Faster AVX2 dot product for IQ2_XS (#5187)

* iq2xs: faster AVX2 dot product

* iq2xs: small AVX2 imrovement

* Speed up computing sign bits in AVX2 iq2_xs dot product

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Peter Reid <peter@peterreid.net>

.devops/server-cuda.Dockerfile

@@ -0,0 +1,32 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/server /server
+
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -0,0 +1,25 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+ARG UBUNTU_VERSION=22.04
+
+FROM intel/hpckit:$ONEAPI_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
+    cmake --build . --config Release --target main server
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -0,0 +1,45 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/server" ]

.devops/server.Dockerfile

@@ -0,0 +1,20 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+RUN make
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.ecrc

@@ -1,4 +1,5 @@
 {
+  "Exclude": ["^\\.gitmodules$"],
   "Disable": {
     "IndentSize": true
   }

.github/workflows/build.yml

@@ -72,7 +72,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -107,7 +107,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -141,7 +141,48 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest -L main --verbose
+
+  ubuntu-22-cmake-sycl:
+    runs-on: ubuntu-22.04
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: add oneAPI to apt
+        shell: bash
+        run: |
+          cd /tmp
+          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
+          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
+
+      - name: install oneAPI dpcpp compiler
+        shell: bash
+        run: |
+          sudo apt update
+          sudo apt install intel-oneapi-compiler-dpcpp-cpp
+
+      - name: install oneAPI MKL library
+        shell: bash
+        run: |
+          sudo apt install intel-oneapi-mkl-devel
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          mkdir build
+          cd build
+          cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake --build . --config Release -j $(nproc)
 
   # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
   #       how to debug it.
@@ -202,7 +243,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -296,6 +337,7 @@ jobs:
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
       SDE_VERSION: 9.33.0-2024-01-07
+      VULKAN_VERSION: 1.3.261.1
 
     strategy:
       matrix:
@@ -312,6 +354,8 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+          - build: 'kompute'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -320,6 +364,12 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Clone Kompute submodule
+        id: clone_kompute
+        if: ${{ matrix.build == 'kompute' }}
+        run: |
+          git submodule update --init kompute
+
       - name: Download OpenCL SDK
         id: get_opencl
         if: ${{ matrix.build == 'clblast' }}
@@ -354,6 +404,15 @@ jobs:
           $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
           & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
 
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'kompute' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
       - name: Build
         id: cmake_build
         run: |
@@ -391,10 +450,11 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
+        # not all machines have native AVX-512
+        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
-          ctest -C Release --verbose --timeout 900
+          ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
@@ -406,7 +466,7 @@ jobs:
           7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
           $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
           cd build
-          & $sde -future -- ctest -C Release --verbose --timeout 900
+          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
 
       - name: Determine tag name
         id: tag

.github/workflows/docker.yml

@@ -28,14 +28,18 @@ jobs:
         config:
           - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
           #                     have disabled them for now until the reason why
           #                     is understood.
           - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3

.gitignore

@@ -27,7 +27,7 @@
 lcov-report/
 gcovr-report/
 
-build*/
+build*
 out/
 tmp/
 
@@ -89,20 +89,3 @@ examples/jeopardy/results.txt
 
 poetry.lock
 poetry.toml
-
-# Test binaries
-/tests/test-grammar-parser
-/tests/test-llama-grammar
-/tests/test-double-float
-/tests/test-grad0
-/tests/test-opt
-/tests/test-quantize-fns
-/tests/test-quantize-perf
-/tests/test-sampling
-/tests/test-tokenizer-0-llama
-/tests/test-tokenizer-0-falcon
-/tests/test-tokenizer-1-llama
-/tests/test-tokenizer-1-bpe
-/tests/test-rope
-/tests/test-backend-ops
-/tests/test-autorelease

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "kompute"]
+	path = kompute
+	url = https://github.com/nomic-ai/kompute.git

CMakeLists.txt

@@ -1,5 +1,6 @@
 cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
+include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
@@ -98,11 +99,15 @@ set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
+option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
+option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
+option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
+option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
@@ -121,8 +126,12 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
+if (LLAMA_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
 
-set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
@@ -409,6 +418,28 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
+if (LLAMA_VULKAN)
+    find_package(Vulkan)
+    if (Vulkan_FOUND)
+        message(STATUS "Vulkan found")
+
+        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
+
+        add_library(ggml-vulkan STATIC ggml-vulkan.cpp ggml-vulkan.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+
+        add_compile_definitions(GGML_USE_VULKAN)
+
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+    else()
+        message(WARNING "Vulkan not found")
+    endif()
+endif()
+
 if (LLAMA_HIPBLAS)
     list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 
@@ -454,6 +485,185 @@ if (LLAMA_HIPBLAS)
     endif()
 endif()
 
+if (LLAMA_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 (LLAMA_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.h ggml-sycl.h)
+    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+endif()
+
+if (LLAMA_KOMPUTE)
+    add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+    if (NOT glslc_executable)
+        message(FATAL_ERROR "glslc not found")
+    endif()
+
+    function(compile_shader)
+      set(options)
+      set(oneValueArgs)
+      set(multiValueArgs SOURCES)
+      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+      foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+        )
+
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+            )
+        else()
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+            )
+        endif()
+      endforeach()
+    endfunction()
+
+    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+        message(STATUS "Kompute found")
+        set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+        add_subdirectory(kompute)
+
+        # Compile our shaders
+        compile_shader(SOURCES
+          kompute-shaders/op_scale.comp
+          kompute-shaders/op_scale_8.comp
+          kompute-shaders/op_add.comp
+          kompute-shaders/op_addrow.comp
+          kompute-shaders/op_mul.comp
+          kompute-shaders/op_silu.comp
+          kompute-shaders/op_relu.comp
+          kompute-shaders/op_gelu.comp
+          kompute-shaders/op_softmax.comp
+          kompute-shaders/op_norm.comp
+          kompute-shaders/op_rmsnorm.comp
+          kompute-shaders/op_diagmask.comp
+          kompute-shaders/op_mul_mat_mat_f32.comp
+          kompute-shaders/op_mul_mat_f16.comp
+          kompute-shaders/op_mul_mat_q8_0.comp
+          kompute-shaders/op_mul_mat_q4_0.comp
+          kompute-shaders/op_mul_mat_q4_1.comp
+          kompute-shaders/op_mul_mat_q6_k.comp
+          kompute-shaders/op_getrows_f16.comp
+          kompute-shaders/op_getrows_q4_0.comp
+          kompute-shaders/op_getrows_q4_1.comp
+          kompute-shaders/op_getrows_q6_k.comp
+          kompute-shaders/op_rope_f16.comp
+          kompute-shaders/op_rope_f32.comp
+          kompute-shaders/op_cpy_f16_f16.comp
+          kompute-shaders/op_cpy_f16_f32.comp
+          kompute-shaders/op_cpy_f32_f16.comp
+          kompute-shaders/op_cpy_f32_f32.comp
+        )
+
+        # Create a custom target for our generated shaders
+        add_custom_target(generated_shaders DEPENDS
+          shaderop_scale.h
+          shaderop_scale_8.h
+          shaderop_add.h
+          shaderop_addrow.h
+          shaderop_mul.h
+          shaderop_silu.h
+          shaderop_relu.h
+          shaderop_gelu.h
+          shaderop_softmax.h
+          shaderop_norm.h
+          shaderop_rmsnorm.h
+          shaderop_diagmask.h
+          shaderop_mul_mat_mat_f32.h
+          shaderop_mul_mat_f16.h
+          shaderop_mul_mat_q8_0.h
+          shaderop_mul_mat_q4_0.h
+          shaderop_mul_mat_q4_1.h
+          shaderop_mul_mat_q6_k.h
+          shaderop_getrows_f16.h
+          shaderop_getrows_q4_0.h
+          shaderop_getrows_q4_1.h
+          shaderop_getrows_q6_k.h
+          shaderop_rope_f16.h
+          shaderop_rope_f32.h
+          shaderop_cpy_f16_f16.h
+          shaderop_cpy_f16_f32.h
+          shaderop_cpy_f32_f16.h
+          shaderop_cpy_f32_f32.h
+        )
+
+        # Create a custom command that depends on the generated_shaders
+        add_custom_command(
+            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            DEPENDS generated_shaders
+            COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+        )
+
+        # Add the stamp to the main sources to ensure dependency tracking
+        set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        add_compile_definitions(GGML_USE_KOMPUTE)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
+    else()
+        message(WARNING "Kompute not found")
+    endif()
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -466,23 +676,25 @@ function(get_flags CCID CCVER)
             (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
             (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
         )
-            set(C_FLAGS ${C_FLAGS} -Wdouble-promotion)
+            list(APPEND C_FLAGS -Wdouble-promotion)
         endif()
     elseif (CCID STREQUAL "GNU")
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
         if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation)
+            list(APPEND CXX_FLAGS -Wno-format-truncation)
         endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
-            set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi)
+            list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
     elseif (CCID MATCHES "Intel")
-        # enable max optimization level when using Intel compiler
-        set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-        set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-        add_link_options(-fuse-ld=lld -static-intel)
+        if (NOT LLAMA_SYCL)
+            # enable max optimization level when using Intel compiler
+            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            add_link_options(-fuse-ld=lld -static-intel)
+        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
@@ -510,16 +722,18 @@ if (LLAMA_ALL_WARNINGS)
     endif()
 endif()
 
+set(CUDA_CXX_FLAGS "")
+
 if (LLAMA_CUBLAS)
     set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
     if (NOT MSVC)
-        set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic)
+        list(APPEND CUDA_FLAGS -Wno-pedantic)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
         set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
         if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
         endif()
 
         execute_process(
@@ -547,13 +761,8 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS)  # pass host compiler flags as a single argument
-        if (NOT CUDA_CXX_FLAGS STREQUAL "")
-            set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS})
-        endif()
+        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
     endif()
-
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
 endif()
 
 if (WIN32)
@@ -618,12 +827,7 @@ if (NOT MSVC)
     endif()
 endif()
 
-function(add_compile_option_cpp ARG)
-    # Adds a compile option to C/C++ only, but not for Cuda.
-    # Use, e.g., for CPU-architecture flags.
-    add_compile_options($<$<COMPILE_LANGUAGE:CXX>:${ARG}>)
-    add_compile_options($<$<COMPILE_LANGUAGE:C>:${ARG}>)
-endfunction()
+set(ARCH_FLAGS "")
 
 if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
@@ -636,19 +840,19 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            add_compile_options(-mfp16-format=ieee)
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
-            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
             # Raspberry Pi 2
-            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            add_compile_options(-mno-unaligned-access)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
@@ -659,7 +863,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
             include(cmake/FindSIMD.cmake)
         endif ()
         if (LLAMA_AVX512)
-            add_compile_option_cpp(/arch:AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
@@ -673,49 +877,61 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            add_compile_option_cpp(/arch:AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
         elseif (LLAMA_AVX)
-            add_compile_option_cpp(/arch:AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
         endif()
     else()
         if (LLAMA_NATIVE)
-            add_compile_option_cpp(-march=native)
+            list(APPEND ARCH_FLAGS -march=native)
         endif()
         if (LLAMA_F16C)
-            add_compile_option_cpp(-mf16c)
+            list(APPEND ARCH_FLAGS -mf16c)
         endif()
         if (LLAMA_FMA)
-            add_compile_option_cpp(-mfma)
+            list(APPEND ARCH_FLAGS -mfma)
         endif()
         if (LLAMA_AVX)
-            add_compile_option_cpp(-mavx)
+            list(APPEND ARCH_FLAGS -mavx)
         endif()
         if (LLAMA_AVX2)
-            add_compile_option_cpp(-mavx2)
+            list(APPEND ARCH_FLAGS -mavx2)
         endif()
         if (LLAMA_AVX512)
-            add_compile_option_cpp(-mavx512f)
-            add_compile_option_cpp(-mavx512bw)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512bw)
         endif()
         if (LLAMA_AVX512_VBMI)
-            add_compile_option_cpp(-mavx512vbmi)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
         endif()
         if (LLAMA_AVX512_VNNI)
-            add_compile_option_cpp(-mavx512vnni)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
     if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        add_compile_options(-mcpu=powerpc64le)
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
     else()
-        add_compile_options(-mcpu=native -mtune=native)
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
         #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
     endif()
 else()
     message(STATUS "Unknown architecture")
 endif()
 
+add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
+add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+
+if (LLAMA_CUBLAS)
+    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
+    list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+endif()
+
 if (MINGW)
     # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
@@ -790,11 +1006,14 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA}   ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_CUDA}    ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL}  ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_METAL}   ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}     ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
@@ -871,7 +1090,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}" "${GGML_HEADERS_VULKAN}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

Makefile

@@ -9,7 +9,7 @@ TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -448,6 +448,19 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
+ifdef LLAMA_VULKAN
+	MK_CPPFLAGS  += -DGGML_USE_VULKAN
+	MK_LDFLAGS += -lvulkan
+	OBJS    += ggml-vulkan.o
+
+ifdef LLAMA_VULKAN_CHECK_RESULTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
+endif
+
+ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_VULKAN
+
 ifdef LLAMA_HIPBLAS
 
 	ifeq ($(wildcard /opt/rocm),)
@@ -619,7 +632,7 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
@@ -748,5 +761,8 @@ tests/test-c.o: tests/test-c.c llama.h
 tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

README.md

@@ -14,7 +14,6 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
-- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406
 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
@@ -63,7 +62,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
-- CUDA, Metal and OpenCL GPU backend support
+- CUDA, Metal, OpenCL, SYCL GPU backend support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
 Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves
@@ -112,6 +111,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
 - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
+- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 
 
 **Bindings:**
@@ -121,7 +121,8 @@ as the main playground for developing new features for the [ggml](https://github
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
@@ -289,7 +290,7 @@ In order to build llama.cpp you have three different options.
         sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
             opencl clblast openblas
 
-            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
         ```
 
     **Notes:** With this packages you can build llama.cpp with OPENBLAS and
@@ -597,15 +598,24 @@ Building the program with BLAS support may lead to some performance improvements
 
   You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
+- #### SYCL
+
+  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+  llama.cpp based on SYCL is used to support Intel GPU (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+  For detailed info, please refer to [llama.cpp for SYCL](README_sycl.md).
+
+
 ### Prepare Data & Run
 
 ```bash
 # obtain the original LLaMA model weights and place them in ./models
 ls ./models
 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
-  # [Optional] for models using BPE tokenizers
-  ls ./models
-  65B 30B 13B 7B vocab.json
+# [Optional] for models using BPE tokenizers
+ls ./models
+65B 30B 13B 7B vocab.json
 
 # install Python dependencies
 python3 -m pip install -r requirements.txt
@@ -613,8 +623,8 @@ python3 -m pip install -r requirements.txt
 # convert the 7B model to ggml FP16 format
 python3 convert.py models/7B/
 
-  # [Optional] for models using BPE tokenizers
-  python convert.py models/7B/ --vocabtype bpe
+# [Optional] for models using BPE tokenizers
+python convert.py models/7B/ --vocabtype bpe
 
 # quantize the model to 4-bits (using q4_0 method)
 ./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0
@@ -930,17 +940,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 * Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
-We have two Docker images available for this project:
+We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 
 - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
@@ -966,6 +979,12 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
 ### Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -975,6 +994,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
 docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -988,6 +1008,7 @@ The resulting images, are essentially the same as the non-CUDA images:
 
 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
 
 #### Usage
 
@@ -996,6 +1017,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ### Contributing

README_sycl.md

@@ -0,0 +1,252 @@
+# llama.cpp for SYCL
+
+[Background](#background)
+
+[OS](#os)
+
+[Intel GPU](#intel-gpu)
+
+[Linux](#linux)
+
+[Environment Variable](#environment-variable)
+
+[Known Issue](#known-issue)
+
+[Todo](#todo)
+
+## Background
+
+SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
+
+oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
+
+Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
+
+To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
+
+The llama.cpp for SYCL is used to support Intel GPUs.
+
+For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
+
+## OS
+
+|OS|Status|Verified|
+|-|-|-|
+|Linux|Support|Ubuntu 22.04|
+|Windows|Ongoing| |
+
+
+## Intel GPU
+
+|Intel GPU| Status | Verified Model|
+|-|-|-|
+|Intel Data Center Max Series| Support| Max 1550|
+|Intel Data Center Flex Series| Support| Flex 170|
+|Intel Arc Series| Support| Arc 770|
+|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
+|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
+
+
+## Linux
+
+### Setup Environment
+
+1. Install Intel GPU driver.
+
+a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
+
+Note: for iGPU, please install the client GPU driver.
+
+b. Add user to group: video, render.
+
+```
+sudo usermod -aG render username
+sudo usermod -aG video username
+```
+
+Note: re-login to enable it.
+
+c. Check
+
+```
+sudo apt install clinfo
+sudo clinfo -l
+```
+
+Output (example):
+
+```
+Platform #0: Intel(R) OpenCL Graphics
+ `-- Device #0: Intel(R) Arc(TM) A770 Graphics
+
+
+Platform #0: Intel(R) OpenCL HD Graphics
+ `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
+```
+
+2. Install Intel® oneAPI Base toolkit.
+
+
+a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+
+Recommend to install to default folder: **/opt/intel/oneapi**.
+
+Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
+
+b. Check
+
+```
+source /opt/intel/oneapi/setvars.sh
+
+sycl-ls
+```
+
+There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
+
+Output (example):
+```
+[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+
+```
+
+2. Build locally:
+
+```
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main only
+#cmake --build . --config Release --target main
+
+#build all binary
+cmake --build . --config Release -v
+
+```
+
+or
+
+```
+./examples/sycl/build.sh
+```
+
+Note:
+
+- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+
+### Run
+
+1. Put model file to folder **models**
+
+2. Enable oneAPI running environment
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. List device ID
+
+Run without parameter:
+
+```
+./build/bin/ls-sycl-device
+
+or
+
+./build/bin/main
+```
+
+Check the ID in startup log, like:
+
+```
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
+```
+
+|Attribute|Note|
+|-|-|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+
+4. Set device ID and execute llama.cpp
+
+Set device ID = 0 by **GGML_SYCL_DEVICE=0**
+
+```
+GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+or run by script:
+
+```
+./examples/sycl/run_llama2.sh
+```
+
+Note:
+
+- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+
+
+5. Check the device ID in output
+
+Like：
+```
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+```
+
+
+## Environment Variable
+
+#### Build
+
+|Name|Value|Function|
+|-|-|-|
+|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
+|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
+|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
+|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
+
+#### Running
+
+
+|Name|Value|Function|
+|-|-|-|
+|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
+|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
+
+## Known Issue
+
+- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
+
+  Miss to enable oneAPI running environment.
+
+  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
+
+
+- Hang during startup
+
+  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
+
+  Solution: add **--no-mmap**.
+
+## Todo
+
+- Support to build in Windows.
+
+- Support multiple cards.

ci/README.md

@@ -22,4 +22,8 @@ bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
 # with CUDA support
 GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with SYCL support
+source /opt/intel/oneapi/setvars.sh
+GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```

ci/run.sh

@@ -10,6 +10,9 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with SYCL support
+# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -22,9 +25,9 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -v $OUT/*.log
-rm -v $OUT/*.exit
-rm -v $OUT/*.md
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
@@ -40,6 +43,14 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
 fi
 
+if [ ! -z ${GG_BUILD_SYCL} ]; then
+    if [ -z ${ONEAPI_ROOT} ]; then
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
+        exit 1
+    fi
+
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON"
+fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -94,7 +105,7 @@ function gg_run_ctest_debug {
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -123,9 +134,9 @@ function gg_run_ctest_release {
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -141,6 +152,61 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+function gg_get_model {
+    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
+    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_3b ]]; then
+        echo -n "$gguf_3b"
+    elif [[ -s $gguf_7b ]]; then
+        echo -n "$gguf_7b"
+    else
+        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
+        exit 1
+    fi
+}
+
+function gg_run_ctest_with_model_debug {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-debug
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_run_ctest_with_model_release {
+    cd ${SRC}
+
+    local model; model=$(gg_get_model)
+    cd build-ci-release
+    set -e
+    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    set +e
+    cd ..
+}
+
+function gg_sum_ctest_with_model_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+function gg_sum_ctest_with_model_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest with model files in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
 # open_llama_3b_v2
 
 function gg_run_open_llama_3b_v2 {
@@ -183,8 +249,6 @@ function gg_run_open_llama_3b_v2 {
 
     wiki_test_60="${path_wiki}/wiki.test-60.raw"
 
-    ./bin/test-autorelease ${model_f16}
-
     ./bin/quantize ${model_f16} ${model_q8_0} q8_0
     ./bin/quantize ${model_f16} ${model_q4_0} q4_0
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1
@@ -507,14 +571,18 @@ function gg_sum_open_llama_7b_v2 {
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
-
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    python3 -m pip install -r ${SRC}/requirements.txt
-    python3 -m pip install --editable gguf-py
+    # Create a fresh python3 venv and enter it
+    python3 -m venv "$MNT/venv"
+    source "$MNT/venv/bin/activate"
+
+    pip install -r ${SRC}/requirements.txt --disable-pip-version-check
+    pip install --editable gguf-py --disable-pip-version-check
 fi
 
 ret=0
@@ -529,6 +597,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
         else
             test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
+        test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi
 

common/common.cpp

@@ -42,6 +42,10 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL))
+#define GGML_USE_CUBLAS_SYCL
+#endif
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -599,9 +603,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.main_gpu = std::stoi(argv[i]);
-#ifndef GGML_USE_CUBLAS
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUBLAS_SYCL
         } else if (arg == "--split-mode" || arg == "-sm") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -618,9 +622,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-#ifndef GGML_USE_CUBLAS
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUBLAS_SYCL
+
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -643,9 +648,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#ifndef GGML_USE_CUBLAS
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n");
-#endif // GGML_USE_CUBLAS
+#ifndef GGML_USE_CUBLAS_SYCL
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting a tensor split has no effect.\n");
+#endif // GGML_USE_CUBLAS_SYCL
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
@@ -1007,7 +1012,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
     printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
     printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
-#endif
+#endif // LLAMA_SUPPORTS_GPU_OFFLOAD
     printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
     printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
     printf("  -gan N, --grp-attn-n N\n");
@@ -1514,7 +1519,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
     fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
     fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
     fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");

common/sampling.cpp

@@ -13,6 +13,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
         // will be empty (default) if there are parse errors
         if (result->parsed_grammar.rules.empty()) {
             fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            delete result;
             return nullptr;
         }
 

convert-hf-to-gguf.py

@@ -201,6 +201,8 @@ class Model:
             return PlamoModel
         if model_architecture == "CodeShellForCausalLM":
             return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -250,6 +252,8 @@ class Model:
             return gguf.MODEL_ARCH.PLAMO
         if arch == "CodeShellForCausalLM":
             return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -572,6 +576,83 @@ class MPTModel(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
+class OrionModel(Model):
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+
 class BaichuanModel(Model):
     def set_vocab(self):
         self._set_vocab_sentencepiece()

convert.py

@@ -334,7 +334,10 @@ class Params:
 class BpeVocab:
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        self.vocab = self.bpe_tokenizer["model"]["vocab"]
+        try:
+            self.vocab = self.bpe_tokenizer["model"]["vocab"]
+        except KeyError:
+            self.vocab = self.bpe_tokenizer
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
             # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.

examples/CMakeLists.txt

@@ -23,6 +23,9 @@ else()
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
+    if (LLAMA_SYCL)
+        add_subdirectory(sycl)
+    endif()
     add_subdirectory(main)
     add_subdirectory(tokenize)
     add_subdirectory(parallel)

examples/infill/infill.cpp

@@ -241,7 +241,7 @@ int main(int argc, char ** argv) {
     LOG("add_bos: %d\n", add_bos);
 
     bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+    if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
         params.input_suffix.erase(0, 1);
         suff_rm_leading_spc = false;
     }

examples/llama-bench/llama-bench.cpp

@@ -562,6 +562,7 @@ struct test {
     static const int build_number;
     static const bool cuda;
     static const bool opencl;
+    static const bool vulkan;
     static const bool metal;
     static const bool gpu_blas;
     static const bool blas;
@@ -643,6 +644,9 @@ struct test {
         if (opencl) {
             return "OpenCL";
         }
+        if (vulkan) {
+            return "Vulkan";
+        }
         if (metal) {
             return "Metal";
         }
@@ -658,7 +662,7 @@ struct test {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "metal", "gpu_blas", "blas",
+            "cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_threads", "type_k", "type_v",
@@ -682,7 +686,7 @@ struct test {
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
+        if (field == "cuda" || field == "opencl"  || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
             field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
             return BOOL;
         }
@@ -710,7 +714,7 @@ struct test {
         }
         std::vector<std::string> values = {
             build_commit, std::to_string(build_number),
-            std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
@@ -738,6 +742,7 @@ const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
+const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();

examples/llava/clip.cpp

@@ -98,6 +98,7 @@ static std::string format(const char * fmt, ...) {
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_UNKNOWN,
 };
@@ -304,10 +305,18 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
-    struct ggml_tensor * mm_0_w;
-    struct ggml_tensor * mm_0_b;
-    struct ggml_tensor * mm_2_w;
-    struct ggml_tensor * mm_2_b;
+    struct ggml_tensor * mm_0_w = NULL;
+    struct ggml_tensor * mm_0_b = NULL;
+    struct ggml_tensor * mm_2_w = NULL;
+    struct ggml_tensor * mm_2_b = NULL;
+
+    // Yi type models with mlp+normalization projection
+    struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
+    struct ggml_tensor * mm_1_b = NULL;
+    struct ggml_tensor * mm_3_w = NULL;
+    struct ggml_tensor * mm_3_b = NULL;
+    struct ggml_tensor * mm_4_w = NULL;
+    struct ggml_tensor * mm_4_b = NULL;
 
     // MobileVLM projection
     struct ggml_tensor * mm_model_mlp_1_w;
@@ -460,6 +469,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     // pre-layernorm
     {
         embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "pre_ln");
 
         embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
@@ -575,6 +585,27 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+
+        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
         }
         else if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projector
@@ -808,6 +839,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         else {
             new_clip->proj_type = PROJECTOR_TYPE_MLP;
         }
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+            if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
+                new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
+            }
+        }
     }
 
 #ifdef GGML_USE_CUBLAS
@@ -956,11 +992,29 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
 
         // LLaVA projection
-        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
             vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
-            vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
-            vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            try {
+                // Yi-type llava
+                vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
+                vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // missing in Yi-type llava
+                vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+                vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
+                vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
+                vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
+            } catch (std::runtime_error & e) {  }
         }
         else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
@@ -1277,7 +1331,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         ".*weight",
     };
 
-    std::vector<uint8_t> read_data(512);
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
     std::vector<int64_t> hist_all(1 << 4, 0);
@@ -1433,6 +1486,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     }
     else if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
+    } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        return ctx->vision_model.mm_3_b->ne[0];
     }
     else {
         std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];

examples/llava/llava-cli.cpp

@@ -148,10 +148,35 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
 
-    // llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
-    eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos);
+    std::string system_prompt, user_prompt;
+    size_t image_pos = prompt.find("<image>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
+        // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string
+        size_t pos = 0;
+        while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) {
+            user_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+        while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) {
+            system_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+
+        printf("system_prompt: %s\n", system_prompt.c_str());
+        printf("user_prompt: %s\n", user_prompt.c_str());
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
+        user_prompt = prompt + "\nASSISTANT:";
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
     llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
 
     // generate the response
 
@@ -162,6 +187,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
         if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
 
         printf("%s", tmp);
         fflush(stdout);

examples/main/main.cpp

@@ -39,6 +39,17 @@ static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
+static bool file_exists(const std::string &path) {
+    std::ifstream f(path.c_str());
+    return f.good();
+}
+
+static bool file_is_empty(const std::string &path) {
+    std::ifstream f;
+    f.exceptions(std::ifstream::failbit | std::ifstream::badbit);
+    f.open(path.c_str(), std::ios::in | std::ios::binary | std::ios::ate);
+    return f.tellg() == 0;
+}
 
 static void write_logfile(
     const llama_context * ctx, const gpt_params & params, const llama_model * model,
@@ -215,12 +226,12 @@ int main(int argc, char ** argv) {
 
     if (!path_session.empty()) {
         LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
-
-        // fopen to check for existing session
-        FILE * fp = std::fopen(path_session.c_str(), "rb");
-        if (fp != NULL) {
-            std::fclose(fp);
-
+        if (!file_exists(path_session)) {
+            LOG_TEE("%s: session file does not exist, will create.\n", __func__);
+        } else if (file_is_empty(path_session)) {
+            LOG_TEE("%s: The session file is empty. A new session will be initialized.\n", __func__);
+        } else {
+            // The file exists and is not empty
             session_tokens.resize(n_ctx);
             size_t n_token_count_out = 0;
             if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
@@ -229,10 +240,7 @@ int main(int argc, char ** argv) {
             }
             session_tokens.resize(n_token_count_out);
             llama_set_rng_seed(ctx, params.seed);
-
-            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
-        } else {
-            LOG_TEE("%s: session file does not exist, will create\n", __func__);
+            LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
         }
     }
 

examples/quantize-stats/quantize-stats.cpp

@@ -378,6 +378,8 @@ int main(int argc, char ** argv) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }
 
+            ggml_quantize_init(type);
+
             error_stats global_stats {};
 
             for (const auto& kv_tensor : tensors) {

examples/quantize/quantize.cpp

@@ -25,6 +25,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
+    { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
     { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
@@ -36,7 +37,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K",   LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
     { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
-    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
+    { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 5.15G, +0.0008 ppl @ LLaMA-v1-7B", },
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "13.00G              @ 7B", },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B", },

examples/server/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

examples/server/README.md

@@ -4,32 +4,34 @@ This example demonstrates a simple HTTP API server and a simple web front end to
 
 Command line options:
 
--   `--threads N`, `-t N`: Set the number of threads to use during generation.
--   `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
--   `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
--   `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
--   `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
--   `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
--   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
--   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
--   `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
--   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
--   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
--   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
--   `--numa`: Attempt optimizations that help on some NUMA systems.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
--   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
--   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
--   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
--   `--port`: Set the port to listen. Default: `8080`.
--   `--path`: path from which to serve static files (default examples/server/public)
--   `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
--   `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
--   `--embedding`: Enable embedding extraction, Default: disabled.
--   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
--   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
--   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
--   `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
+- `--threads N`, `-t N`: Set the number of threads to use during generation.
+- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
+- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
+- `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
+- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
+- `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
+- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
+- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
+- `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
+- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+- `--numa`: Attempt optimizations that help on some NUMA systems.
+- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
+- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
+- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
+- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
+- `--port`: Set the port to listen. Default: `8080`.
+- `--path`: path from which to serve static files (default examples/server/public)
+- `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
+- `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
+- `--embedding`: Enable embedding extraction, Default: disabled.
+- `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
+- `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
+- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
+- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
+- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
+- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
 
 ## Build
 
@@ -51,20 +53,30 @@ server is build alongside everything else from the root of the project
 
 To get started right away, run the following command, making sure to use the correct path for the model you have:
 
-### Unix-based systems (Linux, macOS, etc.):
+### Unix-based systems (Linux, macOS, etc.)
 
 ```bash
 ./server -m models/7B/ggml-model.gguf -c 2048
 ```
 
-### Windows:
+### Windows
 
 ```powershell
 server.exe -m models\7B\ggml-model.gguf -c 2048
 ```
+
 The above command will start a server that by default listens on `127.0.0.1:8080`.
 You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url.
 
+### Docker
+
+```bash
+docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+
+# or, with CUDA:
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+```
+
 ## Testing with CURL
 
 Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.
@@ -111,12 +123,13 @@ node index.js
 ```
 
 ## API Endpoints
-- **GET** `/health`: Returns the current state of the server:
-    - `{"status": "loading model"}` if the model is still being loaded.
-    - `{"status": "error"}` if the model failed to load.
-    - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
 
--   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
+- **GET** `/health`: Returns the current state of the server:
+  - `{"status": "loading model"}` if the model is still being loaded.
+  - `{"status": "error"}` if the model failed to load.
+  - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
+
+- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
@@ -180,14 +193,13 @@ node index.js
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
-### Result JSON:
-
-* Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+### Result JSON
 
+- Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
 
 - `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
 
-```
+```json
 {
   "content": "<the token selected by the model>",
   "probs": [
@@ -203,6 +215,7 @@ node index.js
   ]
 },
 ```
+
 Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
@@ -219,7 +232,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `tokens_evaluated`: Number of tokens evaluated in total from the prompt
 - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
--   **POST** `/tokenize`: Tokenize a given text.
+- **POST** `/tokenize`: Tokenize a given text.
 
     *Options:*
 
@@ -227,13 +240,13 @@ Notice that each `probs` is an array of length `n_probs`.
 
     Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
--   **POST** `/detokenize`: Convert tokens to text.
+- **POST** `/detokenize`: Convert tokens to text.
 
     *Options:*
 
     `tokens`: Set the tokens to detokenize.
 
--   **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
+- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
 
     *Options:*
 
@@ -241,7 +254,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
--   **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
+- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
 
     *Options:*
 
@@ -251,9 +264,9 @@ Notice that each `probs` is an array of length `n_probs`.
 
     It also accepts all the options of `/completion` except `stream` and `prompt`.
 
--   **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
+- **GET** `/props`: Return the required assistant name and anti-prompt to generate the prompt in case you have specified a system prompt for all slots.
 
--   **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
+- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
 
     *Options:*
 
@@ -281,6 +294,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     print(completion.choices[0].message)
     ```
+
     ... or raw HTTP requests:
 
     ```shell
@@ -302,6 +316,40 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
+- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
+
+    *Options:*
+
+    See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
+
+    *Examples:*
+
+  - input as string
+
+    ```shell
+    curl http://localhost:8080/v1/embeddings \
+    -H "Content-Type: application/json" \
+    -H "Authorization: Bearer no-key" \
+    -d '{
+            "input": "hello",
+            "model":"GPT-4",
+            "encoding_format": "float"
+    }'
+    ```
+
+  - `input` as string array
+
+    ```shell
+    curl http://localhost:8080/v1/embeddings \
+    -H "Content-Type: application/json" \
+    -H "Authorization: Bearer no-key" \
+    -d '{
+            "input": ["hello", "world"],
+            "model":"GPT-4",
+            "encoding_format": "float"
+    }'
+    ```
+
 ## More examples
 
 ### Change system prompt on runtime
@@ -353,6 +401,7 @@ python api_like_OAI.py
 ```
 
 After running the API server, you can use it in Python by setting the API base URL.
+
 ```python
 openai.api_base = "http://<Your api-server IP>:port"
 ```

examples/server/oai.hpp

@@ -0,0 +1,223 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+#include "utils.hpp"
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
+using json = nlohmann::json;
+
+inline static json oaicompat_completion_params_parse(
+    const json &body /* openai api json semantics */)
+{
+    json llama_params;
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
+    llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    llama_params["presence_penalty"]  = json_value(body, "presence_penalty", 0.0);
+    llama_params["seed"]              = json_value(body, "seed", LLAMA_DEFAULT_SEED);
+    llama_params["stream"]            = json_value(body, "stream", false);
+    llama_params["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}
+
+inline static json format_embeddings_response_oaicompat(const json &request, const json &embeddings)
+{
+    json res =
+        json{
+            {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", "list"},
+            {"usage",
+                json{{"prompt_tokens", 0},
+                     {"total_tokens", 0}}},
+            {"data", embeddings}
+        };
+    return res;
+}
+

examples/server/utils.hpp

@@ -0,0 +1,508 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+
+#include "../llava/clip.h"
+
+using json = nlohmann::json;
+
+extern bool server_verbose;
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+#if SERVER_VERBOSE != 1
+#define LOG_VERBOSE(MSG, ...)
+#else
+#define LOG_VERBOSE(MSG, ...)                                            \
+    do                                                                   \
+    {                                                                    \
+        if (server_verbose)                                              \
+        {                                                                \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while (0)
+#endif
+
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+
+//
+// parallel
+//
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// TODO: can become bool if we can't find use of more states
+enum slot_state
+{
+    IDLE,
+    PROCESSING,
+};
+
+enum slot_command
+{
+    NONE,
+    LOAD_PROMPT,
+    RELEASE,
+};
+
+struct slot_params
+{
+    bool stream       = true;
+    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+
+    uint32_t seed      = -1; // RNG seed
+    int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t  n_predict = -1; // new tokens to predict
+
+    std::vector<std::string> antiprompt;
+
+    json input_prefix;
+    json input_suffix;
+};
+
+struct slot_image
+{
+    int32_t id;
+
+    bool request_encode_image = false;
+    float * image_embedding = nullptr;
+    int32_t image_tokens = 0;
+
+    clip_image_u8 * img_data;
+
+    std::string prefix_prompt; // before of this image
+};
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
+{
+    nlohmann::ordered_json log
+    {
+        {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
+    };
+
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+//
+// server utils
+//
+
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
+inline std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    return chatml_msgs.str();
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_all_task_finished;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        return id++;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when the batch of tasks is finished
+    void on_all_tasks_finished(std::function<void(void)> callback) {
+        callback_all_task_finished = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // Start the main loop. This call is blocking
+    [[noreturn]]
+    void start_loop() {
+        while (true) {
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("callback_all_task_finished", {});
+                // process and update all the multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is finished
+                callback_all_task_finished();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    condition_tasks.wait(lock, [&]{
+                        return !queue_tasks.empty();
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    void add_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    void remove_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+            LOG_VERBOSE("condition_results unblock", {});
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
+        }
+
+        // should never reach here
+    }
+
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
+    }
+
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
+
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {});
+                queue_results.push_back(result);
+                condition_results.notify_one();
+                return;
+            }
+        }
+    }
+};
+
+//
+// base64 utils (TODO: move to common in the future)
+//
+
+static const std::string base64_chars =
+             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+             "abcdefghijklmnopqrstuvwxyz"
+             "0123456789+/";
+
+static inline bool is_base64(uint8_t c)
+{
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
+    int i = 0;
+    int j = 0;
+    int in_ = 0;
+
+    int in_len = encoded_string.size();
+
+    uint8_t char_array_4[4];
+    uint8_t char_array_3[3];
+
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
+        char_array_4[i++] = encoded_string[in_]; in_++;
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
+                char_array_4[i] = base64_chars.find(char_array_4[i]);
+            }
+
+            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+            for (i = 0; (i < 3); i++)
+            {
+                ret.push_back(char_array_3[i]);
+            }
+            i = 0;
+        }
+    }
+
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
+            char_array_4[j] = 0;
+        }
+
+        for (j = 0; j <4; j++)
+        {
+            char_array_4[j] = base64_chars.find(char_array_4[j]);
+        }
+
+        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+        for (j = 0; (j < i - 1); j++)
+        {
+            ret.push_back(char_array_3[j]);
+        }
+    }
+
+    return ret;
+}
+
+//
+// random string / id
+//
+
+static std::string random_string()
+{
+    static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
+
+    std::random_device rd;
+    std::mt19937 generator(rd());
+
+    std::string result(32, ' ');
+
+    for (int i = 0; i < 32; ++i) {
+        result[i] = str[generator() % str.size()];
+    }
+
+    return result;
+}
+
+static std::string gen_chatcmplid()
+{
+    std::stringstream chatcmplid;
+    chatcmplid << "chatcmpl-" << random_string();
+    return chatcmplid.str();
+}

examples/sycl/CMakeLists.txt

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

examples/sycl/README.md

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

examples/sycl/build.sh

@@ -0,0 +1,20 @@
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+mkdir -p build
+cd build
+source /opt/intel/oneapi/setvars.sh
+
+#for FP16
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
+
+#for FP32
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build example/main only
+#cmake --build . --config Release --target main
+
+#build all binary
+cmake --build . --config Release -v

examples/sycl/ls-sycl-device.cpp

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

examples/sycl/run-llama2.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
+source /opt/intel/oneapi/setvars.sh
+
+if [ $# -gt 0 ]; then
+    export GGML_SYCL_DEVICE=$1
+else
+    export GGML_SYCL_DEVICE=0
+fi
+echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
+#export GGML_SYCL_DEBUG=1
+./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+#./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 5 -e -ngl 33 -t 1 -s 0
+

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1705677747,
-        "narHash": "sha256-eyM3okYtMgYDgmYukoUzrmuoY4xl4FUujnsv/P6I/zI=",
+        "lastModified": 1706191920,
+        "narHash": "sha256-eLihrZAPZX0R6RyM5fYAWeKVNuQPYjAkCUBr+JNvtdE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "bbe7d8f876fbbe7c959c90ba2ae2852220573261",
+        "rev": "ae5c332cbb5827f6b1f02572496b141021de335f",
         "type": "github"
       },
       "original": {

ggml-alloc.c

@@ -335,7 +335,9 @@ bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) {
 }
 
 size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) {
-    return alloc->max_size;
+    // FIXME: changes in the tensor sizes compared to the measure graph may cause allocations to fail
+    // to avoid this, we add a 10% margin to the buffer size
+    return alloc->max_size + alloc->max_size/10;
 }
 
 // graph allocator
@@ -776,38 +778,26 @@ size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph)
 }
 
 // utils
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
-    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
 
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
-
-    size_t nbytes = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        if (t->data == NULL && t->view_src == NULL) {
-            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
-        }
-    }
-
-    if (nbytes == 0) {
-        // all the tensors in the context are already allocated
-#ifndef NDEBUG
-        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
-#endif
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+static bool alloc_tensor_range(struct ggml_context * ctx,
+        struct ggml_tensor * first, struct ggml_tensor * last,
+        ggml_backend_buffer_type_t buft, size_t size,
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
     if (buffer == NULL) {
-        // failed to allocate buffer
 #ifndef NDEBUG
-        fprintf(stderr, "%s: failed to allocate buffer\n", __func__);
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
 #endif
-        return NULL;
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free(*buffers[i]);
+        }
+        free(*buffers);
+        return false;
     }
 
     ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
 
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
         if (t->data == NULL) {
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(tallocr, t);
@@ -824,6 +814,76 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     ggml_tallocr_free(tallocr);
 
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            // tensor is too large to fit in a single buffer
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+        // all the tensors in the context are already allocated
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
     return buffer;
 }
 

ggml-backend-impl.h

@@ -19,6 +19,7 @@ extern "C" {
         const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
         ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
         size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
         size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
         // check if tensor data is in host memory
@@ -63,6 +64,11 @@ extern "C" {
     // do not use directly, use ggml_backend_tensor_copy instead
     bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
+    // buffer that contains a collection of buffers
+    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
     //
     // Backend
     //

ggml-backend.c

@@ -27,10 +27,20 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_alignment(buft);
 }
 
+size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    // get_max_size is optional, defaults to SIZE_MAX
+    if (buft->iface.get_max_size) {
+        return buft->iface.get_max_size(buft);
+    }
+    return SIZE_MAX;
+}
+
 GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
-        return buft->iface.get_alloc_size(buft, tensor);
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
     }
     return ggml_nbytes(tensor);
 }
@@ -55,8 +65,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
                size_t                          size) {
     ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
 
-    GGML_ASSERT(iface.get_base != NULL);
-
     (*buffer) = (struct ggml_backend_buffer) {
         /* .interface = */ iface,
         /* .buft      = */ buft,
@@ -106,6 +114,10 @@ size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
     return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
 }
 
+size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
+}
+
 size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
 }
@@ -120,6 +132,11 @@ bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
 
 void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
     buffer->usage = usage;
+
+    // FIXME: add a generic callback to the buffer interface
+    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
+        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    }
 }
 
 ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
@@ -169,6 +186,10 @@ size_t ggml_backend_get_alignment(ggml_backend_t backend) {
     return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
 }
 
+size_t ggml_backend_get_max_size(ggml_backend_t backend) {
+    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
+}
+
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
@@ -337,11 +358,26 @@ GGML_CALL static void ggml_backend_registry_init(void) {
     ggml_backend_cuda_reg_devices();
 #endif
 
+#ifdef GGML_USE_SYCL
+    extern void ggml_backend_sycl_reg_devices(void);
+    ggml_backend_sycl_reg_devices();
+#endif
+
 #ifdef GGML_USE_METAL
     extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
     extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
     ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
 #endif
+
+#ifdef GGML_USE_VULKAN
+    extern GGML_CALL int ggml_backend_vk_reg_devices(void);
+    ggml_backend_vk_reg_devices();
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+    extern GGML_CALL void ggml_backend_kompute_reg_devices(void);
+    ggml_backend_kompute_reg_devices();
+#endif
 }
 
 GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
@@ -545,6 +581,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
             /* .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,
@@ -600,6 +637,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
             /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
             /* .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,
@@ -756,6 +794,80 @@ GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, v
     GGML_UNUSED(user_data);
 }
 
+// multi-buffer buffer
+
+struct ggml_backend_multi_buffer_context {
+    ggml_backend_buffer_t * buffers;
+    size_t n_buffers;
+};
+
+typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t;
+
+GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_free(ctx->buffers[i]);
+    }
+
+    free(ctx->buffers);
+    free(ctx);
+}
+
+GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_clear(ctx->buffers[i], value);
+    }
+}
+
+static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) {
+    static struct ggml_backend_buffer_i multi_backend_buffer_i = {
+        /* .get_name        = */ ggml_backend_multi_buffer_get_name,
+        /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+        /* .get_base        = */ NULL,
+        /* .init_tensor     = */ NULL,
+        /* .set_tensor      = */ NULL,
+        /* .get_tensor      = */ NULL,
+        /* .cpy_tensor      = */ NULL,
+        /* .clear           = */ ggml_backend_multi_buffer_clear,
+        /* .reset           = */ NULL,
+    };
+
+    return multi_backend_buffer_i;
+}
+
+GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+    ctx->n_buffers = n_buffers;
+    ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
+
+    size_t total_size = 0;
+    for (size_t i = 0; i < n_buffers; i++) {
+        ctx->buffers[i] = buffers[i];
+        total_size += ggml_backend_buffer_get_size(buffers[i]);
+    }
+
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size);
+}
+
+GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+}
+
+GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
+    ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
+    }
+}
+
 
 // scheduler
 

ggml-backend.h

@@ -20,6 +20,7 @@ extern "C" {
     GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
     GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
     GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
+    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
     GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
     GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
     GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
@@ -36,6 +37,7 @@ extern "C" {
     GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
     GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
     GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
     GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
@@ -54,6 +56,7 @@ extern "C" {
     GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
     GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
 
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

ggml-cuda.cu

@@ -191,6 +191,10 @@ static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
 #endif // __has_builtin(__builtin_elementwise_sub_sat)
 }
 
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
 static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
@@ -505,6 +509,14 @@ typedef struct {
 } block_iq2_xs;
 static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
 
+#define QR3_XXS 8
+#define QI3_XXS (QK_K / (4*QR3_XXS))
+typedef struct {
+    half d;
+    uint8_t qs[3*(QK_K/8)];
+} block_iq3_xxs;
+static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
+
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -1613,6 +1625,41 @@ static const __device__ uint64_t iq2xs_grid[512] = {
     0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
 };
 
+static const __device__ uint32_t iq3xxs_grid[256] = {
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
+};
+
 static const __device__ uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
     144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
@@ -1624,6 +1671,43 @@ static const __device__ uint8_t ksigns_iq2xs[128] = {
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 };
 
+//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+static const __device__ uint64_t ksigns64[128] = {
+    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
+    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
+    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
+    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
+    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
+    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
+    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
+    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
+    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
+    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
+    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
+    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
+    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
+    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
+    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
+    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
+    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
+    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
+    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
+    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
+    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
+    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
+    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
+    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
+    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
+    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
+    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
+    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
+    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
+    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
+    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
+    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
+};
+//#endif
+
 static const __device__ uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
 
 inline bool ggml_cuda_supports_mmq(enum ggml_type type) {
@@ -1690,6 +1774,34 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
 
 }
 
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        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
+    assert(false);
+#endif
+
+}
+
 static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@@ -4313,6 +4425,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
 
 static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
 #if QK_K == 256
     const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
 
@@ -4323,20 +4436,22 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const uint8_t ls2 = bq2->scales[ib32] >>  4;
     int sumi1 = 0;
     for (int l = 0; l < 2; ++l) {
-        const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-        const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-        for (int j = 0; j < 8; ++j) {
-            sumi1 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-        }
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
+        sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
         q8 += 8;
     }
     int sumi2 = 0;
     for (int l = 2; l < 4; ++l) {
-        const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-        const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-        for (int j = 0; j < 8; ++j) {
-            sumi2 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-        }
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
+        sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
         q8 += 8;
     }
     const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
@@ -4345,6 +4460,45 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     assert(false);
     return 0.f;
 #endif
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * q3 = bq2->qs + 8*ib32;
+    const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = gas[0] | (gas[1] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0];
+        const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1];
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127));
+        const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]);
+        sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
+        sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
+    return d * sumi;
+#else
+    assert(false);
+    return 0.f;
+#endif
+#else
+    assert(false);
+    return 0.f;
+#endif
 }
 
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
@@ -6381,6 +6535,12 @@ static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k,
     dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
 }
 
+template<typename dst_t>
+static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@@ -6418,6 +6578,8 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_iq2_xxs_cuda;
         case GGML_TYPE_IQ2_XS:
             return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
         default:
@@ -6451,6 +6613,8 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_iq2_xxs_cuda;
         case GGML_TYPE_IQ2_XS:
             return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half>;
         default:
@@ -6663,6 +6827,15 @@ static void mul_mat_vec_iq2_xs_q8_1_cuda(const void * vx, const void * vy, float
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
+static void mul_mat_vec_iq3_xxs_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    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);
+    mul_mat_vec_q<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
 static void ggml_mul_mat_q4_0_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
@@ -8213,6 +8386,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -8235,6 +8409,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -8306,6 +8481,9 @@ static void ggml_cuda_op_mul_mat_vec_q(
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
             break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -9790,8 +9968,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s
     // TODO: mmq/mmv support
 #endif
 
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb1  =  dst->nb[1];
+    const size_t nb11 = src1->nb[1];
+    const size_t nb1  =  dst->nb[1];
 
     const struct ggml_tensor * ids = src0;
     const int32_t id = ((int32_t *) dst->op_params)[0];
@@ -10304,15 +10482,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t
 
     if (ggml_is_quantized(tensor->type)) {
         // initialize padding to 0 to avoid possible NaN values
-        int64_t row_low = 0;
-        int64_t row_high = ggml_nrows(tensor);
-        int64_t nrows_split = row_high - row_low;
-
-        size_t original_size = ggml_nbytes_split(tensor, nrows_split);
+        size_t original_size = ggml_nbytes(tensor);
         size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
 
         if (padded_size > original_size && tensor->view_src == nullptr) {
-            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
+            CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
         }
     }
 }
@@ -10415,12 +10589,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend
 }
 
 GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
-    int64_t row_low = 0;
-    int64_t row_high = ggml_nrows(tensor);
-    int64_t nrows_split = row_high - row_low;
-
-    size_t size = ggml_nbytes_split(tensor, nrows_split);
-
+    size_t size = ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
     if (ggml_is_quantized(tensor->type)) {
@@ -10449,6 +10618,7 @@ 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,
@@ -10724,6 +10894,7 @@ static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface
     /* .get_name         = */ ggml_backend_cuda_split_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
     /* .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,
@@ -10803,6 +10974,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
             /* .get_name         = */ ggml_backend_cuda_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
@@ -10940,7 +11112,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                     return false;
                 }
                 ggml_type a_type = a->type;
-                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS) {
+                if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }

ggml-kompute.h

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

ggml-metal.m

@@ -24,19 +24,7 @@
 
 #define UNUSED(x) (void)(x)
 
-#define GGML_METAL_MAX_KERNELS 256
-
-struct ggml_metal_buffer {
-    const char * name;
-
-    void   * data;
-    size_t   size;
-
-    id<MTLBuffer> metal;
-};
-
 struct ggml_metal_kernel {
-    id<MTLFunction>             function;
     id<MTLComputePipelineState> pipeline;
 };
 
@@ -72,6 +60,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -93,6 +82,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
   //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@@ -110,6 +100,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@@ -124,6 +115,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@@ -138,6 +130,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
     GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@@ -168,14 +161,10 @@ struct ggml_metal_context {
 
     id<MTLDevice>       device;
     id<MTLCommandQueue> queue;
-    id<MTLLibrary>      library;
 
     dispatch_queue_t d_queue;
 
-    int n_buffers;
-    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-
-    struct ggml_metal_kernel kernels[GGML_METAL_MAX_KERNELS];
+    struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT];
 
     bool support_simdgroup_reduction;
     bool support_simdgroup_mm;
@@ -242,26 +231,24 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     // Show all the Metal device instances in the system
     NSArray * devices = MTLCopyAllDevices();
     for (id<MTLDevice> device in devices) {
-        NSString * s = [device name];
-        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
+        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
     }
     [devices release]; // since it was created by a *Copy* C method
 #endif
 
     // Pick and show default Metal device
     id<MTLDevice> device = MTLCreateSystemDefaultDevice();
-    NSString * s = [device name];
-    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
+    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
     struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
-    ctx->n_buffers = 0;
-
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
+    id<MTLLibrary> metal_library;
+
     // load library
     {
         NSBundle * bundle = nil;
@@ -276,7 +263,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:libPath];
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
-            ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
+            metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
@@ -318,7 +305,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
                 //[options setFastMathEnabled:false];
 
-                ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+                metal_library = [ctx->device newLibraryWithSource:src options:options error:&error];
                 if (error) {
                     GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                     return NULL;
@@ -383,8 +370,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     {
         NSError * error = nil;
 
-        for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) {
-            ctx->kernels[i].function = nil;
+        for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
             ctx->kernels[i].pipeline = nil;
         }
 
@@ -396,10 +382,12 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #define GGML_METAL_ADD_KERNEL(e, name, supported) \
         if (supported) { \
             struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
-            kernel->function = [ctx->library newFunctionWithName:@"kernel_"#name]; \
-            kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:kernel->function error:&error]; \
+            id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
+            kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
+            [metal_function release]; \
             if (error) { \
                 GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+                [metal_library release]; \
                 return NULL; \
             } \
         } else { \
@@ -439,6 +427,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,       true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,       true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,               ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,             ctx->support_simdgroup_reduction);
@@ -460,6 +449,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,        ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
       //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,      ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,      ctx->support_simdgroup_reduction);
@@ -477,6 +467,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,     ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,  ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,  ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,         ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,        ctx->support_simdgroup_mm);
@@ -491,6 +482,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,        ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,     ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,     ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,      ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,     ctx->support_simdgroup_mm);
@@ -505,6 +497,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,     ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,  ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,   ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,  ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,              true);
@@ -528,27 +521,17 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,               true);
     }
 
+    [metal_library release];
     return ctx;
 }
 
 static void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
 
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        [ctx->buffers[i].metal release];
+    for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
+        [ctx->kernels[i].pipeline release];
     }
 
-    for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) {
-        if (ctx->kernels[i].pipeline) {
-            [ctx->kernels[i].pipeline release];
-        }
-
-        if (ctx->kernels[i].function) {
-            [ctx->kernels[i].function release];
-        }
-    }
-
-    [ctx->library release];
     [ctx->queue release];
     [ctx->device release];
 
@@ -580,51 +563,30 @@ struct ggml_backend_metal_buffer_context {
 // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
 // Metal buffer based on the host memory pointer
 //
-static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
     //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
 
     const int64_t tsize = ggml_nbytes(t);
 
     ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
 
-    // compatibility with ggml-backend
-    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
-
-        // find the view that contains the tensor fully
-        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
-            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
-
-            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
-            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
-                *offs = (size_t) ioffs;
-
-                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
-
-                return buf_ctx->buffers[i].metal;
-            }
-        }
-
-        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
-
-        return nil;
-    }
+    struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
 
     // find the view that contains the tensor fully
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+    for (int i = 0; i < buf_ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
 
-        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
-        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
+        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+        if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
             *offs = (size_t) ioffs;
 
-            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+            //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
 
-            return ctx->buffers[i].metal;
+            return buf_ctx->buffers[i].metal;
         }
     }
 
-    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
+    GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
 
     return nil;
 }
@@ -817,9 +779,9 @@ static bool ggml_metal_graph_compute(
             const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
             const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
 
-            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
-            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
-            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+            id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
+            id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
+            id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
 
             //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
             //if (src0) {
@@ -1308,6 +1270,7 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
+                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                             }
 
@@ -1436,6 +1399,12 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
                                     } break;
+                                case GGML_TYPE_IQ3_XXS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
+                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1478,6 +1447,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
+                            else if (src0t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
+                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
                             else if (src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1572,6 +1546,7 @@ static bool ggml_metal_graph_compute(
                                 case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32   ].pipeline; break;
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
+                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                             }
 
@@ -1601,7 +1576,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
                             }
@@ -1703,6 +1678,12 @@ static bool ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline;
                                     } break;
+                                case GGML_TYPE_IQ3_XXS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
+                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1746,7 +1727,7 @@ static bool ggml_metal_graph_compute(
                                 struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
 
                                 size_t offs_src_cur = 0;
-                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
+                                id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur);
 
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
                             }
@@ -1761,6 +1742,11 @@ static bool ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
+                            else if (src2t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
+                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
                             else if (src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1801,6 +1787,7 @@ static bool ggml_metal_graph_compute(
                             case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K   ].pipeline; break;
                             case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
                             case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
+                            case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
                             default: GGML_ASSERT(false && "not implemented");
                         }
@@ -2423,6 +2410,16 @@ GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backen
     UNUSED(buft);
 }
 
+GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+    size_t max_size = device.maxBufferLength;
+    ggml_backend_metal_free_device();
+
+    return max_size;
+
+    UNUSED(buft);
+}
+
 GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
     return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
 
@@ -2441,6 +2438,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
             /* .get_name         = */ ggml_backend_metal_buffer_type_get_name,
             /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
+            /* .get_max_size     = */ ggml_backend_metal_buffer_type_get_max_size,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
             /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,

ggml-metal.metal

@@ -2459,6 +2459,12 @@ typedef struct {
 } block_iq2_xs;
 // 74 bytes / block for QK_K = 256, so 2.3125 bpw
 
+typedef struct {
+    half d;
+    uint8_t qs[3*QK_K/8];
+} block_iq3_xxs;
+// 98 bytes / block for QK_K = 256, so 3.0625 bpw
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3681,6 +3687,42 @@ constexpr constant static uint64_t iq2xs_grid[512] = {
     0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
 };
 
+constexpr constant static uint32_t iq3xxs_grid[256] = {
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3c, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043c1c, 0x04043c2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3c04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3c,
+    0x04142c0c, 0x04142c3c, 0x04143c2c, 0x041c040c, 0x041c043c, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3c04, 0x04240c1c, 0x04241c3c, 0x04242424, 0x04242c3c, 0x04243c1c, 0x04243c2c, 0x042c040c,
+    0x042c043c, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043c0c04, 0x043c0c24, 0x043c0c34,
+    0x043c241c, 0x043c340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243c, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143c14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3c0c, 0x0c34042c, 0x0c3c1414,
+    0x0c3c2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343c, 0x140c3c04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3c,
+    0x14141404, 0x14141414, 0x14141c3c, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3c04, 0x141c3c24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143c, 0x142c240c, 0x142c3c24,
+    0x143c040c, 0x143c041c, 0x143c0c34, 0x143c242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043c14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143c14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243c, 0x1c243c14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3c1c1c, 0x1c3c3404, 0x24040424, 0x24040c3c,
+    0x24041c2c, 0x24041c3c, 0x24042c1c, 0x24042c3c, 0x240c3c24, 0x24141404, 0x24141c3c, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043c, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3c04, 0x243c042c, 0x243c0c04, 0x243c0c14, 0x243c1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043c04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143c14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143c, 0x2c243c14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3c1424, 0x2c3c2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3c, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343c041c, 0x343c140c, 0x3c04041c, 0x3c04042c, 0x3c04043c, 0x3c040c04, 0x3c041c14,
+    0x3c042c14, 0x3c0c1434, 0x3c0c2404, 0x3c140c14, 0x3c14242c, 0x3c142c14, 0x3c1c0404, 0x3c1c0c2c,
+    0x3c1c1c1c, 0x3c1c3404, 0x3c24140c, 0x3c24240c, 0x3c2c0404, 0x3c2c0414, 0x3c2c1424, 0x3c341c04,
+};
+
+
 constexpr constant static uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
     144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
@@ -3970,6 +4012,143 @@ kernel void kernel_mul_mv_iq2_xs_f32(
     kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
 }
 
+void kernel_mul_mv_iq3_xxs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq3_xxs * x = (device const block_iq3_xxs *) src0 + ib_row + offset0;
+    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    {
+        int nval = 4;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+#if QK_K == 256
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq3_xxs * xr = x + ibl;
+        device const uint8_t  * q3 = xr->qs + 8 * ib;
+        device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const uint32_t aux32 = gas[0] | (gas[1] << 16);
+            const float d = db * (0.5f + (aux32 >> 28));
+
+            float2 sum = {0};
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
+                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d * (sum[0] + sum[1]);
+
+            dh  += nb*sizeof(block_iq3_xxs)/2;
+            q3  += nb*sizeof(block_iq3_xxs);
+            gas += nb*sizeof(block_iq3_xxs)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+#else
+    // TODO
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
+kernel void kernel_mul_mv_iq3_xxs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+
 //============================= templates and their specializations =============================
 
 // NOTE: this is not dequantizing - we are simply fitting the template
@@ -4287,6 +4466,33 @@ void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4
     }
 }
 
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -4828,6 +5034,7 @@ template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
 
 //
 // matrix-matrix multiplication
@@ -4866,6 +5073,7 @@ template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
 
 //
 // indirect matrix-matrix multiplication
@@ -4916,6 +5124,7 @@ template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
 
 //
 // matrix-vector multiplication
@@ -5818,3 +6027,68 @@ kernel void kernel_mul_mv_id_iq2_xs_f32(
         tiisg,
         sgitg);
 }
+
+[[host_name("kernel_mul_mv_id_iq3_xxs_f32")]]
+kernel void kernel_mul_mv_id_iq3_xxs_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq3_xxs_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}

ggml-opencl.cpp

@@ -714,7 +714,6 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx,
         dst[row] = tmp[0];
     }
 }
-
 );
 
 
@@ -784,6 +783,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float
         dst[row] = tmp[0];
     }
 }
+
 );
 
 
@@ -799,6 +799,18 @@ __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y
 }
 );
 
+std::string add_template = MULTILINE_QUOTE(
+__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) {
+    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
+
+    if (i >= get_global_size(0)) {
+        return;
+    }
+
+    dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky];
+}
+);
+
 #define CL_CHECK(err)                                               \
     do {                                                            \
         cl_int err_ = (err);                                        \
@@ -878,6 +890,7 @@ static std::string generate_kernels() {
         }
         src << mul_kernel << '\n';
     }
+    src << add_template << '\n';
 
     return src.str();
 }
@@ -893,6 +906,7 @@ static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl,
 static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
 static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
 static cl_kernel mul_f32_cl;
+static cl_kernel add_f32_cl;
 static bool fp16_support;
 
 static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
@@ -1100,9 +1114,10 @@ void ggml_cl_init(void) {
     char *ext_buffer = (char *)alloca(ext_str_size + 1);
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
     ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
+    // Disabled due to faulty outputs
     // Check if ext_buffer contains cl_khr_fp16
-    fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
-    fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
+    fp16_support = false;  // strstr(ext_buffer, "cl_khr_fp16") != NULL;
+    // fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
 
     cl_context_properties properties[] = {
         (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
@@ -1150,6 +1165,8 @@ void ggml_cl_init(void) {
 
     // mul kernel
     CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
+
+    CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err));
 }
 
 static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
@@ -1458,6 +1475,70 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
     ggml_cl_mul_f32(src0, src1, dst);
 }
 
+static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+    size_t x_size;
+    size_t d_size;
+
+    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
+    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
+    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
+
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            cl_event ev;
+
+            // copy src0 to device
+            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
+
+            const int64_t i13 = i03%ne13;
+            const int64_t i12 = i02%ne12;
+            const int i1 = i13*ne12*ne11 + i12*ne11;
+
+            cl_int x_offset = 0;
+            cl_int y_offset = i1*ne10;
+            cl_int d_offset = 0;
+
+            size_t global = ne00 * ne01;
+            cl_int ky = ne10 * ne11;
+
+            CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset));
+            CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky));
+            CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
+
+            CL_CHECK(clReleaseEvent(ev));
+            CL_CHECK(clFinish(queue));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
+        }
+    }
+    ggml_cl_pool_free(d_X, x_size);
+    ggml_cl_pool_free(d_D, d_size);
+}
+
+void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cl_add_f32(src0, src1, dst);
+}
+
 static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -2044,6 +2125,15 @@ static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_
     GGML_UNUSED(buffer_type);
 }
 
+static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
+    static size_t max_size = -1;
+    if (max_size == (size_t)-1) {
+        ggml_cl_init();
+        clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_size, NULL);
+    }
+    return max_size;
+}
+
 static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
     //return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
     return ggml_backend_is_cpu(backend);
@@ -2055,6 +2145,7 @@ static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
     /* .get_name         = */ ggml_backend_opencl_buffer_type_name,
     /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
     /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_opencl_buffer_type_get_max_size,
     /* .get_alloc_size   = */ NULL,
     /* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
     /* .is_host          = */ NULL,
@@ -2111,6 +2202,7 @@ ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
             /* .get_name         = */ ggml_backend_opencl_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,

ggml-opencl.h

@@ -10,6 +10,7 @@ extern "C" {
 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-quants.c

@@ -3441,6 +3441,41 @@ static const uint64_t iq2xs_grid[512] = {
     0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
 };
 
+static const uint32_t iq3xxs_grid[256] = {
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
+};
+
 static const uint8_t ksigns_iq2xs[128] = {
       0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
     144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
@@ -3507,6 +3542,38 @@ void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y,
     }
 }
 
+// ====================== 3.0625 bpw (de)-quantization
+
+void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    uint32_t aux32;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * scales_and_signs = qs + QK_K/4;
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, scales_and_signs + 4*ib32, sizeof(uint32_t));
+            const float db = d * (0.5f + (aux32 >> 28)) * 0.5f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + qs[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + qs[2*l+1]);
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qs += 8;
+        }
+    }
+}
+
 //===================================== Q8_K ==============================================
 
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) {
@@ -8458,17 +8525,36 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
 
     const __m128i m4 = _mm_set1_epi8(0xf);
     const __m128i m1 = _mm_set1_epi8(1);
-    const __m128i m511 = _mm_set1_epi16(511);
-    const __m128i m127 = _mm_set1_epi16(127);
+    const __m256i m511 = _mm256_set1_epi16(511);
+    const __m256i mone = _mm256_set1_epi8(1);
 
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
+    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
+    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
+    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
 
     uint64_t aux64;
 
     // somewhat hacky, but gives a significant boost in performance
-    __m128i aux_gindex, aux_sindex;
+    __m256i aux_gindex;
     const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-    const uint16_t * sindex = (const uint16_t *)&aux_sindex;
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
@@ -8483,26 +8569,68 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
 
         __m256i sumi1 = _mm256_setzero_si256();
         __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
+            aux_gindex = _mm256_and_si256(q2_data, m511);
+
+            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
+            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
+
             const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
             const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m128i q2_data = _mm_loadu_si128((const __m128i*)q2);  q2 +=  8;
-            aux_gindex = _mm_and_si128(q2_data, m511);
-            aux_sindex = _mm_and_si128(_mm_srli_epi16(q2_data, 9), m127);
-            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]], iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]], iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
-            const __m256i s2_1 = _mm256_set_epi64x(signs64[sindex[3]], signs64[sindex[2]], signs64[sindex[1]], signs64[sindex[0]]);
-            const __m256i s2_2 = _mm256_set_epi64x(signs64[sindex[7]], signs64[sindex[6]], signs64[sindex[5]], signs64[sindex[4]]);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
+            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
+            const __m256i full_signs_1 = _mm256_set_m128i(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = _mm256_set_m128i(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
+
             const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
             const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
+            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
 
             const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
             const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
 
             sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
             sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
         }
 
         accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
@@ -8551,6 +8679,136 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
 #endif
 }
 
+// TODO
+void ggml_vec_dot_iq3_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_iq3_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
+            const uint32x4_t aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
+            const uint32x4_t aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
+            const uint32x4_t aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
+            const uint32x4_t aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            q3 += 16;
+            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
+            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
+            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
+            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.5f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
 // ================================ IQ2 quantization =============================================
 
 typedef struct {
@@ -9189,3 +9447,436 @@ size_t quantize_iq2_xs(const float * src, void * dst, int nrow, int n_per_row, i
     return nrow * nblock * sizeof(block_iq2_xs);
 }
 
+//
+// ============================================= 3-bit using D4 lattice
+//
+
+typedef struct {
+    uint32_t * grid;
+    int      * map;
+    uint16_t * neighbours;
+} iq3_entry_t;
+
+static iq3_entry_t iq3_data[1] = {
+    {NULL, NULL, NULL},
+};
+
+static inline int iq3_data_index(int grid_size) {
+    (void)grid_size;
+    GGML_ASSERT(grid_size == 256);
+    return 0;
+}
+
+static int iq3_compare_func(const void * left, const void * right) {
+    const int * l = (const int *)left;
+    const int * r = (const int *)right;
+    return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+void iq3xs_init_impl(int grid_size) {
+    const int gindex = iq3_data_index(grid_size);
+    if (iq3_data[gindex].grid) {
+        return;
+    }
+    static const uint16_t kgrid_256[256] = {
+            0,     2,     4,     9,    11,    15,    16,    18,    25,    34,    59,    61,    65,    67,    72,    74,
+           81,    85,    88,    90,    97,   108,   120,   128,   130,   132,   137,   144,   146,   153,   155,   159,
+          169,   175,   189,   193,   199,   200,   202,   213,   248,   267,   287,   292,   303,   315,   317,   321,
+          327,   346,   362,   413,   436,   456,   460,   462,   483,   497,   513,   515,   520,   522,   529,   531,
+          536,   538,   540,   551,   552,   576,   578,   585,   592,   594,   641,   643,   648,   650,   657,   664,
+          698,   704,   706,   720,   729,   742,   758,   769,   773,   808,   848,   852,   870,   889,   901,   978,
+          992,  1024,  1026,  1033,  1035,  1040,  1042,  1046,  1049,  1058,  1089,  1091,  1093,  1096,  1098,  1105,
+         1112,  1139,  1143,  1144,  1152,  1154,  1161,  1167,  1168,  1170,  1183,  1184,  1197,  1217,  1224,  1228,
+         1272,  1276,  1309,  1323,  1347,  1367,  1377,  1404,  1473,  1475,  1486,  1509,  1537,  1544,  1546,  1553,
+         1555,  1576,  1589,  1594,  1600,  1602,  1616,  1625,  1636,  1638,  1665,  1667,  1672,  1685,  1706,  1722,
+         1737,  1755,  1816,  1831,  1850,  1856,  1862,  1874,  1901,  1932,  1950,  1971,  2011,  2032,  2052,  2063,
+         2077,  2079,  2091,  2095,  2172,  2192,  2207,  2208,  2224,  2230,  2247,  2277,  2308,  2345,  2356,  2389,
+         2403,  2424,  2501,  2504,  2506,  2520,  2570,  2593,  2616,  2624,  2630,  2646,  2669,  2700,  2714,  2746,
+         2754,  2795,  2824,  2835,  2839,  2874,  2882,  2905,  2984,  3028,  3042,  3092,  3108,  3110,  3124,  3153,
+         3185,  3215,  3252,  3288,  3294,  3364,  3397,  3434,  3483,  3523,  3537,  3587,  3589,  3591,  3592,  3610,
+         3626,  3670,  3680,  3722,  3749,  3754,  3776,  3789,  3803,  3824,  3857,  3873,  3904,  3906,  3924,  3992,
+    };
+    const int kmap_size = 4096;
+    const int nwant = 2;
+    const uint16_t * kgrid = kgrid_256;
+    uint32_t * kgrid_q3xs;
+    int      * kmap_q3xs;
+    uint16_t * kneighbors_q3xs;
+
+    printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+    uint32_t * the_grid = (uint32_t *)malloc(grid_size*sizeof(uint32_t));
+    for (int k = 0; k < grid_size; ++k) {
+        int8_t * pos = (int8_t *)(the_grid + k);
+        for (int i = 0; i < 4; ++i) {
+            int l = (kgrid[k] >> 3*i) & 0x7;
+            pos[i] = 2*l + 1;
+        }
+    }
+    kgrid_q3xs = the_grid;
+    iq3_data[gindex].grid = the_grid;
+    kmap_q3xs = (int *)malloc(kmap_size*sizeof(int));
+    iq3_data[gindex].map = kmap_q3xs;
+    for (int i = 0; i < kmap_size; ++i) kmap_q3xs[i] = -1;
+    uint32_t aux32;
+    uint8_t * aux8 = (uint8_t *)&aux32;
+    for (int i = 0; i < grid_size; ++i) {
+        aux32 = kgrid_q3xs[i];
+        uint16_t index = 0;
+        for (int k=0; k<4; ++k) {
+            uint16_t q = (aux8[k] - 1)/2;
+            index |= (q << 3*k);
+        }
+        kmap_q3xs[index] = i;
+    }
+    int8_t pos[4];
+    int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+    int num_neighbors = 0, num_not_in_map = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q3xs[i] >= 0) continue;
+        ++num_not_in_map;
+        for (int k = 0; k < 4; ++k) {
+            int l = (i >> 3*k) & 0x7;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+        int n = 0; int d2 = dist2[0];
+        int nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            ++n;
+        }
+        num_neighbors += n;
+    }
+    printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+    kneighbors_q3xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+    iq3_data[gindex].neighbours = kneighbors_q3xs;
+    int counter = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q3xs[i] >= 0) continue;
+        for (int k = 0; k < 4; ++k) {
+            int l = (i >> 3*k) & 0x7;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+        kmap_q3xs[i] = -(counter + 1);
+        int d2 = dist2[0];
+        uint16_t * start = &kneighbors_q3xs[counter++];
+        int n = 0, nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            kneighbors_q3xs[counter++] = dist2[2*j+1];
+            ++n;
+        }
+        *start = n;
+    }
+    free(dist2);
+}
+
+void iq3xs_free_impl(int grid_size) {
+    GGML_ASSERT(grid_size == 256);
+    const int gindex = iq3_data_index(grid_size);
+    if (iq3_data[gindex].grid) {
+        free(iq3_data[gindex].grid);       iq3_data[gindex].grid = NULL;
+        free(iq3_data[gindex].map);        iq3_data[gindex].map  = NULL;
+        free(iq3_data[gindex].neighbours); iq3_data[gindex].neighbours = NULL;
+    }
+}
+
+static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const uint32_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_d2 = FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float d2 = 0;
+        for (int i = 0; i < 4; ++i) {
+            float q = pg[i];
+            float diff = scale*q - xval[i];
+            d2 += weight[i]*diff*diff;
+        }
+        if (d2 < best_d2) {
+            best_d2 = d2; grid_index = neighbours[j];
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 4; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static void quantize_row_iq3_xxs_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
+
+    const int gindex = iq3_data_index(256);
+
+    const uint32_t * kgrid_q3xs      = iq3_data[gindex].grid;
+    const int      * kmap_q3xs       = iq3_data[gindex].map;
+    const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours;
+
+    //GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q3xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q3xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 8;
+
+    const int nbl = n/256;
+
+    block_iq3_xxs * y = vy;
+
+    float scales[QK_K/32];
+    float weight[32];
+    float xval[32];
+    int8_t L[32];
+    int8_t Laux[32];
+    float  waux[32];
+    bool   is_on_grid[8];
+    bool   is_on_grid_aux[8];
+    uint8_t block_signs[8];
+    uint8_t q3[3*(QK_K/8)];
+    uint32_t * scales_and_signs = (uint32_t *)(q3 + QK_K/4);
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(q3, 0, 3*QK_K/8);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float * xb = xbl + 32*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + 32*ib;
+                for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < 32; ++i) weight[i] = xb[i]*xb[i];
+            }
+            for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 4; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+            if (!max) {
+                scales[ib] = 0;
+                memset(L, 0, 32);
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            for (int is = -15; is <= 15; ++is) {
+                float id = (2*kMaxQ-1+is*0.2f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 8; ++k) {
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i);
+                    int grid_index = kmap_q3xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 32; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  8; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 8; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 8; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 3*i);
+                    }
+                    int grid_index = kmap_q3xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k);
+                    }
+                    const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index);
+                    for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2;
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+                // and correspondingly flip quant signs.
+                scale = -scale;
+                for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 8; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i);
+                int grid_index = kmap_q3xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
+                    printf("\n");
+                    GGML_ASSERT(false);
+                }
+                q3[8*ib+k] = grid_index;
+            }
+            scales_and_signs[ib] = block_signs[0] | (block_signs[1] << 7) | (block_signs[2] << 14) | (block_signs[3] << 21);
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(y[ibl].qs, 0, 3*QK_K/8);
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d);
+        float id = 1/d;
+        float sumqx = 0, sumq2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            scales_and_signs[ib] |= ((uint32_t)l << 28);
+            if (false) {
+                const float * xb = xbl + 32*ib;
+                if (quant_weights) {
+                    const float * qw = quant_weights + QK_K*ibl + 32*ib;
+                    for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+                } else {
+                    for (int i = 0; i < 32; ++i) weight[i] = xb[i]*xb[i];
+                }
+                const float db = 0.25f * d * (1 + 2*l);
+                for (int k = 0; k < 8; ++k) {
+                    const int8_t * signs = keven_signs_q2xs + 8*((scales_and_signs[ib] >> 7*(k/2)) & 127) + 4*(k%2);
+                    const float * xk = xb + 4*k;
+                    const float * wk = weight + 4*k;
+                    //const uint8_t * grid = (const uint8_t *)(kgrid_q3xs + q3[8*ib+k]);
+                    const uint8_t * grid = (const uint8_t *)(iq3xxs_grid + q3[8*ib+k]);
+                    float best_mse = 0; int best_index = q3[8*ib+k];
+                    for (int j = 0; j < 4; ++j) {
+                        float diff = db * grid[j] * signs[j] - xk[j];
+                        best_mse += wk[j] * diff * diff;
+                    }
+                    for (int idx = 0; idx < 256; ++idx) {
+                        //grid = (const uint8_t *)(kgrid_q3xs + idx);
+                        grid = (const uint8_t *)(iq3xxs_grid + idx);
+                        float mse = 0;
+                        for (int j = 0; j < 4; ++j) {
+                            float diff = db * grid[j] * signs[j] - xk[j];
+                            mse += wk[j] * diff * diff;
+                        }
+                        if (mse < best_mse) {
+                            best_mse = mse; best_index = idx;
+                        }
+                    }
+                    q3[8*ib+k] = best_index;
+                    //grid = (const uint8_t *)(kgrid_q3xs + best_index);
+                    grid = (const uint8_t *)(iq3xxs_grid + best_index);
+                    for (int j = 0; j < 4; ++j) {
+                        float q = db * grid[j] * signs[j];
+                        sumqx += wk[j] * q * xk[j];
+                        sumq2 += wk[j] * q * q;
+                    }
+                }
+                if (sumq2 > 0) y[ibl].d = GGML_FP32_TO_FP16(d*sumqx/sumq2);
+            }
+        }
+        memcpy(y[ibl].qs, q3, 3*QK_K/8);
+    }
+}
+
+size_t quantize_iq3_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
+    (void)hist;
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int row = 0; row < nrow; ++row) {
+        quantize_row_iq3_xxs_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq3_xxs);
+    }
+    return nrow * nblock * sizeof(block_iq3_xxs);
+}
+
+void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int k) {
+    assert(k % QK_K == 0);
+    block_iq3_xxs * restrict y = vy;
+    quantize_row_iq3_xxs_reference(x, y, k);
+}
+
+void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * restrict y, int k) {
+    assert(k % QK_K == 0);
+    quantize_row_iq3_xxs_impl(x, y, k, NULL);
+}

ggml-quants.h

@@ -166,7 +166,7 @@ typedef struct {
 static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
 
 // (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml dsign, it ends up using
+// Due to the need to use blocks as per ggml design, it ends up using
 // 2.0625 bpw because of the 16-bit scale for each block of 256.
 typedef struct {
     ggml_fp16_t d;
@@ -182,6 +182,15 @@ typedef struct {
 } block_iq2_xs;
 static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
 
+// (Almost) "true" 3-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 3.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_fp16_t d;
+    uint8_t qs[3*QK_K/8];
+} block_iq3_xxs;
+static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
+
 // Quantization
 void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
 void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
@@ -196,6 +205,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
 void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
 void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * restrict y, int k);
 
 void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
 void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
@@ -210,6 +220,7 @@ void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_iq3_xxs(const float * restrict x, void * restrict y, int k);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
@@ -227,6 +238,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
 void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k);
 void dequantize_row_iq2_xs (const block_iq2_xs  * restrict x, float * restrict y, int k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int k);
 
 // Dot product
 void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
@@ -242,12 +254,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx,
 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_iq2_xs_q8_K (int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
 //
 size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 size_t quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
@@ -260,3 +274,5 @@ size_t quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row,
 
 void iq2xs_init_impl(int grid_size);
 void iq2xs_free_impl(int grid_size);
+void iq3xs_init_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);

ggml-sycl.h

@@ -0,0 +1,27 @@
+/*MIT license
+  Copyright (C) 2024 Intel Corporation
+  SPDX-License-Identifier: MIT
+*/
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_SYCL_MAX_DEVICES       16
+#define GGML_SYCL_NAME "SYCL"
+
+GGML_API void   ggml_init_sycl(void);
+GGML_API bool   ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-vulkan.h

@@ -0,0 +1,34 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_VK_NAME "Vulkan"
+
+GGML_API void ggml_vk_init(void);
+
+GGML_API void ggml_vk_preallocate_buffers_graph(struct ggml_tensor * node);
+GGML_API void ggml_vk_preallocate_buffers(void);
+GGML_API void ggml_vk_build_graph(struct ggml_tensor * node, bool last_node);
+GGML_API bool ggml_vk_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#ifdef GGML_VULKAN_CHECK_RESULTS
+void ggml_vk_check_results_1(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#endif
+GGML_API void ggml_vk_graph_cleanup(void);
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(void);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml.c

@@ -248,6 +248,10 @@ inline static void * ggml_aligned_malloc(size_t size) {
 #include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
+#elif defined(GGML_USE_VULKAN)
+#include "ggml-vulkan.h"
+#elif defined(GGML_USE_SYCL)
+#include "ggml-sycl.h"
 #endif
 
 // floating point type used to accumulate sums
@@ -595,6 +599,17 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
     },
+    [GGML_TYPE_IQ3_XXS] = {
+        .type_name                = "iq3_xxs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq3_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_xxs,
+        .from_float               = quantize_row_iq3_xxs,
+        .from_float_reference     = (ggml_from_float_t)quantize_row_iq3_xxs_reference,
+        .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+    },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
         .blck_size                = QK_K,
@@ -2140,6 +2155,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
         case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -2293,6 +2309,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         ggml_init_cublas();
 #elif defined(GGML_USE_CLBLAST)
         ggml_cl_init();
+#elif defined(GGML_USE_VULKAN)
+        ggml_vk_init();
+#elif defined(GGML_USE_SYCL)
+        ggml_init_sycl();
 #endif
 
         ggml_setup_op_has_task_pass();
@@ -7207,6 +7227,17 @@ static void ggml_compute_forward_add_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
+#ifdef GGML_USE_CLBLAST
+    if (src1->backend == GGML_BACKEND_GPU) {
+        // TODO: OpenCL kernel support full broadcast
+        GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
+        if (ith == 0) {
+            ggml_cl_add(src0, src1, dst);
+        }
+        return;
+    }
+#endif
+
     const int nr  = ggml_nrows(src0);
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -7487,7 +7518,12 @@ static void ggml_compute_forward_add(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add_f32(params, src0, src1, dst);
+                if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add_f32(params, src0, src1, dst);
+                }
+                else {
+                    GGML_ASSERT(false);
+                }
             } break;
         case GGML_TYPE_F16:
             {
@@ -7513,6 +7549,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             {
                 ggml_compute_forward_add_q_f32(params, src0, src1, dst);
             } break;
@@ -7779,6 +7816,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             {
                 ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
             } break;
@@ -7898,6 +7936,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
         default:
             {
                 GGML_ASSERT(false);
@@ -7999,7 +8038,7 @@ static void ggml_compute_forward_mul_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-#ifdef GGML_USE_CLBLAST
+#if defined(GGML_USE_CLBLAST)
     if (src1->backend == GGML_BACKEND_GPU) {
         // TODO: OpenCL kernel support full broadcast
         GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
@@ -9954,7 +9993,7 @@ static void ggml_compute_forward_mul_mat(
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
     if (ggml_compute_forward_mul_mat_use_blas(dst)) {
         const int64_t ne_plane      = ne01*ne00;
-        const int64_t desired_wsize = ne13*ne12*ne_plane*sizeof(float);
+        const size_t  desired_wsize = ne13*ne12*ne_plane*sizeof(float);
         UNUSED(desired_wsize);
 
         if (params->type == GGML_TASK_INIT) {
@@ -10649,6 +10688,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             {
                 ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
             } break;
@@ -10828,6 +10868,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
         default:
             {
                 GGML_ASSERT(false);
@@ -11024,6 +11065,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
             {
                 ggml_compute_forward_get_rows_q(params, src0, src1, dst);
             } break;
@@ -11671,6 +11713,7 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -11747,6 +11790,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q6_K:
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -14683,8 +14727,26 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
     GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
     GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
+#elif defined(GGML_USE_VULKAN)
+    const bool skip_cpu = ggml_vk_compute_forward(params, tensor);
+#ifdef GGML_VULKAN_CHECK_RESULTS
+    if (skip_cpu) {
+        ggml_vk_check_results_1(params, tensor);
+    }
+#endif
+    if (skip_cpu) {
+        return;
+    }
+    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
+    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
 #endif // GGML_USE_CUBLAS
 
+#ifdef GGML_USE_SYCL
+    bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
+    if (skip_cpu) {
+        return;
+    }
+#endif // GGML_USE_SYCL
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -16597,7 +16659,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+                n_tasks = MIN(n_threads, ggml_nrows(node->src[0]));
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
@@ -17079,6 +17141,17 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         }
     }
 
+#ifdef GGML_USE_VULKAN
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_vk_preallocate_buffers_graph(cgraph->nodes[i]);
+    }
+    ggml_vk_preallocate_buffers();
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_vk_build_graph(cgraph->nodes[i], i == cgraph->n_nodes - 1);
+    }
+#endif
+
     const int n_threads = cplan->n_threads;
 
     struct ggml_compute_state_shared state_shared = {
@@ -17130,6 +17203,10 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
         }
     }
 
+#ifdef GGML_USE_VULKAN
+    ggml_vk_graph_cleanup();
+#endif
+
     // performance stats (graph)
     {
         int64_t perf_cycles_cur  = ggml_perf_cycles()  - perf_start_cycles;
@@ -18770,6 +18847,7 @@ void ggml_quantize_init(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_IQ2_XXS: iq2xs_init_impl(256); break;
         case GGML_TYPE_IQ2_XS:  iq2xs_init_impl(512); break;
+        case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
         default: // nothing
             break;
     }
@@ -19032,6 +19110,15 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
                 result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
                 GGML_ASSERT(result == row_size * nrows);
             } break;
+        case GGML_TYPE_IQ3_XXS:
+            {
+                GGML_ASSERT(start % QK_K == 0);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
+            } break;
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);
@@ -20264,7 +20351,7 @@ int ggml_cpu_has_wasm_simd(void) {
 }
 
 int ggml_cpu_has_blas(void) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_VULKAN) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_SYCL)
     return 1;
 #else
     return 0;
@@ -20287,8 +20374,24 @@ int ggml_cpu_has_clblast(void) {
 #endif
 }
 
+int ggml_cpu_has_vulkan(void) {
+#if defined(GGML_USE_VULKAN)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sycl(void) {
+#if defined(GGML_USE_SYCL)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_gpublas(void) {
-    return ggml_cpu_has_cublas() || ggml_cpu_has_clblast();
+    return ggml_cpu_has_cublas() || ggml_cpu_has_clblast() || ggml_cpu_has_vulkan() || ggml_cpu_has_sycl();
 }
 
 int ggml_cpu_has_sse3(void) {

ggml.h

@@ -353,6 +353,7 @@ extern "C" {
         GGML_TYPE_Q8_K = 15,
         GGML_TYPE_IQ2_XXS = 16,
         GGML_TYPE_IQ2_XS  = 17,
+        GGML_TYPE_IQ3_XXS = 18,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -389,6 +390,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
     };
 
     // available tensor operations:
@@ -2263,9 +2265,11 @@ extern "C" {
     GGML_API int ggml_cpu_has_blas       (void);
     GGML_API int ggml_cpu_has_cublas     (void);
     GGML_API int ggml_cpu_has_clblast    (void);
+    GGML_API int ggml_cpu_has_vulkan     (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_vsx        (void);
 
     //

gguf-py/gguf/constants.py

@@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum):
     PHI2      = auto()
     PLAMO     = auto()
     CODESHELL = auto()
+    ORION     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PHI2:           "phi2",
     MODEL_ARCH.PLAMO:          "plamo",
     MODEL_ARCH.CODESHELL:      "codeshell",
+    MODEL_ARCH.ORION:          "orion",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
-    ]
+    ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.ORION: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #

gguf-py/gguf/gguf_reader.py

@@ -107,7 +107,7 @@ class GGUFReader:
         offs, tensors_fields = self._build_tensors_fields(offs, tensor_count)
         new_align = self.fields.get('general.alignment')
         if new_align is not None:
-            if new_align.types != [GGUFValueType.UINT64]:
+            if new_align.types != [GGUFValueType.UINT32]:
                 raise ValueError('Bad type for general.alignment field')
             self.alignment = new_align.parts[-1][0]
         padding = offs % self.alignment

kompute-shaders/common.comp

@@ -0,0 +1,102 @@
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#extension GL_EXT_debug_printf : enable
+
+#define QK4_0 32
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+#define TWOPI_F 6.283185307179586f
+
+#define QK_K 256
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
+    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.f * xb.d;
+    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
+    const uint16_t mask1 = mask0 << 8;
+
+    mat4 reg;
+    for (int i=0;i<8;i++) {
+        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
+        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
+        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
+    }
+    return reg;
+}
+
+#define sizeof_block_q4_1 0x14
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
+    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb.m;
+    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
+    const uint16_t mask1 = mask0 << 8;
+
+    mat4 reg;
+    for (int i=0;i<8;i++) {
+        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
+        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
+        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
+    }
+    return reg;
+}
+
+#define sizeof_block_q6_k 210
+struct block_q6_k {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;             // super-block scale
+};
+mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
+    const float16_t d_all = xb.d;
+
+    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
+    const uint qhIndex = 32*(il/8) + 16*(il&1);
+    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
+    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
+    const float16_t ml = float16_t(d_all * sc * 32.f);
+    const float16_t dl = float16_t(d_all * sc * coef);
+    mat4 reg;
+    for (int i = 0; i < 16; ++i) {
+        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
+                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
+        reg[i/4][i%4] = dl * q - ml;
+    }
+    return reg;
+}
+
+
+#define QK8_0 32
+// struct block_q8_0 {
+//     float16_t d;         // delta
+//     int8_t    qs[QK8_0]; // quants
+// };
+#define sizeof_block_q8_0 34

kompute-shaders/op_add.comp

@@ -0,0 +1,58 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1024) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb00;
+    int nb01;
+    int nb02;
+    int nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    int ne13;
+    int nb10;
+    int nb11;
+    int nb12;
+    int nb13;
+    int ne0;
+    int nb0;
+    int nb1;
+    int nb2;
+    int nb3;
+  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
+} pcs;
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint i13 = i03 % pcs.ne13;
+    const uint i12 = i02 % pcs.ne12;
+    const uint i11 = i01 % pcs.ne11;
+
+    int offs = 0; // TMP (see above)
+
+    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
+    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
+    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
+        const uint i10 = i0 % pcs.ne10;
+        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
+    }
+}

kompute-shaders/op_addrow.comp

@@ -0,0 +1,25 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
+    }
+}

kompute-shaders/op_cpy_f16_f16.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute-shaders/op_cpy_f16_f32.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute-shaders/op_cpy_f32_f16.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute-shaders/op_cpy_f32_f32.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}

kompute-shaders/op_diagmask.comp

@@ -0,0 +1,30 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint n_past;
+    int ne00;
+    int ne01;
+} pcs;
+
+void main() {
+    const uint i02 = gl_WorkGroupID.z;
+    const uint i01 = gl_WorkGroupID.y;
+    const uint i00 = gl_WorkGroupID.x;
+
+    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
+
+    if (i00 > pcs.n_past + i01) {
+        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
+    } else {
+        out_[index + pcs.outOff] = in_[index + pcs.inOff];
+    }
+}

kompute-shaders/op_gelu.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 8;
+
+    for (uint x = 0; x < 8; x++) {
+        const uint i = baseIndex + x;
+        const float y = in_[i + pcs.inOff];
+        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
+    }
+}

kompute-shaders/op_getrows.comp

@@ -0,0 +1,17 @@
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    int z = 0;
+    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
+        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
+        const mat4 result = dequantize_block(inIndex, ind%NL);
+        for (uint j = 0; j < 4; ++j) {
+            for (uint k = 0; k < 4; ++k) {
+                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
+                out_[outIndex] = result[j][k];
+                ++z;
+            }
+        }
+    }
+}

kompute-shaders/op_getrows_f16.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    for (int j = 0; j < k; j++) {
+        out_[y + j] = inA[x + j];
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
+}

kompute-shaders/op_getrows_q4_0.comp

@@ -0,0 +1,38 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_0 get_unaligned_block_q4_0(uint index) {
+    block_q4_0 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
+        fres.qs[it] = inA[index+2+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_0 block = get_unaligned_block_q4_0(index);
+    return dequantize_q4_0(block, il);
+}
+
+#include "op_getrows.comp"

kompute-shaders/op_getrows_q4_1.comp

@@ -0,0 +1,39 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_1 get_unaligned_block_q4_1(uint index) {
+    block_q4_1 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    fres.m = u8BufToFloat16(inA, index+2);
+    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
+        fres.qs[it] = inA[index+4+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_1 block = get_unaligned_block_q4_1(index);
+    return dequantize_q4_1(block, il);
+}
+
+#include "op_getrows.comp"

kompute-shaders/op_getrows_q6_k.comp

@@ -0,0 +1,44 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 16
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q6_k get_unaligned_block_q6_k(uint index) {
+    block_q6_k fres;
+    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
+        fres.ql[it] = inA[index + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
+        fres.qh[it] = inA[index + QK_K/2 + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
+        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
+    }
+    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q6_k block = get_unaligned_block_q6_k(index);
+    return dequantize_q6_k(block, il);
+}
+
+#include "op_getrows.comp"

kompute-shaders/op_mul.comp

@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1024) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb00;
+    int nb01;
+    int nb02;
+    int nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    int ne13;
+    int nb10;
+    int nb11;
+    int nb12;
+    int nb13;
+    int ne0;
+    int nb0;
+    int nb1;
+    int nb2;
+    int nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint i13 = i03 % pcs.ne13;
+    const uint i12 = i02 % pcs.ne12;
+    const uint i11 = i01 % pcs.ne11;
+
+    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
+    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
+    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
+        const uint i10 = i0 % pcs.ne10;
+        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
+    }
+}

kompute-shaders/op_mul_mat_f16.comp

@@ -0,0 +1,67 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x_id = 0) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    int ne10;
+    int ne11;
+    int ne12;
+    uint nb10;
+    uint nb11;
+    uint nb12;
+    int ne0;
+    int ne1;
+    uint r2;
+    uint r3;
+} pcs;
+
+#define N_F16_F32 4
+
+void main() {
+    const uint r0 = gl_WorkGroupID.x;
+    const uint rb = gl_WorkGroupID.y*N_F16_F32;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02;
+
+    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
+
+    for (uint row = 0; row < N_F16_F32; ++row) {
+        uint r1 = rb + row;
+        if (r1 >= pcs.ne11) {
+            break;
+        }
+
+        const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB
+
+        float sumf = 0;
+        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+            sumf += float(inA[x+i]) * float(inB[y+i]);
+        }
+
+        const float all_sum = subgroupAdd(sumf);
+        if (subgroupElect()) {
+            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
+        }
+    }
+}

kompute-shaders/op_mul_mat_mat_f32.comp

@@ -0,0 +1,51 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#extension GL_EXT_debug_printf : enable
+
+// device subgroup size
+layout (local_size_x_id = 0) in;
+
+layout(binding = 0) readonly buffer tensorInA { float inA[]; };
+layout(binding = 1) readonly buffer tensorInB { float inB[]; };
+layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout(push_constant) uniform parameter {
+  uint inAOff;
+  uint inBOff;
+  uint outOff;
+  int ne00;
+  int ne01;
+  int ne02;
+  int ne11;
+  int ne12;
+  uint nb01;
+  uint nb02;
+  uint nb11;
+  uint nb12;
+  uint nb1;
+  uint nb2;
+}
+pcs;
+
+
+void main() {
+  uvec3 gid = gl_WorkGroupID;
+
+  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
+  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
+
+  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
+  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
+  float sum = 0.0f;
+  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+      sum += float(inA[x+i]) * float(inB[y+i]);
+  }
+
+  const float all_sum = subgroupAdd(sum);
+  if (subgroupElect()) {
+    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
+  }
+}

kompute-shaders/op_mul_mat_q4_0.comp

@@ -0,0 +1,33 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_0
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_0 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (sumy * -8.f + acc[0] + acc[1]);
+}
+
+#include "op_mul_mv_q_n.comp"

kompute-shaders/op_mul_mat_q4_1.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_1
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_1 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float m = float(u8BufToFloat16(inA, index+2));
+
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
+#include "op_mul_mv_q_n.comp"

kompute-shaders/op_mul_mat_q6_k.comp

@@ -0,0 +1,94 @@
+#version 450
+
+#include "common.comp"
+
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x_id = 0) in;
+layout(local_size_y_id = 1) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int gqa;
+} pcs;
+
+void main() {
+    const uint8_t kmask1 = uint8_t(0x03);
+    const uint8_t kmask2 = uint8_t(0x0C);
+    const uint8_t kmask3 = uint8_t(0x30);
+    const uint8_t kmask4 = uint8_t(0xC0);
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint r2 = gl_WorkGroupID.z;
+
+    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
+    const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0);
+    const uint x = row * nb + offset0; // Based from inA without base offset
+    const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+
+    float sumf = 0;
+
+    // bits of invocation ID for gl_SubgroupSize=32:
+    //  x   x   x   x   x
+    //  4   3   2   1   0
+    // (     tid     ) ix
+    //  ip (   il    )
+
+    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
+    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
+    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
+    const uint ip   = tid/8;        // first or second half of block (0 or 1)
+    const uint il   = tid%8;        // each half has 8 parts, one per scale
+    const uint n    = 4;            // 4 scales at a time (and 4 sums)
+    const uint l0   = n*il;         // offset into half-block, 0..28
+    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    const uint y_offset = 128*ip + l0;
+    const uint q_offset_l = 64*ip + l0;
+    const uint q_offset_h = 32*ip + l0;
+
+    for (uint i = ix; i < nb; i += block_stride) {
+
+        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
+
+        const uint qlIndex = q_offset_l;
+        const uint q2Index = qlIndex + QK_K/8;
+        const uint qhIndex = q_offset_h;
+        const uint y = yy + i * QK_K + y_offset;
+
+        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        for (uint l = 0; l < n; ++l) {
+            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
+            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
+            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
+
+            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
+            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
+            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
+            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
+        }
+
+        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
+        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
+    }
+
+    const float tot = subgroupAdd(sumf);
+    if (subgroupElect()) {
+        out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
+    }
+}

kompute-shaders/op_mul_mat_q8_0.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#include "common.comp"
+
+#include "op_mul_mv_q_n_pre.comp"
+
+#define SIZE_OF_D 2
+
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+#define NB_Q8_0 8
+
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = pcs.ne00/QK8_0;
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
+    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
+
+    float yl[NB_Q8_0];
+    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
+
+    const uint ix = gl_SubgroupInvocationID.x/4;
+    const uint il = gl_SubgroupInvocationID.x%4;
+
+    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
+
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (uint ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = inB[yb + i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
+            float sumq = 0.f;
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
+                sumq += qs_iq * yl[iq];
+            }
+            const float16_t d = u8BufToFloat16(inA, x + block_offset);
+            sumf[row] += sumq*d;
+        }
+
+        yb += NB_Q8_0 * nw;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (subgroupElect() && first_row + row < pcs.ne01) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
+        }
+    }
+}

kompute-shaders/op_mul_mv_q_n.comp

@@ -0,0 +1,48 @@
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint x = offset0; // Based from inA without base offset
+    const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+
+    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
+
+    const uint ix = gl_SubgroupInvocationID/2;
+    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
+
+    uint yb = y + ix * BLOCKS_IN_QUANT + il;
+
+    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
+    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
+    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
+
+    for (uint ib = ix; ib < nb; ib += 16) {
+        for (int row = 0; row < N_ROWS; row++) {
+            const uint block_index = x + ib + row * nb;
+            sumf[row] += block_q_n_dot_y(block_index, yb, il);
+        }
+
+        yb += BLOCKS_IN_QUANT * 16;
+    }
+
+    for (int row = 0; row < N_ROWS; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (first_row + row < pcs.ne01 && subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
+        }
+    }
+}

kompute-shaders/op_mul_mv_q_n_pre.comp

@@ -0,0 +1,22 @@
+layout(local_size_x_id = 0) in;
+layout(local_size_y = 1) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int  ne00;
+    int  ne01;
+    int  ne02;
+    int  ne10;
+    int  ne12;
+    int  ne0;
+    int  ne1;
+    uint r2;
+    uint r3;
+} pcs;

kompute-shaders/op_norm.comp

@@ -0,0 +1,84 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 256) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+    // MEAN
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float mean = sum[0];
+
+    // recenter
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float variance = sum[0];
+
+    const float scale = 1.0f/sqrt(variance + pcs.eps);
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] *= scale;
+    }
+}

kompute-shaders/op_relu.comp

@@ -0,0 +1,21 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
+    }
+}

kompute-shaders/op_rmsnorm.comp

@@ -0,0 +1,53 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 512) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+
+    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
+
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] * scale;
+    }
+}

kompute-shaders/op_rope_f16.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    const bool is_neox = (pcs.mode & 2) != 0;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_orig_ctx, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    const int p = inB[pcs.inBOff + i2];
+
+    float theta = float(p);
+
+    if (!is_neox) {
+        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+1]);
+
+            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
+        }
+    } else {
+        const float inv_ndims = -1.f/pcs.n_dims;
+        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
+            const uint cur_rot = ic;
+
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint i0 = ic/2;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+pcs.n_dims/2]);
+
+            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
+        }
+
+        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
+            const uint i0 = ic;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data + 0] = inA[src + 0];
+            out_[dst_data + 1] = inA[src + 1];
+        }
+    }
+}

kompute-shaders/op_rope_f32.comp

@@ -0,0 +1,73 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    const bool is_neox = (pcs.mode & 2) != 0;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_orig_ctx, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    const int p = inB[pcs.inBOff + i2];
+
+    float theta = float(p);
+
+    if (!is_neox) {
+        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+1];
+
+            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
+        }
+    } else {
+        const float inv_ndims = -1.f/pcs.n_dims;
+        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
+            const uint cur_rot = ic;
+
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint i0 = ic/2;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+pcs.n_dims/2];
+
+            out_[dst_data] = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        }
+
+        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
+            const uint i0 = ic;
+
+            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data + 0] = inA[src + 0];
+            out_[dst_data + 1] = inA[src + 1];
+        }
+    }
+}

kompute-shaders/op_scale.comp

@@ -0,0 +1,19 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+}

kompute-shaders/op_scale_8.comp

@@ -0,0 +1,23 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 8;
+
+    for (uint x = 0; x < 8; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+    }
+}

kompute-shaders/op_silu.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        const float y = in_[i + pcs.inOff];
+        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
+    }
+}

kompute-shaders/op_softmax.comp

@@ -0,0 +1,56 @@
+// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
+
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x_id = 0) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    float scale;
+    int mask;
+} pcs;
+
+void main() {
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
+    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
+    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
+    const uint pdst = extra_off + pcs.outOff; // Based from out_
+
+    // parallel max
+    float localMax = uintBitsToFloat(0xFF800000);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f));
+    }
+    float max_ = subgroupMax(localMax);
+
+    // parallel sum
+    float localSum = 0.0f;
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_);
+        localSum += exp_psrc0;
+        out_[pdst + i00] = exp_psrc0;
+    }
+
+    const float sum = subgroupAdd(localSum);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        out_[pdst + i00] /= sum;
+    }
+}

kompute-shaders/rope_common.comp

@@ -0,0 +1,67 @@
+#include "common.comp"
+
+// TODO: use a local size of 32 or more (Metal uses 1024)
+layout(local_size_x = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int n_dims;
+    int mode;
+    int n_orig_ctx;
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+float rope_yarn_ramp(const float low, const float high, const float 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.
+void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], float i0, float ext_factor, float mscale,
+    out float cos_theta, out 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[0], corr_dims[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 * log(1.0f / freq_scale);
+    }
+    cos_theta = cos(theta) * mscale;
+    sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
+    return n_dims * log(n_orig_ctx / (n_rot * TWOPI_F)) / (2 * log(base));
+}
+
+void rope_yarn_corr_dims(
+    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, out float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
+}

llama.cpp

@@ -11,6 +11,12 @@
 #  include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #  include "ggml-opencl.h"
+#elif defined(GGML_USE_VULKAN)
+#  include "ggml-vulkan.h"
+#elif defined(GGML_USE_SYCL)
+#  include "ggml-sycl.h"
+#elif defined(GGML_USE_KOMPUTE)
+#   include "ggml-kompute.h"
 #endif
 
 #ifdef GGML_USE_METAL
@@ -52,6 +58,7 @@
 #include <algorithm>
 #include <array>
 #include <cassert>
+#include <cfloat>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
@@ -196,6 +203,7 @@ enum llm_arch {
     LLM_ARCH_PHI2,
     LLM_ARCH_PLAMO,
     LLM_ARCH_CODESHELL,
+    LLM_ARCH_ORION,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -217,6 +225,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHI2,            "phi2"      },
     { LLM_ARCH_PLAMO,           "plamo"     },
     { LLM_ARCH_CODESHELL,       "codeshell" },
+    { LLM_ARCH_ORION,           "orion"     },
 };
 
 enum llm_kv {
@@ -641,6 +650,25 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ORION,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
 
     {
         LLM_ARCH_UNKNOWN,
@@ -1132,10 +1160,10 @@ struct llama_mlock {
     #ifdef __APPLE__
         #define MLOCK_SUGGESTION \
             "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
     #else
         #define MLOCK_SUGGESTION \
-            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+            "Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
     #endif
 
     bool raw_lock(const void * addr, size_t size) const {
@@ -1256,8 +1284,14 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer
     if (host_buffer) {
         buft = ggml_backend_cuda_host_buffer_type();
     }
+#elif defined(GGML_USE_SYCL)
+    buft = ggml_backend_sycl_host_buffer_type();
 #elif defined(GGML_USE_CPU_HBM)
     buft = ggml_backend_cpu_hbm_buffer_type();
+#elif defined(GGML_USE_VULKAN)
+    if (host_buffer) {
+        buft = ggml_backend_vk_host_buffer_type();
+    }
 #endif
 
     if (buft == nullptr) {
@@ -1275,8 +1309,17 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
     buft = ggml_backend_metal_buffer_type();
 #elif defined(GGML_USE_CUBLAS)
     buft = ggml_backend_cuda_buffer_type(gpu);
+#elif defined(GGML_USE_VULKAN)
+    buft = ggml_backend_vk_buffer_type();
+#elif defined(GGML_USE_SYCL)
+    buft = ggml_backend_sycl_buffer_type(gpu);
 #elif defined(GGML_USE_CLBLAST)
     buft = ggml_backend_opencl_buffer_type();
+#elif defined(GGML_USE_KOMPUTE)
+    buft = ggml_backend_kompute_buffer_type(gpu);
+    if (buft == nullptr) {
+        LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
+    }
 #endif
 
     if (buft == nullptr) {
@@ -1332,6 +1375,7 @@ enum e_model {
     MODEL_7B,
     MODEL_8B,
     MODEL_13B,
+    MODEL_14B,
     MODEL_15B,
     MODEL_30B,
     MODEL_34B,
@@ -2323,6 +2367,7 @@ struct llama_model_loader {
                 case GGML_TYPE_Q6_K:    ftype = LLAMA_FTYPE_MOSTLY_Q6_K;    break;
                 case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
                 case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
+                case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2671,6 +2716,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
         case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small";
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XSS - 3.0625 bpw";
 
         default: return "unknown, may not work";
     }
@@ -2683,6 +2729,7 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_7B:     return "7B";
         case MODEL_8B:     return "8B";
         case MODEL_13B:    return "13B";
+        case MODEL_14B:    return "14B";
         case MODEL_15B:    return "15B";
         case MODEL_30B:    return "30B";
         case MODEL_34B:    return "34B";
@@ -2950,7 +2997,15 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_ORION:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_14B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -3933,6 +3988,38 @@ static bool llm_load_tensors(
                         layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff});
                     }
                 } break;
+            case LLM_ARCH_ORION:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+
+
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -4029,7 +4116,7 @@ static bool llm_load_tensors(
 }
 
 // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
+static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
     try {
         llama_model_loader ml(fname, params.use_mmap, params.kv_overrides);
 
@@ -4050,6 +4137,22 @@ static int llama_model_load(const std::string & fname, llama_model & model, cons
             return 0;
         }
 
+#ifdef GGML_USE_KOMPUTE
+        if (params.n_gpu_layers > 0 && (
+            !(model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON)
+            || !(
+                model.ftype == LLAMA_FTYPE_ALL_F32 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_F16 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1
+            )
+        )) {
+            // TODO(cebtenzzre): propagate this error outside of llama_load_model_from_file
+            LLAMA_LOG_WARN("%s: disabling Kompute due to unsupported model arch or quantization\n", __func__);
+            params.n_gpu_layers = 0;
+        }
+#endif
+
         if (!llm_load_tensors(
             ml, model, params.n_gpu_layers, params.split_mode,  params.main_gpu, params.tensor_split, params.use_mlock,
             params.progress_callback, params.progress_callback_user_data
@@ -4563,6 +4666,126 @@ struct llm_build_context {
             ctx0 = nullptr;
         }
     }
+    struct ggml_cgraph * build_orion() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                // if (model.layers[il].bq) {
+                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                //     cb(Qcur, "Qcur", il);
+                // }
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                // if (model.layers[il].bk) {
+                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                //     cb(Kcur, "Kcur", il);
+                // }
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                // if (model.layers[il].bv) {
+                //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                //     cb(Vcur, "Vcur", il);
+                // }
+
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, cur,
+                    model.layers[il].ffn_up,   NULL,
+                    model.layers[il].ffn_gate, NULL,
+                    model.layers[il].ffn_down, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+
 
     struct ggml_cgraph * build_llama() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
@@ -6520,6 +6743,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_codeshell();
             } break;
+        case LLM_ARCH_ORION:
+            {
+                result = llm.build_orion();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -6652,7 +6879,7 @@ static int llama_decode_internal(
     }
 
     const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 1;
-    if (ggml_cpu_has_cublas() && fully_offloaded) {
+    if ((ggml_cpu_has_cublas() || ggml_cpu_has_vulkan()) && fully_offloaded) {
         n_threads = 1;
     }
 
@@ -7946,6 +8173,11 @@ void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * c
 }
 
 void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
+    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
+    // if (k >= (int32_t)candidates->size) {
+    //     return;
+    // }
+
     const int64_t t_start_sample_us = ggml_time_us();
 
     k = std::max(k, (int) min_keep);
@@ -7956,10 +8188,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can
         auto comp = [](const llama_token_data & a, const llama_token_data & b) {
             return a.logit > b.logit;
         };
-        if (k == (int) candidates->size) {
-            std::sort(candidates->data, candidates->data + candidates->size, comp);
-        } else {
+        if (k <= 128) {
             std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
+        } else {
+            constexpr int   nbuckets     = 128;
+            constexpr float bucket_low   = -10.0f;
+            constexpr float bucket_high  =  10.0f;
+            constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
+            constexpr float bucker_inter = -bucket_low * bucket_scale;
+
+            std::vector<int> bucket_idx(candidates->size);
+            std::vector<int> histo(nbuckets, 0);
+
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                const float val = candidates->data[i].logit;
+                int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+                ib = std::max(0, std::min(nbuckets-1, ib));
+                bucket_idx[i] = ib;
+                ++histo[ib];
+            }
+            int nhave = 0;
+            int ib = nbuckets - 1;
+            for ( ; ib >= 0; --ib) {
+                nhave += histo[ib];
+                if (nhave >= k) break;
+            }
+            std::vector<llama_token_data> tmp_tokens(nhave);
+            auto ptr = tmp_tokens.data();
+            std::vector<llama_token_data*> bucket_ptrs;
+            bucket_ptrs.reserve(nbuckets - ib);
+            for (int j = nbuckets - 1; j >= ib; --j) {
+                bucket_ptrs.push_back(ptr);
+                ptr += histo[j];
+            }
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                int j = bucket_idx[i];
+                if (j >= ib) {
+                    *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i];
+                }
+            }
+
+            ptr = tmp_tokens.data();
+            int ndone = 0;
+            for (int j = nbuckets-1; j > ib; --j) {
+                std::sort(ptr, ptr + histo[j], comp);
+                ptr += histo[j];
+                ndone += histo[j];
+            }
+            std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
+
+            std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data));
+
         }
         candidates->sorted = true;
     }
@@ -8007,21 +8286,56 @@ void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * can
         return;
     }
 
-    llama_sample_softmax(ctx, candidates);
-
     const int64_t t_start_sample_us = ggml_time_us();
 
-    float scale = candidates->data[0].p; // scale by max prob
-    size_t i = 1; // first token always matches
+    bool min_p_applied = false;
 
-    for (; i < candidates->size; ++i) {
-        if (candidates->data[i].p < p * scale && i >= min_keep) {
-            break; // prob too small
+    // if the candidates aren't sorted, try the unsorted implementation first
+    if (!candidates->sorted) {
+        std::vector<llama_token_data> filtered_tokens;
+
+        float max_logit = -FLT_MAX;
+        for (size_t i = 0; i < candidates->size; ++i) {
+            max_logit = std::max(max_logit, candidates->data[i].logit);
+        }
+        const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max
+
+        for (size_t i = 0; i < candidates->size; ++i) {
+            if (candidates->data[i].logit >= min_logit) {
+                filtered_tokens.push_back(candidates->data[i]);
+            }
+        }
+
+        // if we have enough values the operation was a success
+        if (filtered_tokens.size() >= min_keep) {
+            memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            candidates->size = filtered_tokens.size();
+            min_p_applied = true;
         }
     }
 
-    // Resize the output vector to keep only the matching tokens
-    candidates->size = i;
+    // if the candidates are sorted or the unsorted implementation failed, use this implementation
+    if (!min_p_applied) {
+        // Sort the logits in descending order
+        if (!candidates->sorted) {
+            std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+            candidates->sorted = true;
+        }
+
+        const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max
+        size_t i = 1; // first token always matches
+
+        for (; i < candidates->size; ++i) {
+            if (candidates->data[i].logit < min_logit && i >= min_keep) {
+                break; // prob too small
+            }
+        }
+
+        // Resize the output vector to keep only the matching tokens
+        candidates->size = i;
+    }
 
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
@@ -8925,6 +9239,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (new_type != GGML_TYPE_Q8_0) {
             new_type = GGML_TYPE_Q6_K;
         }
+    } else if (name == "token_embd.weight") {
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+            new_type = GGML_TYPE_Q2_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_Q4_K;
+        }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
         if (name.find("attn_v.weight") != std::string::npos) {
             if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
@@ -8935,7 +9256,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K;
             ++qs.i_ffn_down;
         }
-        else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K;
     } else if (name.find("attn_v.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
             new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
@@ -8943,6 +9263,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
             new_type = GGML_TYPE_Q4_K;
         }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
             new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
@@ -8980,6 +9303,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
             if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
         }
+        //else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+        //    if (i_layer < n_layer/8) new_type = GGML_TYPE_Q5_K;
+        //}
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
             new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
                      : arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
@@ -9011,13 +9337,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
     } else if (name.find("attn_output.weight") != std::string::npos) {
         if (arch != LLM_ARCH_FALCON) {
             if (qs.model.hparams.n_expert == 8) {
-                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS ||
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
                     new_type = GGML_TYPE_Q5_K;
                 }
             } else {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_Q3_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
             }
@@ -9060,7 +9387,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
     bool convert_incompatible_tensor = false;
     if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
         new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K ||
-        new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS) {
+        new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS ||
+        new_type == GGML_TYPE_IQ3_XXS) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -9074,6 +9402,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         switch (new_type) {
             case GGML_TYPE_IQ2_XXS:
             case GGML_TYPE_IQ2_XS:
+            case GGML_TYPE_IQ3_XXS:
             case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break;
             case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break;
             case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
@@ -9115,6 +9444,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS:quantized_type = GGML_TYPE_IQ3_XXS; break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -9950,6 +10280,36 @@ struct llama_context * llama_new_context_with_model(
                 }
             }
         }
+#elif defined(GGML_USE_VULKAN)
+        if (model->n_gpu_layers > 0) {
+            ggml_backend_t backend = ggml_backend_vk_init();
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_SYCL)
+        if (model->n_gpu_layers > 0) {
+            ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_KOMPUTE)
+        if (model->n_gpu_layers > 0) {
+            auto * backend = ggml_backend_kompute_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
 #endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {

llama.h

@@ -6,6 +6,9 @@
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
 #define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
+#elif defined(GGML_USE_SYCL)
+#include "ggml-sycl.h"
+#define LLAMA_MAX_DEVICES GGML_SYCL_MAX_DEVICES
 #else
 #define LLAMA_MAX_DEVICES 1
 #endif // GGML_USE_CUBLAS
@@ -46,7 +49,8 @@
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION 4
 
-#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
+    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
@@ -108,6 +112,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ3_XXS       = 23, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };

pocs/vdot/vdot.cpp

@@ -243,7 +243,6 @@ int main(int argc, char** argv) {
     if (useQ4_1) q41.resize(n4);
     else q40.resize(n4);
     std::vector<block_q8_0> q8(n8);
-    std::vector<int64_t> H(16, 0);
     double sumt = 0, sumt2 = 0, maxt = 0;
     double sumqt = 0, sumqt2 = 0, maxqt = 0;
     double sum = 0, sumq = 0, exactSum = 0;

scripts/ci-run.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+set -euo pipefail
+this=$(realpath "$0"); readonly this
+cd "$(dirname "$this")"
+shellcheck "$this"
+
+if (( $# != 1 && $# != 2  )); then
+    cat >&2 <<'EOF'
+usage:
+    ci-run.sh <tmp_dir> [<cache_dir>]
+
+This script wraps ci/run.sh:
+* If <tmp_dir> is a ramdisk, you can reduce writes to your SSD. If <tmp_dir> is not a ramdisk, keep in mind that total writes will increase by the size of <cache_dir>.
+    (openllama_3b_v2: quantized models are about 30GB)
+* Persistent model and data files are synced to and from <cache_dir>,
+    excluding generated .gguf files.
+    (openllama_3b_v2: persistent files are about 6.6GB)
+* <cache_dir> defaults to  ~/.cache/llama.cpp
+EOF
+    exit 1
+fi
+
+cd .. # => llama.cpp repo root
+
+tmp="$1"
+mkdir -p "$tmp"
+tmp=$(realpath "$tmp")
+echo >&2 "Using tmp=$tmp"
+
+cache="${2-$HOME/.cache/llama.cpp}"
+mkdir -p "$cache"
+cache=$(realpath "$cache")
+echo >&2 "Using cache=$cache"
+
+_sync() {
+    local from="$1"; shift
+    local to="$1"; shift
+
+    echo >&2 "Syncing from $from to $to"
+    mkdir -p "$from" "$to"
+    rsync -a "$from" "$to" --delete-during "$@"
+}
+
+_sync "$(realpath .)/" "$tmp/llama.cpp"
+_sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/"
+
+cd "$tmp/llama.cpp"
+bash ci/run.sh ci-out ci-mnt
+
+_sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P

scripts/run-with-preset.py

@@ -46,7 +46,7 @@ Formatting considerations:
 - To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings.
 - To define a tensor split, pass a list of floats.
 """
-usage = "run_with_preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
+usage = "run-with-preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
 epilog = ("  --<ARG_NAME> specify additional CLI ars to be passed to the binary (override all preset files). "
           "Unknown args will be ignored.")