K/2];
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
 
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL

ggml/src/ggml-cuda.cu

@@ -43,19 +43,59 @@
 #include <mutex>
 #include <stdint.h>
 #include <stdio.h>
+#include <stdarg.h>
+#include <stdlib.h>
 #include <string>
 #include <vector>
 
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
+static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    GGML_UNUSED(level);
+    GGML_UNUSED(user_data);
+    fprintf(stderr, "%s", msg);
+}
+
+ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback;
+void * ggml_cuda_log_user_data = NULL;
+
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_cuda_log_callback = log_callback;
+    ggml_cuda_log_user_data = user_data;
+}
+
+#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+GGML_ATTRIBUTE_FORMAT(2, 3)
+static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) {
+    if (ggml_cuda_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data);
+        } else {
+            std::vector<char> buffer2(len + 1);  // vsnprintf adds a null terminator
+            va_end(args);
+            va_start(args, format);
+            vsnprintf(&buffer2[0], buffer2.size(), format, args);
+            ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data);
+        }
+        va_end(args);
+    }
+}
+
 [[noreturn]]
 void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
     int id = -1; // in case cudaGetDevice fails
     cudaGetDevice(&id);
 
-    fprintf(stderr, "CUDA error: %s\n", msg);
-    fprintf(stderr, "  current device: %d, in function %s at %s:%d\n", id, func, file, line);
-    fprintf(stderr, "  %s\n", stmt);
+    GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
+    GGML_CUDA_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
+    GGML_CUDA_LOG_ERROR("  %s\n", stmt);
     // abort with GGML_ASSERT to get a stack trace
     GGML_ASSERT(!"CUDA error");
 }
@@ -79,6 +119,20 @@ int ggml_cuda_get_device() {
     return id;
 }
 
+static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
+    ggml_cuda_set_device(device);
+#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
+    auto res = hipMallocManaged(ptr, size);
+    if (res == hipSuccess) {
+        // if error we "need" to know why...
+        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+    }
+    return res;
+#else
+    return cudaMalloc(ptr, size);
+#endif
+}
+
 static ggml_cuda_device_info ggml_cuda_init() {
 #ifdef __HIP_PLATFORM_AMD__
     // Workaround for a rocBLAS bug when using multiple graphics cards:
@@ -91,24 +145,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
     cudaError_t err = cudaGetDeviceCount(&info.device_count);
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
+        GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
         return info;
     }
 
     GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
 
     int64_t total_vram = 0;
-#if defined(GGML_CUDA_FORCE_MMQ)
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
+#ifdef GGML_CUDA_FORCE_MMQ
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    yes\n", __func__);
 #else
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
-#endif
-#if defined(CUDA_USE_TENSOR_CORES)
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    no\n", __func__);
+#endif // GGML_CUDA_FORCE_MMQ
+#ifdef GGML_CUDA_FORCE_CUBLAS
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__);
 #else
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
-#endif
-    fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
+#endif // GGML_CUDA_FORCE_CUBLAS
+    GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
@@ -129,18 +183,20 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         cudaDeviceProp prop;
         CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-        fprintf(stderr, "  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_CUDA_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
 
+        info.devices[id].nsm   = prop.multiProcessorCount;
+        info.devices[id].smpb  = prop.sharedMemPerBlock;
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+        info.devices[id].smpbo = prop.sharedMemPerBlock;
         info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
 #else
+        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-        info.devices[id].smpb = prop.sharedMemPerBlock;
-        info.devices[id].nsm  = prop.multiProcessorCount;
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -231,12 +287,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         size_t look_ahead_size = (size_t) (1.05 * size);
         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
         ggml_cuda_set_device(device);
-        CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
+        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
-        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
-                (uint32_t)(max_size/1024/1024), (uint32_t)(pool_size/1024/1024), (uint32_t)(size/1024/1024));
+        GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
+                           (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
 #endif
         return ptr;
     }
@@ -250,7 +306,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
                 return;
             }
         }
-        fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
+        GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
         ggml_cuda_set_device(device);
         CUDA_CHECK(cudaFree(ptr));
         pool_size -= size;
@@ -489,6 +545,10 @@ GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_bu
     return ctx->name.c_str();
 }
 
+static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_buffer_type_name;
+}
+
 GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
 
@@ -497,9 +557,11 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
     size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
 
     void * dev_ptr;
-    cudaError_t err = cudaMalloc(&dev_ptr, size);
+    cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err));
+        // clear the error
+        cudaGetLastError();
+        GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
         return nullptr;
     }
 
@@ -529,24 +591,12 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backen
     GGML_UNUSED(buft);
 }
 
-GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    if (!ggml_backend_is_cuda(backend)) {
-        return false;
-    }
-
-    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
-
-    return buft_ctx->device == cuda_ctx->device;
-}
-
 static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_cuda_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
 };
 
@@ -577,88 +627,22 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
 
 // cuda split buffer
 
-static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
-    int64_t min_compute_capability = INT_MAX;
-    int64_t max_compute_capability = INT_MIN;
+static int64_t get_row_rounding(const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
+    int64_t row_rounding = 0;
     for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
-        if (tensor_split[id] < (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
-            if (min_compute_capability > ggml_cuda_info().devices[id].cc) {
-                min_compute_capability = ggml_cuda_info().devices[id].cc;
-            }
-            if (max_compute_capability < ggml_cuda_info().devices[id].cc) {
-                max_compute_capability = ggml_cuda_info().devices[id].cc;
-            }
+        if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
+            continue;
         }
-    }
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 32;
-        case GGML_TYPE_Q3_K:
-            return min_compute_capability < CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-        case GGML_TYPE_IQ2_XXS:
-        case GGML_TYPE_IQ2_XS:
-        case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ1_M:
-        case GGML_TYPE_IQ4_NL:
-        case GGML_TYPE_IQ4_XS:
-        case GGML_TYPE_IQ3_S:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        default:
-            GGML_ASSERT(false);
+        const int cc = ggml_cuda_info().devices[id].cc;
+        row_rounding = std::max(row_rounding, (int64_t)get_mmq_y_host(cc));
     }
-#else
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_IQ2_XXS:
-        case GGML_TYPE_IQ2_XS:
-        case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ1_M:
-        case GGML_TYPE_IQ4_NL:
-        case GGML_TYPE_IQ4_XS:
-        case GGML_TYPE_IQ3_S:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q6_K:
-            return 64;
-        default:
-            GGML_ASSERT(false);
-    }
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    return row_rounding;
 }
 
 static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
     const int64_t nrows = ggml_nrows(tensor);
-    const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
+    const int64_t rounding = get_row_rounding(tensor_split);
 
     *row_low = id == 0 ? 0 : nrows*tensor_split[id];
     *row_low -= *row_low % rounding;
@@ -756,7 +740,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
         // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
         ggml_cuda_set_device(id);
         char * buf;
-        CUDA_CHECK(cudaMalloc(&buf, size));
+        CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));
 
         // set padding to 0 to avoid possible NaN values
         if (size > original_size) {
@@ -873,6 +857,10 @@ GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_back
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_name;
+}
+
 GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
     // instead, we allocate them for each tensor separately in init_tensor
@@ -916,12 +904,6 @@ GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_
     return total_size;
 }
 
-GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_cuda(backend);
-
-    GGML_UNUSED(buft);
-}
-
 GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
     return false;
 
@@ -934,7 +916,6 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface
     /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
     /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
     /* .get_alloc_size   = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend,
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
@@ -1002,8 +983,8 @@ static void * ggml_cuda_host_malloc(size_t size) {
     if (err != cudaSuccess) {
         // clear the error
         cudaGetLastError();
-        fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
-            size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
         return nullptr;
     }
 
@@ -1034,7 +1015,6 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
             /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
         /* .context  = */ nullptr,
@@ -1357,10 +1337,30 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
     GGML_UNUSED(main_device);
 }
 
+static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
+    void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
+
+#if !defined(GGML_USE_HIPBLAS)
+    // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
+    cudaMemcpy3DPeerParms p = {};
+    p.dstDevice = dstDevice;
+    p.dstPtr = make_cudaPitchedPtr(dst, dpitch, dpitch, height);
+    p.srcDevice = srcDevice;
+    p.srcPtr = make_cudaPitchedPtr(src, spitch, spitch, height);
+    p.extent = make_cudaExtent(width, height, 1);
+    return cudaMemcpy3DPeerAsync(&p, stream);
+#else
+    // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
+    GGML_UNUSED(dstDevice);
+    GGML_UNUSED(srcDevice);
+    return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
+#endif // !defined(GGML_USE_HIPBLAS)
+}
+
 static void ggml_cuda_op_mul_mat(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
-    const bool convert_src1_to_q8_1) {
+    quantize_cuda_t quantize_src1) {
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -1417,7 +1417,9 @@ static void ggml_cuda_op_mul_mat(
     }
 
     struct dev_data {
-        ggml_cuda_pool_alloc<char>  src0_dd_alloc;
+        int cc;
+
+        ggml_cuda_pool_alloc<char>   src0_dd_alloc;
         ggml_cuda_pool_alloc<float> src1_ddf_alloc;
         ggml_cuda_pool_alloc<char>  src1_ddq_alloc;
         ggml_cuda_pool_alloc<float>   dst_dd_alloc;
@@ -1436,6 +1438,8 @@ static void ggml_cuda_op_mul_mat(
     int used_devices = 0;
 
     for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        dev[id].cc = ggml_cuda_info().devices[id].cc;
+
         // by default, use all rows
         dev[id].row_low  = 0;
         dev[id].row_high = ne01;
@@ -1443,7 +1447,7 @@ static void ggml_cuda_op_mul_mat(
         // for multi GPU, get the row boundaries from tensor split
         // and round to mul_mat_q tile sizes
         if (split) {
-            const int64_t rounding = get_row_rounding(src0->type, tensor_split);
+            const int64_t rounding = get_row_rounding(tensor_split);
 
             if (id != 0) {
                 dev[id].row_low  = ne01*tensor_split[id];
@@ -1486,11 +1490,15 @@ static void ggml_cuda_op_mul_mat(
             dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1));
         }
 
-        if (convert_src1_to_q8_1) {
-            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
+        if (quantize_src1) {
+            size_t src_1_ddq_size = nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs;
+            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                src_1_ddq_size += get_mmq_x_max_host(dev[id].cc)*sizeof(block_q8_1_mmq);
+            }
+            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);
 
             if (src1_on_device && src1_is_contiguous) {
-                quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
                 CUDA_CHECK(cudaGetLastError());
             }
         }
@@ -1536,7 +1544,12 @@ static void ggml_cuda_op_mul_mat(
                 const int64_t i03 = i0 / ne12;
                 const int64_t i02 = i0 % ne12;
 
-                const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                size_t src1_ddq_i_offset = i0*ne11 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                    src1_ddq_i_offset += src1_col_0 * sizeof(block_q8_1_mmq);
+                } else {
+                    src1_ddq_i_offset += src1_col_0 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                }
 
                 // for split tensors the data begins at i0 == i0_offset_low
                 char  *  src0_dd_i =  dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
@@ -1553,10 +1566,17 @@ static void ggml_cuda_op_mul_mat(
                 // copy src0, src1 to device if necessary
                 if (src1_is_contiguous) {
                     if (id != ctx.device) {
-                        if (convert_src1_to_q8_1) {
+                        if (quantize_src1) {
                             char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
-                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, ctx.device,
-                                                            src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
+                            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                                const size_t pitch = ne11*sizeof(block_q8_1_mmq);
+                                const size_t width = src1_ncols*sizeof(block_q8_1_mmq);
+                                const size_t height = src1_padded_col_size/(4*QK8_1);
+                                CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(src1_ddq_i, id, pitch, src1_ddq_i_source, ctx.device, pitch, width, height, stream));
+                            } else {
+                                CUDA_CHECK(cudaMemcpyPeerAsync(
+                                    src1_ddq_i, id, src1_ddq_i_source, ctx.device, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
+                            }
                         } else {
                             float * src1_ddf_i_source = (float *) src1->data;
                             src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
@@ -1571,8 +1591,8 @@ static void ggml_cuda_op_mul_mat(
                     GGML_ASSERT(false);
                 }
 
-                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
-                    quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                if (quantize_src1 && !src1_is_contiguous) {
+                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
 
@@ -1597,22 +1617,8 @@ static void ggml_cuda_op_mul_mat(
                         float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                         GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
                         dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
-#if !defined(GGML_USE_HIPBLAS)
-                        // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
-                        cudaMemcpy3DPeerParms p = {};
-                        p.dstDevice = ctx.device;
-                        p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols);
-                        p.srcDevice = id;
-                        p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols);
-                        p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1);
-                        CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream));
-#else
-                        // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
-                        CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float),
-                                                        dst_dd_i, row_diff*sizeof(float),
-                                                        row_diff*sizeof(float), src1_ncols,
-                                                        cudaMemcpyDeviceToDevice, stream));
-#endif
+                        CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(
+                            dhf_dst_i, ctx.device, ne0*sizeof(float), dst_dd_i, id, row_diff*sizeof(float), row_diff*sizeof(float), src1_ncols, stream));
                     } else {
                         float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
                         GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
@@ -1814,7 +1820,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         }
     }
 #else
-    if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
+    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
@@ -1867,9 +1873,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
 
-    int64_t min_compute_capability = INT_MAX;
+    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
+    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    bool any_gpus_with_slow_fp16 = false;
 
-    bool any_pascal_with_slow_fp16 = false;
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
         auto & tensor_split = buft_ctx->tensor_split;
@@ -1879,55 +1893,18 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            if (min_compute_capability > ggml_cuda_info().devices[id].cc) {
-                min_compute_capability = ggml_cuda_info().devices[id].cc;
-            }
-            if (ggml_cuda_info().devices[id].cc == 610) {
-                any_pascal_with_slow_fp16 = true;
-            }
+            const int cc            = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_vec_q       = use_mul_mat_vec_q       && cc >= MIN_CC_DP4A;
+            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
         }
     } else {
-        min_compute_capability    = ggml_cuda_info().devices[ctx.device].cc;
-        any_pascal_with_slow_fp16 = ggml_cuda_info().devices[ctx.device].cc == 610;
+        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_vec_q       = use_mul_mat_vec_q       && cc >= MIN_CC_DP4A;
+        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
     }
 
-    // check data types and tensor shapes for custom matrix multiplication kernels:
-    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
-
-    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-
-    bool              use_mul_mat_q =  ggml_cuda_supports_mmq(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
-    const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
-
-#ifdef CUDA_USE_TENSOR_CORES
-    use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
-#endif // CUDA_USE_TENSOR_CORES
-
-#else
-
-    // fp16 performance is good on Volta or newer and on P100 (compute capability 6.0)
-    const bool fp16_performance_good = min_compute_capability >= CC_PASCAL && !any_pascal_with_slow_fp16;
-
-    // mmvq and mmq need the __dp4a instruction which on NVIDIA is only available for CC >= 6.1
-    use_mul_mat_vec_q = use_mul_mat_vec_q && min_compute_capability >= MIN_CC_DP4A;
-    use_mul_mat_q     = use_mul_mat_q     && min_compute_capability >= MIN_CC_DP4A;
-
-#ifdef CUDA_USE_TENSOR_CORES
-    // when tensor cores are available, use them for large batch size
-    // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-    use_mul_mat_q     = use_mul_mat_q     && (!fp16_performance_good || src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
-#endif // CUDA_USE_TENSOR_CORES
-
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
     // if mmvq is available it's a better choice than dmmv:
 #ifndef GGML_CUDA_FORCE_DMMV
     use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
@@ -1941,23 +1918,24 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // KQ single-batch
+    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
         ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
-    } else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // KQV single-batch
+    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
         ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || fp16_performance_good) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
+               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        // KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, quantize_mmq_q8_1_cuda);
     } else {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, nullptr);
     }
 }
 
@@ -2246,7 +2224,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             break;
         case GGML_OP_MUL_MAT:
             if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
-                fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
+                GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
                 return false;
             } else {
                 ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
@@ -2261,6 +2239,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SQR:
             ggml_cuda_op_sqr(ctx, dst);
             break;
+        case GGML_OP_SQRT:
+            ggml_cuda_op_sqrt(ctx, dst);
+            break;
         case GGML_OP_CLAMP:
             ggml_cuda_op_clamp(ctx, dst);
             break;
@@ -2300,7 +2281,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
 
     cudaError_t err = cudaGetLastError();
     if (err != cudaSuccess) {
-        fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst));
+        GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
         CUDA_CHECK(err);
     }
 
@@ -2468,15 +2449,15 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 
     bool use_cuda_graph = true;
     bool cuda_graph_update_required = false;
-    // pointer to CUDA cpy kernel, which is required to identify
+    // vector of pointers to CUDA cpy kernels, which are required to identify
     // kernel parameters which need updated in the graph for each token
-    void * ggml_cuda_cpy_fn_ptr = nullptr;
+    std::vector<void *> ggml_cuda_cpy_fn_ptrs;
 
     if (cuda_ctx->cuda_graph->graph == nullptr) {
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
             cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
 #endif
         }
     }
@@ -2523,14 +2504,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
             if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
                 use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to split buffer\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
 #endif
             }
 
             if (node->op == GGML_OP_MUL_MAT_ID) {
                 use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
 #endif
             }
 
@@ -2539,16 +2520,17 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                 // Changes in batch size or context size can cause changes to the grid size of some kernels.
                 use_cuda_graph = false;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
 #endif
             }
 
             if (node->op == GGML_OP_CPY) {
                 // store the copy op parameter which changes with each token.
                 cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
-                if (ggml_cuda_cpy_fn_ptr == nullptr) {
-                    // store a pointer to the copy op CUDA kernel to identify it later
-                    ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                // store a pointer to each copy op CUDA kernel to identify it later
+                void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
+                    ggml_cuda_cpy_fn_ptrs.push_back(ptr);
                 }
             }
 
@@ -2558,7 +2540,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         }
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
-        if (cuda_graph_update_required) {
+        if (use_cuda_graph && cuda_graph_update_required) {
             cuda_ctx->cuda_graph->number_consecutive_updates++;
         } else {
             cuda_ctx->cuda_graph->number_consecutive_updates = 0;
@@ -2567,7 +2549,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
             cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
 #endif
         }
     }
@@ -2605,7 +2587,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
-                    fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+                    GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                 }
                 GGML_ASSERT(ok);
             }
@@ -2624,7 +2606,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                 use_cuda_graph = false;
                 cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to failed graph capture\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
 #endif
             } else {
                 graph_evaluated_or_captured = true; // CUDA graph has been captured
@@ -2645,10 +2627,8 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 
         if (cuda_graph_update_required) {
             // Extract nodes from graph
-            if (cuda_ctx->cuda_graph->num_nodes == 0) {
-                // First call with null argument gets number of nodes in graph
-                CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
-            }
+            // First call with null argument gets number of nodes in graph
+            CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
             // Subsequent call with non-null argument gets nodes
             cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
             cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
@@ -2678,7 +2658,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
             int k = 0;
             for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
-                if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) {
+                if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
                     char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
                     cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
                     CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
@@ -2691,7 +2671,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
         if (stat == cudaErrorGraphExecUpdateFailure) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: CUDA graph update failed\n", __func__);
+            GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
 #endif
             // The pre-existing graph exec cannot be updated due to violated constraints
             // so instead clear error and re-instantiate
@@ -2725,7 +2705,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
-                    return true;
+                    return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
             }
@@ -2825,11 +2805,14 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
+            return true;
         case GGML_OP_ROPE:
+            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
@@ -2846,10 +2829,14 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
             return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
 #else
-            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+            if (op->src[0]->ne[0] == 128) {
                 return true;
             }
-            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+            if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA &&
+                op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;
@@ -2858,6 +2845,20 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
     GGML_UNUSED(backend);
 }
 
+GGML_CALL static bool ggml_backend_cuda_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cuda_split(buft)) {
+        return true;
+    }
+
+    if (ggml_backend_buft_is_cuda(buft)) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+        ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+        return buft_ctx->device == cuda_ctx->device;
+    }
+
+    return false;
+}
+
 GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
@@ -2930,9 +2931,11 @@ static ggml_backend_i ggml_backend_cuda_interface = {
     /* .synchronize             = */ ggml_backend_cuda_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .supports_op             = */ ggml_backend_cuda_supports_op,
+    /* .supports_buft           = */ ggml_backend_cuda_supports_buft,
     /* .offload_op              = */ ggml_backend_cuda_offload_op,
     /* .event_new               = */ ggml_backend_cuda_event_new,
     /* .event_free              = */ ggml_backend_cuda_event_free,
@@ -2948,13 +2951,13 @@ static ggml_guid_t ggml_backend_cuda_guid() {
 
 GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
     if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
-        fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
+        GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
         return nullptr;
     }
 
     ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
     if (ctx == nullptr) {
-        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__);
         return nullptr;
     }
 
@@ -2998,8 +3001,8 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
         // clear the error
         cudaGetLastError();
 
-        fprintf(stderr, "%s: warning: failed to register %.2f MiB of pinned memory: %s\n", __func__,
-                size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
         return false;
     }
     return true;

ggml/src/ggml-cuda/argsort.cu

@@ -73,6 +73,7 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
     const dim3 block_nums(1, nrows, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
+    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
     GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
 
     if (order == GGML_SORT_ORDER_ASC) {

ggml/src/ggml-cuda/common.cuh

@@ -79,13 +79,8 @@
 #define cudaHostRegisterReadOnly hipHostRegisterReadOnly
 #define cudaHostUnregister hipHostUnregister
 #define cudaLaunchHostFunc hipLaunchHostFunc
-#ifdef GGML_HIP_UMA
-#define cudaMalloc hipMallocManaged
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
-#else
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#endif
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync
@@ -144,29 +139,13 @@
 #define CC_PASCAL     600
 #define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
 #define CC_VOLTA      700
+#define CC_TURING     750
 #define CC_AMPERE     800
 #define CC_OFFSET_AMD 1000000
 #define CC_RDNA1      (CC_OFFSET_AMD + 1010)
 #define CC_RDNA2      (CC_OFFSET_AMD + 1030)
 #define CC_RDNA3      (CC_OFFSET_AMD + 1100)
 
-// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
-// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
-// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
-// -  7B quantum model: +100-200 MB
-// - 13B quantum model: +200-400 MB
-//
-//#define GGML_CUDA_FORCE_MMQ
-
-// TODO: improve this to be correct for more hardware
-//       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
-#if !defined(GGML_CUDA_FORCE_MMQ)
-#define CUDA_USE_TENSOR_CORES
-#endif
-
-#define MMVQ_MAX_BATCH_SIZE  8 // max batch size to use MMVQ kernels
-#define  MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
-
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
 #if defined(_MSC_VER)
@@ -315,20 +294,50 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #endif
     return c;
 }
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
 #endif // defined(GGML_USE_HIPBLAS)
 
-#define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#define FP16_AVAILABLE
+#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
 
-#define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
+#define FAST_FP16_AVAILABLE
+#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-static bool fast_fp16_available(const int cc) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#define FP16_MMA_AVAILABLE
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#define INT8_MMA_AVAILABLE
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+
+static constexpr bool fast_fp16_available(const int cc) {
     return cc >= CC_PASCAL && cc != 610;
 }
 
-static bool fp16_mma_available(const int cc) {
+static constexpr bool fp16_mma_available(const int cc) {
     return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
 }
 
+static constexpr bool int8_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
+}
+
 [[noreturn]]
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
@@ -370,7 +379,7 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 }
 
 static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-#if FP16_AVAILABLE
+#ifdef FP16_AVAILABLE
 
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
@@ -403,7 +412,7 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 }
 
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
-#if FP16_AVAILABLE
+#ifdef FP16_AVAILABLE
 
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
@@ -463,6 +472,159 @@ static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -
 
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
+static __device__ __forceinline__ float get_alibi_slope(
+    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return powf(base, exph);
+}
+
+template <ggml_type type>
+struct ggml_cuda_type_traits;
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_F16> {
+    static constexpr int qk = 1;
+    static constexpr int qr = 1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
+    static constexpr int qk = QK4_0;
+    static constexpr int qr = QR4_0;
+    static constexpr int qi = QI4_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_1> {
+    static constexpr int qk = QK4_1;
+    static constexpr int qr = QR4_1;
+    static constexpr int qi = QI4_1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_0> {
+    static constexpr int qk = QK5_0;
+    static constexpr int qr = QR5_0;
+    static constexpr int qi = QI5_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_1> {
+    static constexpr int qk = QK5_1;
+    static constexpr int qr = QR5_1;
+    static constexpr int qi = QI5_1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
+    static constexpr int qk = QK8_0;
+    static constexpr int qr = QR8_0;
+    static constexpr int qi = QI8_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_K;
+    static constexpr int qi = QI2_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q3_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_K;
+    static constexpr int qi = QI3_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR4_K;
+    static constexpr int qi = QI4_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR5_K;
+    static constexpr int qi = QI5_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q6_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR6_K;
+    static constexpr int qi = QI6_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XXS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_XXS;
+    static constexpr int qi = QI2_XXS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_XS;
+    static constexpr int qi = QI2_XS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_S;
+    static constexpr int qi = QI2_S;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ3_XXS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_XXS;
+    static constexpr int qi = QI3_XXS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR1_S;
+    static constexpr int qi = QI1_S;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR1_M;
+    static constexpr int qi = QI1_M;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ4_NL> {
+    static constexpr int qk = QK4_NL;
+    static constexpr int qr = QR4_NL;
+    static constexpr int qi = QI4_NL;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ4_XS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR4_XS;
+    static constexpr int qi = QI4_XS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ3_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_S;
+    static constexpr int qi = QI3_S;
+};
 
 //////////////////////
 
@@ -473,6 +635,7 @@ struct ggml_cuda_device_info {
         int     cc;                 // compute capability
         int     nsm;                // number of streaming multiprocessors
         size_t  smpb;               // max. shared memory per block
+        size_t  smpbo;              // max. shared memory per block (with opt-in)
         bool    vmm;                // virtual memory support
         size_t  vmm_granularity;    // granularity of virtual memory
         size_t  total_vram;

ggml/src/ggml-cuda/concat.cu

@@ -0,0 +1,196 @@
+#include "concat.cuh"
+
+// contiguous kernels
+static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+
+    if (nidx < ne00) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * gridDim.y;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            (nidx - ne00) +
+            blockIdx.y * (ne0 - ne00) +
+            blockIdx.z * (ne0 - ne00) * gridDim.y;
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+
+    if (blockIdx.y < ne01) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * ne01;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            (blockIdx.y - ne01) * ne0 +
+            blockIdx.z * ne0 * (gridDim.y - ne01);
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void concat_f32_dim2(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;
+    }
+
+    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, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    if (dim == 0) {
+        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
+        return;
+    }
+    if (dim == 1) {
+        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
+        return;
+    }
+    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+}
+
+// non-contiguous kernel (slow)
+static __global__ void concat_f32_non_cont(
+        const char * src0,
+        const char * src1,
+              char * dst,
+           int64_t   ne00,
+           int64_t   ne01,
+           int64_t   ne02,
+           int64_t   ne03,
+          uint64_t   nb00,
+          uint64_t   nb01,
+          uint64_t   nb02,
+          uint64_t   nb03,
+           int64_t /*ne10*/,
+           int64_t /*ne11*/,
+           int64_t /*ne12*/,
+           int64_t /*ne13*/,
+          uint64_t   nb10,
+          uint64_t   nb11,
+          uint64_t   nb12,
+          uint64_t   nb13,
+           int64_t   ne0,
+           int64_t /*ne1*/,
+           int64_t /*ne2*/,
+           int64_t /*ne3*/,
+          uint64_t   nb0,
+          uint64_t   nb1,
+          uint64_t   nb2,
+          uint64_t   nb3,
+          int32_t   dim) {
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+    const float * x;
+
+    for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+        }
+
+        float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
+    }
+}
+
+
+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];
+
+    cudaStream_t stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const float * src0_d = (const float *)src0->data;
+        const float * src1_d = (const float *)src1->data;
+
+        float * dst_d = (float *)dst->data;
+
+        if (dim != 3) {
+            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),
+                        src0->ne[0], src0->ne[1], src0->ne[2],
+                        dst->ne[0],  dst->ne[1],  dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            CUDA_CHECK(cudaMemcpyAsync(dst_d,           src0_d, size0, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
+        }
+    } else {
+        dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
+        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+                (const char *)src0->data,
+                (const char *)src1->data,
+                (      char *)dst->data,
+                src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+                dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
+                dst->nb[0],  dst->nb[1],  dst->nb[2],  dst->nb[3], dim);
+    }
+}

ggml/src/ggml-cuda/convert.cu

@@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     const block_q2_K * x = (const block_q2_K *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t n   = tid/32;
     const int64_t l   = tid - 32*n;
     const int64_t is  = 8*n + l/16;
@@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
     y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int64_t is = tid/16;  // 0 or 1
-    const int64_t il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
 }
 
 template<typename dst_t>
@@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const int64_t i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
 
-#if QK_K == 256
     const int64_t r = threadIdx.x/4;
     const int64_t tid = r/2;
     const int64_t is0 = r%2;
@@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * hm = x[i].hmask;
 
     for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int64_t tid = threadIdx.x;
-    const int64_t is  = tid/16;  // 0 or 1
-    const int64_t il  = tid%16;  // 0...15
-    const int64_t im  = il/8;    // 0...1
-    const int64_t in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
 }
 
-#if QK_K == 256
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
     if (j < 4) {
         d = q[j] & 63; m = q[j + 4] & 63;
@@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
         m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
     }
 }
-#endif
 
 template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 32 threads
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/8;
@@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
         y[l + 0] = d1 * (q[l] & 0xF) - m1;
         y[l +32] = d2 * (q[l] >>  4) - m2;
     }
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
 }
 
 template<typename dst_t>
@@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/16;   // il is in 0...3
@@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     hm <<= 1;
     y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
     y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t q = x[i].qs[tid];
-    const int64_t im = tid/8;  // 0...3
-    const int64_t in = tid%8;  // 0...7
-    const int64_t is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
 }
 
 template<typename dst_t>
@@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const block_q6_K * x = (const block_q6_K *) vx;
 
     const int64_t i = blockIdx.x;
-#if QK_K == 256
 
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
@@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
     y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
     y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/16;         // 0 or 1
-    const int64_t il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
 }
 
 template<typename dst_t>
@@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     const block_iq1_s * x = (const block_iq1_s  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     const block_iq1_m * x = (const block_iq1_m  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
-
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
         y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
-
 }
 
-#if QK_K != 64
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const int64_t i   = blockIdx.x;
@@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 }
-#endif
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
 template<typename dst_t>
 static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k
 template<typename dst_t>
 static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t
 template<typename dst_t>
 static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = (k + QK_K - 1) / QK_K;
-#if QK_K == 64
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-#else
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template <typename src_t, typename dst_t>

ggml/src/ggml-cuda/dmmv.cu

@@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
@@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
         tmp += dall * sum1 - dmin * sum2;
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const float2 dall = __half22float2(x[i].dm);
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x * sum1 - dall.y * sum2;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
-
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
@@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
         tmp += d * sum;
 
     }
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 
     const block_q4_K * x = (const block_q4_K *)vx + ib0;
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 #endif
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
         tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
     }
 
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 
@@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     const block_q6_K * x = (const block_q6_K *)vx + ib0;
 
-#if QK_K == 256
-
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
@@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     }
 
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 
@@ -573,10 +422,22 @@ static __device__ void convert_f16(const void * vx, const int64_t ib, const int
     v.y = x[ib + iqs + 1];
 }
 
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
+        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
+        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
+        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
+        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
+        type == GGML_TYPE_F16 ? convert_f16 :
+        nullptr;
+}
+
+template <ggml_type type>
 static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    // qk = quantized weights per x block
-    // qr = number of quantized weights per data value in x block
+    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
+    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
+    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
+
     const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
 
     if (row >= nrows) {
@@ -644,7 +505,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
     // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -653,7 +514,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -662,7 +523,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -671,7 +532,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -680,7 +541,7 @@ static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -731,7 +592,7 @@ static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, floa
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<1, 1, convert_f16>
+    dequantize_mul_mat_vec<GGML_TYPE_F16>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 

ggml/src/ggml-cuda/fattn-common.cuh

@@ -0,0 +1,741 @@
+#pragma once
+
+#include "common.cuh"
+#include "convert.cuh"
+#include "vecdotq.cuh"
+
+#include <cstdint>
+
+#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.
+
+typedef void (* fattn_kernel_t)(
+        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 nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3);
+
+typedef half (*vec_dot_KQ_f16_t)(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
+typedef float (*vec_dot_KQ_f32_t)(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+
+    const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    half sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_0;
+        const int shift = k_KQ & (QI8_1/2);
+
+        const int v = (get_int_from_uint8(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = __dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 sum2 = __half2half2(K_q4_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE];
+            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2) /* *8/QI8_1 == 1 */);
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            sum += (T) (__half2float(K_q4_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (8/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y));
+        }
+    }
+
+    return sum;
+#else
+    GGML_UNUSED(K_c);
+    GGML_UNUSED(Q_v);
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+    NO_DEVICE_CODE;
+#endif  // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+
+    const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        const int v = (get_int_from_uint8_aligned(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = __dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 d4d8_m4s8 = K_q4_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE];
+            const half2 sumid4d8_m4s8scaled = d4d8_m4s8 * make_half2(sumi, 1.0f/QI8_1);
+            sum += (T) (__low2half(sumid4d8_m4s8scaled) + __high2half(sumid4d8_m4s8scaled));
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            const float sumid4d8   =  __low2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi;
+            const float m4s8scaled = __high2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1;
+
+            sum += (T) (sumid4d8 + m4s8scaled);
+        }
+    }
+
+    return sum;
+#else
+    GGML_UNUSED(K_c);
+    GGML_UNUSED(Q_v);
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+    NO_DEVICE_CODE;
+#endif  // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+
+    const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_0;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v = (get_int_from_uint8(K_q5_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int vh = get_int_from_uint8(K_q5_0[ib].qh, 0) >> (iqs8 * QI5_0);
+        v |= (vh <<  4) & 0x00000010; // 0 ->  4
+        v |= (vh << 11) & 0x00001000; // 1 -> 12
+        v |= (vh << 18) & 0x00100000; // 2 -> 20
+        v |= (vh << 25) & 0x10000000; // 3 -> 28
+
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = __dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 sum2 = __half2half2(K_q5_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE];
+            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2)*__float2half(2.0f)) /* *16/QI8_1 == 2 */;
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            sum += (T) (__half2float(K_q5_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (16/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y));
+        }
+    }
+
+    return sum;
+#else
+    GGML_UNUSED(K_c);
+    GGML_UNUSED(Q_v);
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+    NO_DEVICE_CODE;
+#endif  // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+
+    const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_1;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v = (get_int_from_uint8(K_q5_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int vh = get_int_from_uint8(K_q5_1[ib].qh, 0) >> (iqs8 * QI5_1);
+        v |= (vh <<  4) & 0x00000010; // 0 ->  4
+        v |= (vh << 11) & 0x00001000; // 1 -> 12
+        v |= (vh << 18) & 0x00100000; // 2 -> 20
+        v |= (vh << 25) & 0x10000000; // 3 -> 28
+
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = __dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 d5d8_m5s8 = K_q5_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE];
+            const half2 sumid5d8_m5s8scaled = d5d8_m5s8 * make_half2(sumi, 1.0f/QI8_1);
+            sum += (T) (__low2half(sumid5d8_m5s8scaled) + __high2half(sumid5d8_m5s8scaled));
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            const float sumid5d8   =  __low2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi;
+            const float m5s8scaled = __high2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1;
+
+            sum += (T) (sumid5d8 + m5s8scaled);
+        }
+    }
+
+    return sum;
+#else
+    GGML_UNUSED(K_c);
+    GGML_UNUSED(Q_v);
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+    NO_DEVICE_CODE;
+#endif  // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template <typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+
+    const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib  = k_KQ / QI8_0;
+        const int iqs = k_KQ % QI8_0;
+
+        const int v = get_int_from_int8(K_q8_0[ib].qs, iqs);
+
+        T Q_d;
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+            Q_d = __low2half(Q_ds[k_KQ_0/WARP_SIZE]);
+        } else {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+            Q_d = Q_ds[k_KQ_0/WARP_SIZE].x;
+        }
+
+        sum += vec_dot_q8_0_q8_1_impl<T, 1>(&v, &Q_q8[k_KQ_0/WARP_SIZE], K_q8_0[ib].d, Q_d);
+    }
+
+    return sum;
+#else
+    GGML_UNUSED(K_c);
+    GGML_UNUSED(Q_v);
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+    NO_DEVICE_CODE;
+#endif  // __CUDA_ARCH__ >= MIN_CC_DP4A
+}
+
+template <typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) {
+
+    const half2 * K_h2 = (const half2 *) K_c;
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        const half2 * Q_h2 = (const half2 *) Q_v;
+
+        half2 sum2 = make_half2(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_KQ];
+            sum2 += K_ik * Q_h2[k_KQ_0/WARP_SIZE];
+        }
+
+        return __low2half(sum2) + __high2half(sum2);
+    }
+#endif // FP16_AVAILABLE
+
+    const float2 * Q_f2 = (const float2 *) Q_v;
+
+    float sum = 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_KQ];
+        sum +=  __low2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].x;
+        sum += __high2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].y;
+    }
+
+    return sum;
+}
+
+template <typename Tds>
+static __device__ __forceinline__ void quantize_q8_1_to_shared(
+    const float * __restrict__ x, const float scale, int * __restrict__ yq32, void * __restrict__ yds) {
+
+    float vals[sizeof(int)] = {0.0f};
+#pragma unroll
+    for (int l = 0; l < sizeof(int); ++l) {
+        vals[l] = scale * x[4*threadIdx.x + l];
+    }
+
+    float amax = fabsf(vals[0]);
+    float sum  = vals[0];
+#pragma unroll
+    for (int l = 1; l < sizeof(int); ++l) {
+        amax = fmaxf(amax, fabsf(vals[l]));
+        sum += vals[l];
+    }
+#pragma unroll
+    for (int mask = QI8_1/2; mask > 0; mask >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, 32));
+        sum +=             __shfl_xor_sync(0xFFFFFFFF, sum,  mask, 32);
+    }
+
+    const float d = amax / 127;
+    int q32 = 0;
+    int8_t * q8 = (int8_t *) &q32;
+
+    if (d != 0.0f) {
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            q8[l] = roundf(vals[l] / d);
+        }
+    }
+
+    yq32[threadIdx.x] = q32;
+    if (threadIdx.x % QI8_1 == 0) {
+        if (std::is_same<Tds, half2>::value) {
+            ((half2  *) yds)[threadIdx.x/QI8_1] =  make_half2(d, sum);
+        } else {
+            ((float2 *) yds)[threadIdx.x/QI8_1] = make_float2(d, sum);
+        }
+    }
+}
+
+typedef half  (*dequantize_1_f16_t)(const void *, const int64_t);
+typedef float (*dequantize_1_f32_t)(const void *, const int64_t);
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q4_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int64_t ib    =  i          /  QK4_0;
+    const int     iqs   =  i          % (QK4_0/2);
+    const int     shift = (i % QK4_0) / (QK4_0/2);
+
+    const T   d  = x[ib].d;
+    const int q0 = x[ib].qs[iqs];
+    const int q  = ((q0 >> (4*shift)) & 0x0F) - 8;
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q4_1(const void * __restrict__ vx, const int64_t i) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int64_t ib    =  i          /  QK4_1;
+    const int     iqs   =  i          % (QK4_1/2);
+    const int     shift = (i % QK4_1) / (QK4_1/2);
+
+    const half2 dm = x[ib].dm;
+    const int   q0 = x[ib].qs[iqs];
+    const int   q  = ((q0 >> (4*shift)) & 0x0F);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return __low2half(dm)*((half) q) + __high2half(dm);
+    }
+#endif // FP16_AVAILABLE
+
+    return __low2float(dm)*((float) q) + __high2float(dm);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q5_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const int64_t ib    =  i          /  QK5_0;
+    const int     idq   =  i          %  QK5_0;
+    const int     iqs   =  i          % (QK5_0/2);
+    const int     shift = (i % QK5_0) / (QK5_0/2);
+
+    const T   d   = x[ib].d;
+    const int ql0 = x[ib].qs[iqs];
+    const int qh0 = get_int_from_uint8(x[ib].qh, 0);
+    const int ql  = ((ql0 >> (4*shift)) & 0x0F);
+    const int qh  = ((qh0 >> idq) << 4) & 0x10;
+    const int q   = (ql | qh) - 16;
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q5_1(const void * __restrict__ vx, const int64_t i) {
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const int64_t ib    =  i          /  QK5_1;
+    const int     idq   =  i          %  QK5_1;
+    const int     iqs   =  i          % (QK5_1/2);
+    const int     shift = (i % QK5_1) / (QK5_1/2);
+
+    const half2 dm  = x[ib].dm;
+    const int   ql0 = x[ib].qs[iqs];
+    const int   qh0 = get_int_from_uint8_aligned(x[ib].qh, 0);
+    const int   ql  = ((ql0 >> (4*shift)) & 0x0F);
+    const int   qh  = ((qh0 >> idq) << 4) & 0x10;
+    const int   q   = (ql | qh);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return __low2half(dm)*((half) q) + __high2half(dm);
+    }
+#endif // FP16_AVAILABLE
+
+    return __low2float(dm)*((float) q) + __high2float(dm);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const int64_t ib  = i / QK8_0;
+    const int     iqs = i % QK8_0;
+
+    const T   d = x[ib].d;
+    const int q = x[ib].qs[iqs];
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ vx, const int64_t i) {
+    const half * x = (const half *) vx;
+
+    return x[i];
+}
+
+template <int D>
+constexpr __device__ vec_dot_KQ_f16_t get_vec_dot_KQ_f16(ggml_type type_K) {
+    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<half, D> :
+        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<half, D> :
+        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<half, D> :
+        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<half, D> :
+        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<half, D> :
+        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<half, D> :
+        nullptr;
+}
+
+template <int D>
+constexpr __device__ vec_dot_KQ_f32_t get_vec_dot_KQ_f32(ggml_type type_K) {
+    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<float, D> :
+        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<float, D> :
+        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<float, D> :
+        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<float, D> :
+        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<float, D> :
+        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<float, D> :
+        nullptr;
+}
+
+constexpr __device__ dequantize_1_f16_t get_dequantize_1_f16(ggml_type type_V) {
+    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<half> :
+        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<half> :
+        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<half> :
+        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<half> :
+        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<half> :
+        type_V == GGML_TYPE_F16 ? dequantize_1_f16<half> :
+        nullptr;
+}
+
+constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
+    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<float> :
+        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<float> :
+        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<float> :
+        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<float> :
+        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<float> :
+        type_V == GGML_TYPE_F16 ? dequantize_1_f16<float> :
+        nullptr;
+}
+
+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;
+}
+
+static void on_no_fattn_vec_case(const int D) {
+    if (D == 64) {
+        fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
+        fprintf(stderr, "By default only f16 KV cache is supported.\n");
+        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
+        GGML_ASSERT(false);
+    } else if (D == 128) {
+        fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
+        fprintf(stderr, "Supported combinations:\n");
+        fprintf(stderr, "  - K == q4_0, V == q4_0,  4.50 BPV\n");
+        fprintf(stderr, "  - K == q8_0, V == q8_0,  8.50 BPV\n");
+        fprintf(stderr, "  - K == f16,  V == f16,  16.00 BPV\n");
+        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
+        GGML_ASSERT(false);
+    } else {
+        fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
+        fprintf(stderr, "Only f16 is supported.\n");
+        GGML_ASSERT(false);
+    }
+}
+
+template <int D, int parallel_blocks>
+void launch_fattn(
+    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel,
+    const int nwarps, const int cols_per_block, const bool need_f16_K, const bool need_f16_V
+) {
+    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->type == GGML_TYPE_F32);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
+    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
+                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t main_stream = ctx.stream();
+
+    ggml_cuda_pool_alloc<half>   K_f16(pool);
+    ggml_cuda_pool_alloc<half>   V_f16(pool);
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    char * K_data = (char *) K->data;
+    size_t nb11 = K->nb[1];
+    size_t nb12 = K->nb[2];
+    size_t nb13 = K->nb[3];
+
+    char * V_data = (char *) V->data;
+    size_t nb21 = V->nb[1];
+    size_t nb22 = V->nb[2];
+    size_t nb23 = V->nb[3];
+
+    if (need_f16_K && K->type != GGML_TYPE_F16) {
+        K_f16.alloc(ggml_nelements(K));
+        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
+        to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
+        K_data = (char *) K_f16.ptr;
+
+        const size_t bs = ggml_blck_size(K->type);
+        const size_t ts = ggml_type_size(K->type);
+
+        nb11 = nb11*bs*sizeof(half)/ts;
+        nb12 = nb12*bs*sizeof(half)/ts;
+        nb13 = nb13*bs*sizeof(half)/ts;
+    }
+
+    if (need_f16_V && V->type != GGML_TYPE_F16) {
+        V_f16.alloc(ggml_nelements(V));
+        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+        to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+        V_data = (char *) V_f16.ptr;
+
+        const size_t bs = ggml_blck_size(V->type);
+        const size_t ts = ggml_type_size(V->type);
+
+        nb21 = nb21*bs*sizeof(half)/ts;
+        nb22 = nb22*bs*sizeof(half)/ts;
+        nb23 = nb23*bs*sizeof(half)/ts;
+    }
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    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);
+
+    fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
+        (const char *) Q->data,
+        K_data,
+        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],
+        nb11, nb12, nb13,
+        nb21, nb22, nb23,
+        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());
+}

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -0,0 +1,319 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F16 64
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_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 nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef 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 half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
+    half2 * KQ2 = (half2 *) KQ;
+
+    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
+
+    half kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -HALF_MAX_HALF;
+    }
+    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
+
+    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ half2 Q_h2[ncols][D/2];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
+            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        half kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#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;
+
+                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+            }
+        }
+
+        __syncthreads();
+
+        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
+            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
+            half2 Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
+                const half2 val = h2exp(diff);
+                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
+                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
+            half2  V_k[(D/2)/WARP_SIZE][2];
+            half2 KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
+                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val /= __half2half2(kqsum_j);
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}