cmake : do not search for curl libraries by ourselves (#14613 )

* cmake : do not search for curl libraries by ourselves * run : do not search for curl libraries by ourselves
SYCL: Initial set_rows kernel implementation (#14562 )
2026-02-05 13:53:23 +02:00 · 2025-07-10 15:29:05 +03:00 · 2025-07-10 09:29:38 +01:00 · 2025-07-10 10:00:20 +03:00 · 2025-07-10 08:19:37 +03:00 · 2025-07-10 08:19:13 +03:00
314 changed files with 21396 additions and 13912 deletions
--- a/.devops/tools.sh
+++ b/.devops/tools.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e

 # Read the first argument into a variable
--- a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -40,7 +40,7 @@ body:
    attributes:
        label: GGML backends
        description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
        multiple: true
    validations:
      required: true
--- a/.github/ISSUE_TEMPLATE/011-bug-results.yml
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -42,7 +42,7 @@ body:
    attributes:
        label: GGML backends
        description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
        multiple: true
    validations:
      required: true
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -1,10 +1,4 @@
 # https://github.com/actions/labeler
-Kompute:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-kompute.h
-            - ggml/src/ggml-kompute/**
-            - README-kompute.md
 Apple Metal:
    - changed-files:
        - any-glob-to-any-file:
@@ -93,3 +87,8 @@ Ascend NPU:
            - ggml/include/ggml-cann.h
            - ggml/src/ggml-cann/**
            - docs/backend/CANN.md
+OpenCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-opencl.h
+            - ggml/src/ggml-opencl/**
--- a/.github/workflows/build-cmake-pkg.yml
+++ b/.github/workflows/build-cmake-pkg.yml
@@ -0,0 +1,51 @@
+name: Build relocatable cmake package
+on:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  linux:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential tcl
+
+      - name: Build
+        run: |
+          PREFIX="$(pwd)"/inst
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
+          cmake --build build --config Release
+          cmake --install build --prefix "$PREFIX" --config Release
+
+          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
+          tclsh <<'EOF'
+          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
+          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          set build(version) "0.0.$build(number)"
+
+          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
+          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
+                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
+                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+
+          puts -nonewline "Checking llama-config.cmake version... "
+          foreach check $checks {
+              if {![regexp -expanded -- $check $llamaconfig]} {
+                  puts "\"$check\" failed!"
+                  exit 1
+              }
+          }
+          puts "success."
+          EOF
+
+          cd examples/simple-cmake-pkg
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
+          cmake --build build
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -5,10 +5,43 @@ on:
  push:
    branches:
      - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]
+
  pull_request:
    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]

 concurrency:
  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -51,7 +84,8 @@ jobs:
            -DCMAKE_BUILD_RPATH="@loader_path" \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_METAL_EMBED_LIBRARY=OFF \
+            -DGGML_METAL_SHADER_DEBUG=ON \
            -DGGML_RPC=ON
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

@@ -308,7 +342,7 @@ jobs:
          cd build
          export GGML_VK_VISIBLE_DEVICES=0
          # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 3600
+          ctest -L main --verbose --timeout 4200

  ubuntu-22-cmake-hip:
    runs-on: ubuntu-22.04
@@ -478,6 +512,9 @@ jobs:
  build-linux-cross:
    uses: ./.github/workflows/build-linux-cross.yml

+  build-cmake-pkg:
+    uses: ./.github/workflows/build-cmake-pkg.yml
+
  macOS-latest-cmake-ios:
    runs-on: macos-latest

@@ -628,7 +665,7 @@ jobs:
          ./build-xcframework.sh

  windows-msys2:
-    runs-on: windows-latest
+    runs-on: windows-2025

    strategy:
      fail-fast: false
@@ -678,28 +715,31 @@ jobs:
            cmake --build build --config ${{ matrix.build }} -j $(nproc)

  windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025

    env:
      OPENBLAS_VERSION: 0.3.23
      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2

    strategy:
      matrix:
        include:
          - build: 'cpu-x64 (static)'
+            arch: 'x64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
          - build: 'openblas-x64'
+            arch: 'x64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
          - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            arch: 'x64'
+            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
          - build: 'llvm-arm64'
+            arch: 'arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
          - build: 'llvm-arm64-opencl-adreno'
+            arch: 'arm64'
            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
-         # - build: 'kompute-x64'
-         #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'

    steps:
      - name: Clone
@@ -713,12 +753,6 @@ jobs:
          variant: ccache
          evict-old-files: 1d

-      - name: Clone Kompute submodule
-        id: clone_kompute
-        if: ${{ matrix.build == 'kompute-x64' }}
-        run: |
-          git submodule update --init ggml/src/ggml-kompute/kompute
-
      - name: Download OpenBLAS
        id: get_openblas
        if: ${{ matrix.build == 'openblas-x64' }}
@@ -734,9 +768,9 @@ jobs:

      - name: Install Vulkan SDK
        id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'vulkan-x64' }}
        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"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
          & "$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"
@@ -769,6 +803,8 @@ jobs:
      - name: libCURL
        id: get_libcurl
        uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}

      - name: Build
        id: cmake_build
@@ -778,6 +814,7 @@ jobs:
          cmake -S . -B build ${{ matrix.defines }} `
            -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release

      - name: Add libopenblas.dll
        id: add_libopenblas_dll
@@ -788,7 +825,7 @@ jobs:

      - name: Test
        id: cmake_test
-        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
+        if: ${{ matrix.arch == 'x64' }}
        run: |
          cd build
          ctest -L main -C Release --verbose --timeout 900
@@ -893,7 +930,7 @@ jobs:
          cmake --build build --config Release

  windows-latest-cmake-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022

    defaults:
      run:
@@ -927,7 +964,7 @@ jobs:

  windows-latest-cmake-hip:
    if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-latest
+    runs-on: windows-2022

    steps:
      - name: Clone
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -49,7 +49,8 @@ jobs:
        run: |
          sysctl -a
          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -103,7 +104,8 @@ jobs:
          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DGGML_METAL=OFF \
            -DGGML_RPC=ON
@@ -160,6 +162,8 @@ jobs:
        id: cmake_build
        run: |
          cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DGGML_BACKEND_DL=ON \
            -DGGML_NATIVE=OFF \
            -DGGML_CPU_ALL_VARIANTS=ON \
@@ -211,6 +215,8 @@ jobs:
        id: cmake_build
        run: |
          cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
            -DGGML_BACKEND_DL=ON \
            -DGGML_NATIVE=OFF \
            -DGGML_CPU_ALL_VARIANTS=ON \
@@ -235,7 +241,7 @@ jobs:
          name: llama-bin-ubuntu-vulkan-x64.zip

  windows-cpu:
-    runs-on: windows-latest
+    runs-on: windows-2025

    strategy:
      matrix:
@@ -271,7 +277,7 @@ jobs:
        env:
          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
            -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
            -DGGML_NATIVE=OFF ^
@@ -288,7 +294,7 @@ jobs:
          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
        run: |
          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
          7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*

      - name: Upload artifacts
@@ -298,11 +304,11 @@ jobs:
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

  windows:
-    runs-on: windows-latest
+    runs-on: windows-2025

    env:
      OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2

    strategy:
      matrix:
@@ -332,7 +338,7 @@ jobs:
        id: get_vulkan
        if: ${{ matrix.backend == 'vulkan' }}
        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"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
          & "$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 $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -448,7 +454,7 @@ jobs:
          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip

  windows-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022

    defaults:
      run:
@@ -520,7 +526,7 @@ jobs:
          name: llama-bin-win-sycl-x64.zip

  windows-hip:
-    runs-on: windows-latest
+    runs-on: windows-2022

    strategy:
      matrix:
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +0,0 @@
-[submodule "kompute"]
-	path = ggml/src/ggml-kompute/kompute
-	url = https://github.com/nomic-ai/kompute.git
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -95,7 +95,7 @@ endif()
 if (NOT DEFINED LLAMA_BUILD_COMMIT)
    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
 endif()
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})

 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
@@ -120,7 +120,6 @@ endfunction()

 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
-llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
--- a/build-xcframework.sh
+++ b/build-xcframework.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Options
 IOS_MIN_OS_VERSION=16.4
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # sample usage:
 #
@@ -39,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`

-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"

 if [ ! -z ${GG_BUILD_METAL} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
@@ -779,7 +779,7 @@ function gg_run_rerank_tiny {
    model_f16="${path_models}/ggml-model-f16.gguf"

    # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log

    # sample output
    # rerank score 0:    0.029
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -86,8 +86,7 @@ if (LLAMA_CURL)
    endif()
    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
    include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
 endif ()

 if (LLAMA_LLGUIDANCE)
@@ -112,13 +111,13 @@ if (LLAMA_LLGUIDANCE)

    ExternalProject_Add(llguidance_ext
        GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.20 (+ fix to build on GCC 15):
-        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
+        # v1.0.1:
+        GIT_TAG d795912fedc7d393de740177ea9ea761e7905774
        PREFIX ${CMAKE_BINARY_DIR}/llguidance
        SOURCE_DIR ${LLGUIDANCE_SRC}
        BUILD_IN_SOURCE TRUE
        CONFIGURE_COMMAND ""
-        BUILD_COMMAND cargo build --release
+        BUILD_COMMAND cargo build --release --package llguidance
        INSTALL_COMMAND ""
        BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
        UPDATE_COMMAND ""
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -988,10 +988,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
        params.tensor_buft_overrides.push_back({nullptr, nullptr});
    }

-    if (params.reranking && params.embedding) {
-        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
-    }
-
    if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
        throw std::runtime_error(string_format(
            "error: the supplied chat template is not supported: %s%s\n",
@@ -2710,6 +2706,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.embd_sep = value;
        }
    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
+    add_opt(common_arg(
+        {"--cls-separator"}, "STRING",
+        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
+        [](common_params & params, const std::string & value) {
+            params.cls_sep = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
    add_opt(common_arg(
        {"--host"}, "HOST",
        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -2731,6 +2734,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.public_path = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
+    add_opt(common_arg(
+        {"--api-prefix"}, "PREFIX",
+        string_format("prefix path the server serves from, without the trailing slash (default: %s)", params.api_prefix.c_str()),
+        [](common_params & params, const std::string & value) {
+            params.api_prefix = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
    add_opt(common_arg(
        {"--no-webui"},
        string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
@@ -2747,9 +2757,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
    add_opt(common_arg(
        {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
        [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
    add_opt(common_arg(
@@ -2790,6 +2801,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.ssl_file_cert = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
+    add_opt(common_arg(
+        {"--chat-template-kwargs"}, "STRING",
+        string_format("sets additional params for the json template parser"),
+        [](common_params & params, const std::string &  value) {
+            auto parsed = json::parse(value);
+            for (const auto & item : parsed.items()) {
+                params.default_template_kwargs[item.key()] = item.value().dump();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
    add_opt(common_arg(
        {"-to", "--timeout"}, "N",
        string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -3213,6 +3234,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.speculative.model.path = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
+    add_opt(common_arg(
+        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for K for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_k)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_k = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
+    add_opt(common_arg(
+        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for V for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_v)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_v = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));

    add_opt(common_arg(
        {"-mv", "--model-vocoder"}, "FNAME",
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -17,6 +17,8 @@
 #include <string>
 #include <vector>

+using json = nlohmann::ordered_json;
+
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
    auto time = std::chrono::system_clock::to_time_t(now);
    auto local_time = *std::localtime(&time);
@@ -140,6 +142,7 @@ struct templates_params {
    bool add_generation_prompt = true;
    bool enable_thinking = true;
    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    json extra_context;
 };

 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -720,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const

 static std::string apply(
    const common_chat_template & tmpl,
-    const nlohmann::ordered_json & messages,
-    const nlohmann::ordered_json & tools,
-    bool add_generation_prompt,
-    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
+    const struct templates_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt)
 {
    minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages;
-    tmpl_inputs.tools = tools;
-    tmpl_inputs.add_generation_prompt = add_generation_prompt;
-    tmpl_inputs.extra_context = extra_context;
+    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
+    if (tools_override) {
+        tmpl_inputs.tools = *tools_override;
+    } else {
+        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
+    }
+    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
+    tmpl_inputs.extra_context = inputs.extra_context;
+    if (additional_context) {
+        tmpl_inputs.extra_context.merge_patch(*additional_context);
+    }
    // TODO: add flag to control date/time, if only for testing purposes.
    // tmpl_inputs.now = std::chrono::system_clock::now();

@@ -828,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
        inputs.messages,
        "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");

-    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
    data.format = COMMON_CHAT_FORMAT_GENERIC;
    return data;
 }
@@ -904,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
    data.preserved_tokens = {
        "[TOOL_CALLS]",
    };
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
    data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
    return data;
 }
@@ -934,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
            adjusted_messages.push_back(msg);
        }
    }
-    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
    data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
    if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
        if (!inputs.enable_thinking) {
@@ -1122,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
    } else {
        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
    }
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
        {"date_string", format_time(inputs.now, "%d %b %Y")},
        {"tools_in_user_message", false},
        {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1187,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w

 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
    common_chat_params data;
-    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    auto prompt = apply(tmpl, inputs);

    // Hacks to fix the official (broken) prompt.
    // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1282,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
    LOG_DBG("%s\n", __func__);
    common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
        {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
        {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
    });
@@ -1338,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
    // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
    // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
    common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
    data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
    if (inputs.tools.is_array() && !inputs.tools.empty()) {
        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1465,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
    }

-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
    // TODO: if (has_raw_python)
    return data;
 }
@@ -1498,14 +1508,15 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
    common_chat_params data;

-    json additional_context = {
+    json extra_context = json {
        {"enable_thinking", inputs.enable_thinking},
    };
+    extra_context.update(inputs.extra_context);

-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
    data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
    if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!inputs.enable_thinking) {
+        if (!extra_context["enable_thinking"]) {
            data.prompt += "</think>";
        } else {
            data.thinking_forced_open = true;
@@ -1691,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {

 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
    common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
    data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
    data.grammar_lazy = false;
    if (!inputs.json_schema.is_null()) {
@@ -1722,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
    params.enable_thinking = inputs.enable_thinking;
    params.grammar = inputs.grammar;
    params.now = inputs.now;
+
+    params.extra_context = json::object();
+    for (auto el : inputs.chat_template_kwargs) {
+        params.extra_context[el.first] = json::parse(el.second);
+    }
+
    if (!inputs.json_schema.empty()) {
        params.json_schema = json::parse(inputs.json_schema);
    }
@@ -1838,7 +1855,7 @@ static common_chat_params common_chat_templates_apply_legacy(
    if (res < 0) {
        // if the custom "tmpl" is not supported, we throw an error
        // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-        throw std::runtime_error("this custom template is not supported");
+        throw std::runtime_error("this custom template is not supported, try using --jinja");
    }

    // if it turns out that our buffer is too small, we resize it
--- a/common/chat.h
+++ b/common/chat.h
@@ -7,6 +7,7 @@
 #include <chrono>
 #include <string>
 #include <vector>
+#include <map>

 struct common_chat_templates;

@@ -125,6 +126,7 @@ struct common_chat_templates_inputs {
    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
    bool enable_thinking = true;
    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    std::map<std::string, std::string> chat_template_kwargs;
 };

 struct common_chat_params {
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -706,11 +706,17 @@ bool fs_validate_filename(const std::string & filename) {
        // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
+
        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;

 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif

        filename_utf32 = converter.from_bytes(filename);
@@ -767,6 +773,9 @@ bool fs_validate_filename(const std::string & filename) {
    return true;
 }

+#include <iostream>
+
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -784,9 +793,16 @@ bool fs_create_directory_with_parents(const std::string & path) {
    // process path from front to back, procedurally creating directories
    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
        const std::wstring subpath = wpath.substr(0, pos_slash);
-        const wchar_t * test = subpath.c_str();

-        const bool success = CreateDirectoryW(test, NULL);
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
        if (!success) {
            const DWORD error = GetLastError();

@@ -800,8 +816,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                return false;
            }
        }
-
-        pos_slash += 1;
    }

    return true;
@@ -897,34 +911,6 @@ struct common_init_result common_init_from_params(common_params & params) {

    const llama_vocab * vocab = llama_model_get_vocab(model);

-    if (params.reranking) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
    auto cparams = common_context_params_to_llama(params);

    llama_context * lctx = llama_init_from_model(model, cparams);
@@ -966,6 +952,35 @@ struct common_init_result common_init_from_params(common_params & params) {
        }
    }

+    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
    // load and optionally apply lora adapters
    for (auto & la : params.lora_adapters) {
        llama_adapter_lora_ptr lora;
@@ -1143,11 +1158,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.op_offload        = !params.no_op_offload;
    cparams.swa_full          = params.swa_full;

-    if (params.reranking) {
-        cparams.embeddings    = true;
-        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
-    }
-
    cparams.type_k = params.cache_type_k;
    cparams.type_v = params.cache_type_v;

@@ -1280,6 +1290,9 @@ std::vector<llama_token> common_tokenize(
    int n_tokens = text.length() + 2 * add_special;
    std::vector<llama_token> result(n_tokens);
    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    if (n_tokens == std::numeric_limits<int32_t>::min()) {
+        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
+    }
    if (n_tokens < 0) {
        result.resize(-n_tokens);
        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
--- a/common/common.h
+++ b/common/common.h
@@ -8,6 +8,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <map>
 #include <sstream>

 #ifdef _WIN32
@@ -199,6 +200,9 @@ struct common_params_speculative {
    float   p_split      =  0.1f; // speculative decoding split probability
    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)

+    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
+    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
+
    struct cpu_params cpuparams;
    struct cpu_params cpuparams_batch;

@@ -355,7 +359,7 @@ struct common_params {
    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
    std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
    std::string embd_sep   = "\n";  // separator of embeddings
-    bool reranking         = false; // enable reranking support on server
+    std::string cls_sep    = "\t";  // separator of classification sequences

    // server params
    int32_t port           = 8080;         // server listens on this network port
@@ -366,6 +370,7 @@ struct common_params {

    std::string hostname      = "127.0.0.1";
    std::string public_path   = "";                                                                         // NOLINT
+    std::string api_prefix    = "";                                                                         // NOLINT
    std::string chat_template = "";                                                                         // NOLINT
    bool use_jinja = false;                                                                                 // NOLINT
    bool enable_chat_template = true;
@@ -378,6 +383,8 @@ struct common_params {
    std::string ssl_file_key  = "";                                                                         // NOLINT
    std::string ssl_file_cert = "";                                                                         // NOLINT

+    std::map<std::string, std::string> default_template_kwargs;
+
    // "advanced" endpoints are disabled by default for better security
    bool webui            = true;
    bool endpoint_slots   = false;
--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -41,49 +41,6 @@ static std::string build_repetition(const std::string & item_rule, int min_items
    return result;
 }

-/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
-class string_view {
-    const std::string & _str;
-    const size_t _start;
-    const size_t _end;
-public:
-    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
-
-    size_t size() const {
-        return _end - _start;
-    }
-
-    size_t length() const {
-        return size();
-    }
-
-    operator std::string() const {
-        return str();
-    }
-
-    std::string str() const {
-        return _str.substr(_start, _end - _start);
-    }
-
-    string_view substr(size_t pos, size_t len = std::string::npos) const {
-        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
-    }
-
-    char operator[](size_t pos) const {
-        auto index = _start + pos;
-        if (index >= _end) {
-            throw std::out_of_range("string_view index out of range");
-        }
-        return _str[_start + pos];
-    }
-
-    bool operator==(const string_view & other) const {
-        std::string this_str = *this;
-        std::string other_str = other;
-        return this_str == other_str;
-    }
-};
-
 static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
    auto has_min = min_value != std::numeric_limits<int>::min();
    auto has_max = max_value != std::numeric_limits<int>::max();
@@ -112,14 +69,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
        }
        out << "}";
    };
-    std::function<void(const string_view &, const string_view &)> uniform_range =
-        [&](const string_view & from, const string_view & to) {
+    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
+        [&](const std::string_view & from, const std::string_view & to) {
            size_t i = 0;
            while (i < from.length() && i < to.length() && from[i] == to[i]) {
                i++;
            }
            if (i > 0) {
-                out << "\"" << from.substr(0, i).str() << "\"";
+                out << "\"" << from.substr(0, i) << "\"";
            }
            if (i < from.length() && i < to.length()) {
                if (i > 0) {
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -128,6 +128,7 @@ models = [
    {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
    {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
+    {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -137,6 +138,12 @@ pre_computed_hashes = [
    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
+    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
+    # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
 ]


--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -757,7 +757,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
+| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
--- a/docs/build-s390x.md
+++ b/docs/build-s390x.md
@@ -0,0 +1,246 @@
+> [!IMPORTANT]
+> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
+
+# Build llama.cpp locally (for s390x)
+
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
+**To get the code:**
+
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+```
+
+## CPU Build with BLAS
+
+Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
+
+```bash
+cmake -S . -B build             \
+    -DCMAKE_BUILD_TYPE=Release  \
+    -DGGML_BLAS=ON              \
+    -DGGML_BLAS_VENDOR=OpenBLAS
+
+cmake --build build --config Release -j $(nproc)
+```
+
+**Notes**:
+
+-   For faster repeated compilation, install [ccache](https://ccache.dev/)
+-   By default, VXE/VXE2 is enabled. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_VXE=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   By default, NNPA is enabled when available. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_NNPA=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   For debug builds:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Debug    \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS
+    cmake --build build --config Debug -j $(nproc)
+    ```
+
+-   For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DBUILD_SHARED_LIBS=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+## Getting GGUF Models
+
+All models need to be converted to Big-Endian. You can achieve this in three cases:
+
+1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
+
+    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
+
+2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
+
+    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
+
+    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile model-name-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        model-directory/
+    ```
+
+    For example,
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile granite-3.3-2b-instruct-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        granite-3.3-2b-instruct/
+    ```
+
+3. **Convert existing GGUF Little-Endian model to Big-Endian**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
+    ```
+
+    For example,
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
+    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
+    ```
+
+    **Notes:**
+
+    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
+
+## IBM Accelerators
+
+### 1. SIMD Acceleration
+
+Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 2. NNPA Vector Intrinsics Acceleration
+
+Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 3. zDNN Accelerator
+
+_Only available in IBM z16 or later system. No direction at the moment._
+
+### 4. Spyre Accelerator
+
+_No direction at the moment._
+
+## Performance Tuning
+
+### 1. Virtualization Setup
+
+It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
+
+Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
+
+### 2. IFL (Core) Count
+
+It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
+
+Note: IFL count does not equate to vCPU count.
+
+### 3. SMT vs NOSMT (Simultaneous Multithreading)
+
+It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
+
+### 4. BLAS vs NOBLAS
+
+IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
+
+## Frequently Asked Questions (FAQ)
+
+1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
+
+    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
+
+    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
+
+2. I'm getting extremely poor performance when running inference on a model
+
+    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
+
+3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
+
+    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
+
+## Getting Help on IBM Z & LinuxONE
+
+1. **Bugs, Feature Requests**
+
+    Please file an issue in llama.cpp and ensure that the title contains "s390x".
+
+2. **Other Questions**
+
+    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
+
+## Appendix A: Hardware Support Matrix
+
+|         | Support | Minimum Compiler Version |
+| ------- | ------- | ------------------------ |
+| IBM z15 | ✅      |                          |
+| IBM z16 | ✅      |                          |
+| IBM z17 | ✅      | GCC 15.1.0               |
+
+-   ✅ - supported and verified to run as intended
+-   🚫 - unsupported, we are unlikely able to provide support
+
+## Appendix B: SIMD Support Matrix
+
+|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
+| ---------- | ----------- | ---- | ---- | ----- |
+| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP16       | ✅          | ✅   | ❓   | ❓    |
+| BF16       | 🚫          | 🚫   | ❓   | ❓    |
+| Q4_0       | ✅          | ✅   | ❓   | ❓    |
+| Q4_1       | ✅          | ✅   | ❓   | ❓    |
+| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| Q8_0       | ✅          | ✅   | ❓   | ❓    |
+| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
+| Q3_K       | ✅          | ✅   | ❓   | ❓    |
+| Q4_K       | ✅          | ✅   | ❓   | ❓    |
+| Q5_K       | ✅          | ✅   | ❓   | ❓    |
+| Q6_K       | ✅          | ✅   | ❓   | ❓    |
+| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
+| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
+| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
+| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
+| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+
+-   ✅ - acceleration available
+-   🚫 - acceleration unavailable, will still run using scalar implementation
+-   ❓ - acceleration unknown, please contribute if you can test it yourself
--- a/docs/build.md
+++ b/docs/build.md
@@ -1,6 +1,6 @@
 # Build llama.cpp locally

-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).

 The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.

@@ -557,6 +557,10 @@ ninja

 To read documentation for how to build on Android, [click here](./android.md)

+## IBM Z & LinuxONE
+
+To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
+
 ## Notes about GPU-accelerated backends

 The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.
--- a/docs/development/HOWTO-add-model.md
+++ b/docs/development/HOWTO-add-model.md
@@ -83,20 +83,22 @@ NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the conv

 ### 2. Define the model architecture in `llama.cpp`

-The model params and tensors layout must be defined in `llama.cpp`:
-1. Define a new `llm_arch`
-2. Define the tensors layout in `LLM_TENSOR_NAMES`
-3. Add any non-standard metadata in `llm_load_hparams`
-4. Create the tensors for inference in `llm_load_tensors`
-5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
+The model params and tensors layout must be defined in `llama.cpp` source files:
+1. Define a new `llm_arch` enum value in `src/llama-arch.h`.
+2. In `src/llama-arch.cpp`:
+    - Add the architecture name to the `LLM_ARCH_NAMES` map.
+    - Add the tensor mappings to the `LLM_TENSOR_NAMES` map.
+3. Add any non-standard metadata loading in the `llama_model_loader` constructor in `src/llama-model-loader.cpp`.
+4. If the model has a RoPE operation, add a case for the architecture in `llama_model_rope_type` function in `src/llama-model.cpp`.

 NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorch` dimensions.

 ### 3. Build the GGML graph implementation

-This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
-
-Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
+This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `src/llama-model.cpp`.
+Create a new struct that inherits from `llm_graph_context` and implement the graph-building logic in its constructor.
+Have a look at existing implementations like `llm_build_llama`, `llm_build_dbrx` or `llm_build_bert`.
+Then, in the `llama_model::build_graph` method, add a case for your architecture to instantiate your new graph-building struct.

 Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.

--- a/docs/docker.md
+++ b/docs/docker.md
@@ -25,6 +25,9 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)

 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, ROCm or MUSA library, you'll need to build the images locally for now).

--- a/examples/Miku.sh
+++ b/examples/Miku.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e

 AI_NAME="${AI_NAME:-Miku}"
--- a/examples/chat-13B.sh
+++ b/examples/chat-13B.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 set -e

--- a/examples/chat-persistent.sh
+++ b/examples/chat-persistent.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 set -euo pipefail

--- a/examples/chat-vicuna.sh
+++ b/examples/chat-vicuna.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 set -e

--- a/examples/chat.sh
+++ b/examples/chat.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 #
 # Temporary script - will be removed in the future
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -133,10 +133,36 @@ int main(int argc, char ** argv) {
    // max batch size
    const uint64_t n_batch = params.n_batch;

+    // get added sep and eos token, if any
+    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
+    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
+
    // tokenize the prompts and trim
    std::vector<std::vector<int32_t>> inputs;
    for (const auto & prompt : prompts) {
-        auto inp = common_tokenize(ctx, prompt, true, true);
+        std::vector<llama_token> inp;
+
+        // split classification pairs and insert expected separator tokens
+        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
+            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
+            std::string final_prompt;
+
+            for (size_t i = 0; i < pairs.size(); i++) {
+                final_prompt += pairs[i];
+                if (i != pairs.size() - 1) {
+                    if (!added_eos_token.empty()) {
+                        final_prompt += added_eos_token;
+                    }
+                    if (!added_sep_token.empty()) {
+                        final_prompt += added_sep_token;
+                    }
+                }
+            }
+
+            inp = common_tokenize(ctx, final_prompt, true, true);
+        } else {
+            inp = common_tokenize(ctx, prompt, true, true);
+        }
        if (inp.size() > n_batch) {
            LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                    __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -145,11 +171,11 @@ int main(int argc, char ** argv) {
        inputs.push_back(inp);
    }

-    // check if the last token is SEP
+    // check if the last token is SEP/EOS
    // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
    for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
-            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
+        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
+            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
            LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
        }
    }
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -55,6 +55,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                    } else if (type == GGML_TYPE_F32) {
                        v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I64) {
+                        v = (float) *(int64_t *) &data[i];
                    } else if (type == GGML_TYPE_I32) {
                        v = (float) *(int32_t *) &data[i];
                    } else if (type == GGML_TYPE_I16) {
@@ -134,6 +136,11 @@ static bool run(llama_context * ctx, const common_params & params) {

    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);

+    if (tokens.empty()) {
+        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
+        return false;
+    }
+
    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
        LOG_ERR("%s : failed to eval\n", __func__);
        return false;
--- a/examples/gritlm/gritlm.cpp
+++ b/examples/gritlm/gritlm.cpp
@@ -41,12 +41,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve

        // add input to batch (this increments n_tokens)
        for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
        }

        // clear previous kv_cache values (irrelevant for embeddings)
        llama_memory_clear(llama_get_memory(ctx), true);
-        llama_set_embeddings(ctx, true);
        llama_set_causal_attn(ctx, false);

        // run model
@@ -103,7 +102,6 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
    llama_token eos_token = llama_vocab_eos(vocab);

    llama_memory_clear(llama_get_memory(ctx), true);
-    llama_set_embeddings(ctx, false);
    llama_set_causal_attn(ctx, true);

    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -166,6 +164,8 @@ int main(int argc, char * argv[]) {
    llama_model_params mparams = common_model_params_to_llama(params);
    llama_context_params cparams = common_context_params_to_llama(params);

+    cparams.embeddings = true;
+
    llama_backend_init();

    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,6 +213,8 @@ int main(int argc, char * argv[]) {
        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
    }

+    llama_set_embeddings(ctx, false);
+
    // ### Generation ###
    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
    {
--- a/examples/jeopardy/jeopardy.sh
+++ b/examples/jeopardy/jeopardy.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e

 MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin
--- a/examples/reason-act.sh
+++ b/examples/reason-act.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 cd `dirname $0`
 cd ..
--- a/examples/server-llama2-13B.sh
+++ b/examples/server-llama2-13B.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 set -e

--- a/examples/simple-chat/simple-chat.cpp
+++ b/examples/simple-chat/simple-chat.cpp
@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
    auto generate = [&](const std::string & prompt) {
        std::string response;

-        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1;

        // tokenize the prompt
        const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,15 +113,16 @@ int main(int argc, char ** argv) {
        while (true) {
            // check if we have enough space in the context to evaluate this batch
            int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
            if (n_ctx_used + batch.n_tokens > n_ctx) {
                printf("\033[0m\n");
                fprintf(stderr, "context size exceeded\n");
                exit(0);
            }

-            if (llama_decode(ctx, batch)) {
-                GGML_ABORT("failed to decode\n");
+            int ret = llama_decode(ctx, batch);
+            if (ret != 0) {
+                GGML_ABORT("failed to decode, ret = %d\n", ret);
            }

            // sample the next token
--- a/examples/sycl/build.sh
+++ b/examples/sycl/build.sh
@@ -1,4 +1,4 @@
-
+#!/usr/bin/env bash
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
--- a/examples/sycl/run-llama2.sh
+++ b/examples/sycl/run-llama2.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
--- a/examples/sycl/run-llama3.sh
+++ b/examples/sycl/run-llama3.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash

 #  MIT license
 #  Copyright (C) 2025 Intel Corporation
--- a/examples/ts-type-to-grammar.sh
+++ b/examples/ts-type-to-grammar.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
 # python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -131,6 +131,7 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_NNPA             "ggml: enable nnpa"             ON)

 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -172,6 +173,7 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
+option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -179,7 +181,6 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
-option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -264,7 +265,6 @@ set(GGML_PUBLIC_HEADERS
    include/ggml-cann.h
    include/ggml-cpp.h
    include/ggml-cuda.h
-    include/ggml-kompute.h
    include/ggml-opt.h
    include/ggml-metal.h
    include/ggml-rpc.h
@@ -358,6 +358,13 @@ write_basic_package_version_file(
    VERSION ${GGML_INSTALL_VERSION}
    COMPATIBILITY SameMajorVersion)

+target_compile_definitions(ggml-base PRIVATE
+    GGML_VERSION="${GGML_INSTALL_VERSION}"
+    GGML_COMMIT="${GGML_BUILD_COMMIT}"
+)
+message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
+message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")
+
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
              ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
@@ -367,6 +374,8 @@ if (MSVC)
        /wd4005  # Macro redefinition
        /wd4244  # Conversion from one type to another type, possible loss of data
        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+        /wd4305  # Conversion from 'type1' to 'type2', possible loss of data
+        /wd4566  # Conversion from 'char' to 'wchar_t', possible loss of data
        /wd4996  # Disable POSIX deprecation warnings
        /wd4702  # Unreachable code warnings
    )
@@ -386,4 +395,46 @@ if (MSVC)
    disable_msvc_warnings(ggml-cpu-skylakex)
    disable_msvc_warnings(ggml-cpu-icelake)
    disable_msvc_warnings(ggml-cpu-alderlake)
+
+    if (GGML_BUILD_EXAMPLES)
+        disable_msvc_warnings(common-ggml)
+        disable_msvc_warnings(common)
+
+        disable_msvc_warnings(mnist-common)
+        disable_msvc_warnings(mnist-eval)
+        disable_msvc_warnings(mnist-train)
+
+        disable_msvc_warnings(gpt-2-ctx)
+        disable_msvc_warnings(gpt-2-alloc)
+        disable_msvc_warnings(gpt-2-backend)
+        disable_msvc_warnings(gpt-2-sched)
+        disable_msvc_warnings(gpt-2-quantize)
+        disable_msvc_warnings(gpt-2-batched)
+
+        disable_msvc_warnings(gpt-j)
+        disable_msvc_warnings(gpt-j-quantize)
+
+        disable_msvc_warnings(magika)
+        disable_msvc_warnings(yolov3-tiny)
+        disable_msvc_warnings(sam)
+
+        disable_msvc_warnings(simple-ctx)
+        disable_msvc_warnings(simple-backend)
+    endif()
+
+    if (GGML_BUILD_TESTS)
+        disable_msvc_warnings(test-mul-mat)
+        disable_msvc_warnings(test-arange)
+        disable_msvc_warnings(test-backend-ops)
+        disable_msvc_warnings(test-cont)
+        disable_msvc_warnings(test-conv-transpose)
+        disable_msvc_warnings(test-conv-transpose-1d)
+        disable_msvc_warnings(test-conv1d)
+        disable_msvc_warnings(test-conv2d)
+        disable_msvc_warnings(test-conv2d-dw)
+        disable_msvc_warnings(test-customop)
+        disable_msvc_warnings(test-dup)
+        disable_msvc_warnings(test-opt)
+        disable_msvc_warnings(test-pool)
+    endif ()
 endif()
--- a/ggml/cmake/common.cmake
+++ b/ggml/cmake/common.cmake
@@ -36,8 +36,7 @@ function(ggml_get_system_arch)
            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
        set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
-            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
        set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
        set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -339,7 +339,7 @@ extern "C" {
    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);

    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);

    // Tensor initialization
    GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -101,6 +101,7 @@ extern "C" {
    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
    GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
+    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);

@@ -133,6 +134,7 @@ extern "C" {

    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);

+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
    GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
    GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
    GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
--- a/ggml/include/ggml-kompute.h
+++ b/ggml/include/ggml-kompute.h
@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-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_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -314,6 +314,13 @@
 extern "C" {
 #endif

+    // Function type used in fatal error callbacks
+    typedef void (*ggml_abort_callback_t)(const char * error_message);
+
+    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+    // Returns the old callback for chaining
+    GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
+
    GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
    GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);

@@ -470,6 +477,7 @@ extern "C" {
        GGML_OP_TRANSPOSE,
        GGML_OP_GET_ROWS,
        GGML_OP_GET_ROWS_BACK,
+        GGML_OP_SET_ROWS,
        GGML_OP_DIAG,
        GGML_OP_DIAG_MASK_INF,
        GGML_OP_DIAG_MASK_ZERO,
@@ -481,14 +489,16 @@ extern "C" {
        GGML_OP_CONV_TRANSPOSE_1D,
        GGML_OP_IM2COL,
        GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D,
        GGML_OP_CONV_2D_DW,
        GGML_OP_CONV_TRANSPOSE_2D,
        GGML_OP_POOL_1D,
        GGML_OP_POOL_2D,
        GGML_OP_POOL_2D_BACK,
-        GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_UPSCALE,
        GGML_OP_PAD,
        GGML_OP_PAD_REFLECT_1D,
+        GGML_OP_ROLL,
        GGML_OP_ARANGE,
        GGML_OP_TIMESTEP_EMBEDDING,
        GGML_OP_ARGSORT,
@@ -518,6 +528,8 @@ extern "C" {
        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
        GGML_OP_OPT_STEP_ADAMW,

+        GGML_OP_GLU,
+
        GGML_OP_COUNT,
    };

@@ -541,6 +553,16 @@ extern "C" {
        GGML_UNARY_OP_COUNT,
    };

+    enum ggml_glu_op {
+        GGML_GLU_OP_REGLU,
+        GGML_GLU_OP_GEGLU,
+        GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_GEGLU_ERF,
+        GGML_GLU_OP_GEGLU_QUICK,
+
+        GGML_GLU_OP_COUNT,
+    };
+
    enum ggml_object_type {
        GGML_OBJECT_TYPE_TENSOR,
        GGML_OBJECT_TYPE_GRAPH,
@@ -626,6 +648,9 @@ extern "C" {

    // misc

+    GGML_API const char * ggml_version(void);
+    GGML_API const char * ggml_commit(void);
+
    GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
    GGML_API int64_t ggml_time_ms(void);
    GGML_API int64_t ggml_time_us(void);
@@ -656,6 +681,7 @@ extern "C" {
    GGML_API const char * ggml_op_symbol(enum ggml_op   op);

    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name

    GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
@@ -686,6 +712,9 @@ extern "C" {
    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
    GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);

+    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
+    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
+
    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);

@@ -757,6 +786,7 @@ extern "C" {
    GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);

    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);

    GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
    GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1085,6 +1115,89 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

+    // gated linear unit ops
+    // A: n columns, r rows,
+    // result is n / 2 columns, r rows,
+    // expects gate in second half of row, unless swapped is true
+    GGML_API struct ggml_tensor * ggml_glu(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_glu_op     op,
+             bool                 swapped);
+
+    GGML_API struct ggml_tensor * ggml_reglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_reglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // A: n columns, r rows,
+    // B: n columns, r rows,
+    GGML_API struct ggml_tensor * ggml_glu_split(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             struct ggml_tensor * b,
+             enum ggml_glu_op     op);
+
+    GGML_API struct ggml_tensor * ggml_reglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
    // normalize along rows
    GGML_API struct ggml_tensor * ggml_norm(
            struct ggml_context * ctx,
@@ -1184,6 +1297,19 @@ extern "C" {
            struct ggml_tensor  * a,
            float                 s);

+    // x = s * a + b
+    GGML_API struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    GGML_API struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
    // b -> view(a,offset,nb1,nb2,3), return modified a
    GGML_API struct ggml_tensor * ggml_set(
            struct ggml_context * ctx,
@@ -1374,6 +1500,23 @@ extern "C" {
            struct ggml_tensor  * b,  // row indices
            struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape

+    // a TD  [n_embd, ne1,    ne2,    ne3]
+    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
+    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
+    //
+    // undefined behavior if destination rows overlap
+    //
+    // broadcast:
+    //   ne2 % ne11 == 0
+    //   ne3 % ne12 == 0
+    //
+    // return view(a)
+    GGML_API struct ggml_tensor * ggml_set_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // destination
+            struct ggml_tensor  * b,  // source
+            struct ggml_tensor  * c); // row indices
+
    GGML_API struct ggml_tensor * ggml_diag(
        struct ggml_context     * ctx,
        struct ggml_tensor      * a);
@@ -1411,8 +1554,14 @@ extern "C" {
            struct ggml_context * ctx,
            struct ggml_tensor  * a);

+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
    // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
    // max_bias = 0.0f for no ALiBi
    GGML_API struct ggml_tensor * ggml_soft_max_ext(
            struct ggml_context * ctx,
@@ -1722,6 +1871,17 @@ extern "C" {
            struct ggml_tensor  * b,
            int                   stride);

+    GGML_API struct ggml_tensor * ggml_conv_2d_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+            struct ggml_tensor  * b,   // input data [W, H, C, N]
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
+
    enum ggml_op_pool {
        GGML_OP_POOL_MAX,
        GGML_OP_POOL_AVG,
@@ -1764,6 +1924,12 @@ extern "C" {
    enum ggml_scale_mode {
        GGML_SCALE_MODE_NEAREST  = 0,
        GGML_SCALE_MODE_BILINEAR = 1,
+
+        GGML_SCALE_MODE_COUNT
+    };
+
+    enum ggml_scale_flag {
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
    };

    // interpolate
@@ -1776,14 +1942,26 @@ extern "C" {

    // interpolate
    // interpolate scale to specified dimensions
-    GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
            int                   ne0,
            int                   ne1,
            int                   ne2,
            int                   ne3,
-            enum ggml_scale_mode  mode);
+            enum ggml_scale_mode  mode),
+        "use ggml_interpolate instead");
+
+    // Up- or downsamples the input to the specified size.
+    // 2D scale modes (eg. bilinear) are applied to the first two dimensions.
+    GGML_API struct ggml_tensor * ggml_interpolate(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            uint32_t              mode); // ggml_scale_mode [ | ggml_scale_flag...]

    // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
    GGML_API struct ggml_tensor * ggml_pad(
@@ -1801,6 +1979,17 @@ extern "C" {
            int                   p0,
            int                   p1);

+    // Move tensor elements by an offset given for each dimension. Elements that
+    // are shifted beyond the last position are wrapped around to the beginning.
+    GGML_API struct ggml_tensor * ggml_roll(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   shift0,
+            int                   shift1,
+            int                   shift2,
+            int                   shift3);
+
+
    // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
    // timesteps: [N,]
    // return: [N, dim]
@@ -1835,11 +2024,17 @@ extern "C" {

 #define GGML_KQ_MASK_PAD 64

-    // q:    [n_embd_k, n_batch,     n_head,    1]
-    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
-    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1878,7 +2073,8 @@ extern "C" {
            struct ggml_tensor  * dt,
            struct ggml_tensor  * A,
            struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);

    // partition into non-overlapping windows with padding if needed
    // example:
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -286,6 +286,10 @@ function(ggml_add_cpu_backend_variant tag_name)
        foreach (feat ${ARGN})
            set(GGML_INTERNAL_${feat} ON)
        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
    endif()

    ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -311,18 +315,45 @@ if (GGML_CPU_ALL_VARIANTS)
            # MSVC doesn't support AMX
            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
        endif()
-    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
-        # Many of these features are optional so we build versions with popular
-        # combinations and name the backends based on the version they were
-        # first released with
-        ggml_add_cpu_backend_variant(armv8.0_1)
-        ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
-        ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-        ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
-        ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
-        ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
-        ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
-        ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            # Many of these features are optional so we build versions with popular
+            # combinations and name the backends based on the version they were
+            # first released with
+            ggml_add_cpu_backend_variant(armv8.0_1)
+            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
+            # Android-specific backends with SoC-compatible feature sets
+            ggml_add_cpu_backend_variant(android_armv8.0_1)
+            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
+        elseif (APPLE)
+            ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
+            ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)
+            ggml_add_cpu_backend_variant(apple_m4             DOTPROD MATMUL_INT8 NOSVE SME)
+        else()
+            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(power0)
+            ggml_add_cpu_backend_variant(power7_1       POWER7)
+            ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)
+            ggml_add_cpu_backend_variant(power8_1       POWER8)
+            ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
+            ggml_add_cpu_backend_variant(power9         POWER9  VSX)
+            ggml_add_cpu_backend_variant(power10        POWER10 VSX)
+            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
+        else()
+            message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
    else()
        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
    endif()
@@ -334,7 +365,6 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
-ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -61,14 +61,13 @@
 #include "ggml-cann.h"
 #endif

-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif

 namespace fs = std::filesystem;
@@ -91,6 +90,8 @@ static std::string path_str(const fs::path & path) {

 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif

 #ifdef _WIN32
@@ -184,9 +185,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
        register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
        register_backend(ggml_backend_cpu_reg());
 #endif
@@ -570,7 +568,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
    ggml_backend_load_best("cann", silent, dir_path);
    ggml_backend_load_best("cuda", silent, dir_path);
    ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
    ggml_backend_load_best("metal", silent, dir_path);
    ggml_backend_load_best("rpc", silent, dir_path);
    ggml_backend_load_best("sycl", silent, dir_path);
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -817,8 +817,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
        }
        if (sched->debug > 1) {
            ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
            for (int j = 0; j < GGML_MAX_SRC; j++) {
                struct ggml_tensor * src = node->src[j];
                if (src == NULL) {
@@ -1826,7 +1827,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
    ggml_free(copy.ctx_unallocated);
 }

-bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node) {
    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
    if (copy.buffer == NULL) {
        return false;
@@ -1837,28 +1838,45 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t

    assert(g1->n_nodes == g2->n_nodes);

-    for (int i = 0; i < g1->n_nodes; i++) {
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
+    if (test_node != nullptr) {
+        // Compute the whole graph and only test the output for a specific tensor
+        ggml_backend_graph_compute(backend1, g1);
+        ggml_backend_graph_compute(backend2, g2);

-        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-        ggml_backend_graph_compute(backend1, &g1v);
-        ggml_backend_graph_compute(backend2, &g2v);
-
-        if (ggml_is_view_op(t1->op)) {
-            continue;
+        int test_node_idx = -1;
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            if (t1 == test_node) {
+                test_node_idx = i;
+                break;
+            }
        }
+        GGML_ASSERT(test_node_idx != -1);

-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
+        callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
+    } else {
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            struct ggml_tensor * t2 = g2->nodes[i];
+
+            assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+            struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+            struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+            ggml_backend_graph_compute(backend1, &g1v);
+            ggml_backend_graph_compute(backend2, &g2v);
+
+            if (ggml_is_view_op(t1->op)) {
+                continue;
+            }
+
+            // compare results, calculate rms etc
+            if (!callback(i, t1, t2, user_data)) {
+                break;
+            }
        }
    }
-
    ggml_backend_graph_copy_free(copy);

    return true;
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -65,8 +65,9 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v2.h>
+#include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_zero.h>
 #include <float.h>

 #include <cmath>
@@ -804,10 +805,11 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
        nb[i] = nb[i - 1] * ne[i - 1];
    }

-    ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
    aclTensor* zero =
        ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
    return zero;
+    GGML_UNUSED(n_bytes);
 }

 /**
@@ -2654,6 +2656,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
        memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
    }

+#ifdef ASCEND_310P
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    if (src0->type == GGML_TYPE_F16) {
+        src0_row.type = GGML_TYPE_F32;
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = ori_src0_nb[0];
+    src0_row.nb[1] = ori_src0_nb[1];
+    src0_row.nb[2] = ori_src0_nb[1];
+    src0_row.nb[3] = ori_src0_nb[1];
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            src0_row.data = src0_tmp_ptr;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+#endif
+
    std::vector<aclTensor*> src0_tensor_vec;
    std::vector<aclTensor*> src1_tensor_vec;
    std::vector<aclTensor*> dst_tensor_vec;
@@ -2701,9 +2764,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
    }

    size_t GROUP_SIZE = 128;
-    // GroupedMatmulV2 required tensor_list.size < 128
+    // GroupedMatmulV3 required tensor_list.size < 128
    for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV2
+        // split and call GroupedMatmulV3
        size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
        std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
        std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2713,7 +2776,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
        aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
        aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());

-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
            nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);

        ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -359,7 +359,7 @@ struct ggml_backend_cann_context {
        ggml_cann_set_device(device);
        description = aclrtGetSocName();

-        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
+        async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
            device, async_mode ? "ON" : "OFF");
    }
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2086,6 +2086,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                    return false;
            }
        } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
        case GGML_OP_CPY: {
            ggml_tensor *src = op->src[0];
            if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@@ -2182,12 +2188,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
        case GGML_OP_MUL:
        case GGML_OP_DIV:
        case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
        case GGML_OP_SQR:
        case GGML_OP_SQRT:
        case GGML_OP_CLAMP:
        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
        case GGML_OP_SUM_ROWS:
        case GGML_OP_ARGSORT:
        case GGML_OP_ACC:
@@ -2205,6 +2209,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
        case GGML_OP_PAD_REFLECT_1D:
        case GGML_OP_COUNT_EQUAL:
            return true;
+        case GGML_OP_SCALE:
+            float bias;
+            memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
+            return bias == 0.0f; // TODO: support bias != 0.0f
+        case GGML_OP_SOFT_MAX:
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
        case GGML_OP_FLASH_ATTN_EXT:{
            // derived from [ggml-cuda.cu]
            if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2227,6 +2239,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                // DeepSeek MLA
                return false;
            }
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
            if (op->src[0]->ne[3] != 1) {
                return false;
            }
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -5,7 +5,7 @@ function(ggml_add_cpu_backend_features cpu_name arch)
    # build, using set_source_files_properties() to set the arch flags is not possible
    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
-    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
@@ -158,48 +158,48 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                if (GGML_CPU_ARM_ARCH)
                    list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
                elseif(GGML_CPU_ALL_VARIANTS)
-                    if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-                        # Begin with the lowest baseline
-                        set(ARM_MCPU "armv8-a")
-                        set(ARCH_TAGS "")
-                        set(ARCH_DEFINITIONS "")
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+                    set(ARCH_DEFINITIONS "")

-                        # When a feature is selected, bump the MCPU to the first
-                        # version that supported it
-                        if (GGML_INTERNAL_DOTPROD)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
-                        endif()
-                        if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+fp16")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
-                        endif()
-                        if (GGML_INTERNAL_SVE)
-                            set(ARM_MCPU "armv8.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sve")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
-                        endif()
-                        if (GGML_INTERNAL_MATMUL_INT8)
-                            set(ARM_MCPU "armv8.6-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
-                        endif()
-                        if (GGML_INTERNAL_SVE2)
-                            set(ARM_MCPU "armv8.6-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sve2")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
-                        endif()
-                        if (GGML_INTERNAL_SME)
-                            set(ARM_MCPU "armv9.2-a")
-                            set(ARCH_TAGS "${ARCH_TAGS}+sme")
-                            list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
-                        endif()
-
-                        list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
-                        ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
                    endif()
+                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                    endif()
+                    if (GGML_INTERNAL_SVE)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                    endif()
+                    if (GGML_INTERNAL_MATMUL_INT8)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                    endif()
+                    if (GGML_INTERNAL_SVE2)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                    endif()
+                    if (GGML_INTERNAL_NOSVE)
+                        set(ARCH_TAGS "${ARCH_TAGS}+nosve")
+                    endif()
+                    if (GGML_INTERNAL_SME)
+                        set(ARM_MCPU "armv9.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                    endif()
+                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                endif()
            endif()

@@ -388,6 +388,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            else()
                list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
            endif()
+        elseif(GGML_CPU_ALL_VARIANTS)
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            # When a feature is selected, bump the MCPU to the first
+            # version that supported it
+            foreach(PVER RANGE 7 11)
+                if(DEFINED GGML_INTERNAL_POWER${PVER})
+                    set(POWERPC_MCPU "power${PVER}")
+                    list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
+                endif()
+            endforeach()
+            if (GGML_INTERNAL_VSX)
+                list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
+                list(APPEND ARCH_FLAGS -mvsx)
+            endif()
+
+            if (DEFINED POWERPC_MCPU)
+                list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
+            endif()
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
        else()
            if (GGML_CPU_POWERPC_CPUTYPE)
                list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
@@ -427,6 +448,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

        # TODO: Separation to determine activation of VX/VXE/VXE2
        if (${S390X_M} MATCHES "8561|8562")
+            set(GGML_NNPA OFF)
            message(STATUS "z15 target")
            list(APPEND ARCH_FLAGS -march=z15)
        elseif (${S390X_M} MATCHES "3931")
@@ -443,7 +465,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        endif()

        if (GGML_VXE)
+            message(STATUS "VX/VXE/VXE2 enabled")
            list(APPEND ARCH_FLAGS -mvx -mzvector)
+            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+        endif()
+
+        if (GGML_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
        endif()
    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
        message(STATUS "Wasm detected")
@@ -465,9 +494,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

        # Fetch KleidiAI sources:
        include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
        set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
+        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")

        if (POLICY CMP0135)
            cmake_policy(SET CMP0135 NEW)
@@ -560,4 +589,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
    if (EMSCRIPTEN)
        set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
    endif()
+
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
+    endif()
 endfunction()
--- a/ggml/src/ggml-cpu/amx/mmq.cpp
+++ b/ggml/src/ggml-cpu/amx/mmq.cpp
@@ -8,6 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
@@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_

        // Quantize these floats
        const float iscale = 127.f / amax;
-        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
+        y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
        const float id = ( amax != 0.0f ) ? iscale : 0.f;
        const __m512 vscale = _mm512_set1_ps(id);

@@ -1090,7 +1091,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));

        for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));

            __m512 vsum;
@@ -1113,8 +1114,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
        const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));

        for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
-            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));

            __m512 vsum;
@@ -1137,7 +1138,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));

        for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));

            __m512 vsum;
@@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
                    va[k] = _mm512_set1_epi32(a_ptr[k]);
                    vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
                }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
            }

            // load b
@@ -1498,8 +1499,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
                for (int k = 0; k < 8; ++k) {
                    va[k] = _mm512_set1_epi32(a_ptr[k]);
                }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
-                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
            }

            // load b
@@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
                    va[k] = _mm512_set1_epi32(a_ptr[k]);
                    va[k] = _mm512_add_epi8(va[k], off);
                }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
            }

            // load b
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -0,0 +1,184 @@
+#pragma once
+
+// Rename `_generic` functions if no native implementation is available.
+// This effectively selects the generic implementation.
+
+#if defined(GGML_CPU_GENERIC)
+// quants.c
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
+// repack.cpp
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__POWERPC__) || defined(__powerpc__)
+// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__loongarch64)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__riscv)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__s390x__)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__wasm__)
+// quants.c
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
--- a/ggml/src/ggml-cpu/arch/arm/quants.c
+++ b/ggml/src/ggml-cpu/arch/arm/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -62,7 +63,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        for (int j = 0; j < 8; j++) {
            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
@@ -104,7 +105,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        int32x4_t accv = vdupq_n_s32(0);

@@ -120,7 +121,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
            accv = vaddq_s32(accv, vi);
        }

-        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vaddvq_s32(accv));
    }
 #else
    GGML_UNUSED(nb);
@@ -194,10 +195,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);

            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
            };
            float32x4_t scale = vld1q_f32(_scale);

@@ -274,10 +275,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                    // dot product
                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
-                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
-                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                }

                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -313,9 +314,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

                    // dot product
                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                }

                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -354,9 +355,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

                    // dot product
                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                }

                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
@@ -404,8 +405,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);

-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -423,7 +424,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -464,10 +465,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
            const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i];

            float32_t summs_t[4] = {
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s)
            };
            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));

@@ -490,10 +491,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

            // mmla into int32x4_t
            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
            };
            float32x4_t scale = vld1q_f32(_scale);

@@ -539,7 +540,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];

-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

@@ -562,8 +563,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);

-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@@ -582,7 +583,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -666,10 +667,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -694,7 +695,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -739,8 +740,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

-        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);

        // extract the 5th bit via lookup table ((b) << 4)
        memcpy(&qh0, x0->qh, sizeof(qh0));
@@ -784,10 +785,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@@ -812,7 +813,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -864,10 +865,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);

            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
            };
            float32x4_t scale = vld1q_f32(_scale);

@@ -934,10 +935,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
-                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
-                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                }

                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
@@ -960,9 +961,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);

                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                }

                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -1002,8 +1003,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);

                    // scale creation
-                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
-                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+                    const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d);

                    // duplicate deq1 in first half of vector and deq2 in second half of vector
                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
@@ -1043,11 +1044,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));

        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
    }

    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -1059,7 +1060,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -1217,7 +1218,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

 #if defined(__ARM_FEATURE_DOTPROD)
        sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            }
        }

-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
    }

    *s = sumf;
@@ -1362,7 +1363,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

 #if defined(__ARM_FEATURE_DOTPROD)
        sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1393,7 +1394,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            }
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        sumf += (float) sumi * d;
    }
@@ -1425,9 +1426,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    switch (vector_length) {
        case 128:
            for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);

                const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1570,9 +1571,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        case 256:
        case 512:
            for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);

                const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1671,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sum = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1742,8 +1743,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -1805,7 +1806,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q3_sv = x[i].qs;
        const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask;
@@ -1981,7 +1982,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].hmask;
@@ -2112,7 +2113,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2258,18 +2259,18 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
                const float32x4_t dmins = {
-                    GGML_FP16_TO_FP32(x0->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x0->dmin) * y1->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d,
                };
                vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);

                const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                };
                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
            }
@@ -2289,8 +2290,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;
    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));

@@ -2377,8 +2378,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));

@@ -2478,9 +2479,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2520,8 +2521,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));

@@ -2630,9 +2631,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2827,10 +2828,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);

                const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                };

                visum = vsubq_s32(visum, vibias);
@@ -2858,7 +2859,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    svuint8_t q6h_1, q6h_2, q6h_3, q6h_4;

    for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q6 = x[i].ql;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3011,7 +3012,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q6 = x[i].ql;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3128,7 +3129,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -3199,7 +3200,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        float sumf1 = 0, sumf2 = 0;
@@ -3234,7 +3235,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        int32_t bsum = 0;
@@ -3284,7 +3285,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        const uint8x8_t scales8 = vld1_u8(x[i].scales);
@@ -3329,7 +3330,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3398,7 +3399,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    float sumf = 0;
    for (int i = 0; i < nb; ++i) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3458,7 +3459,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    float sumf = 0;
    for (int i = 0; i < nb; i++) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const int8_t  * q8 = y[i].qs;
        const uint8_t * qs = x[i].qs;
        const uint8_t * qh = x[i].qh;
@@ -3521,7 +3522,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3557,7 +3558,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3630,7 +3631,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3691,7 +3692,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3786,7 +3787,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

        }

-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
    }

    *s = sumf;
@@ -3817,7 +3818,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
@@ -3905,7 +3906,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

        }

-        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
    }

    *s = sumf;
@@ -3952,7 +3953,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qh += 2;
        }

-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
    }

    *s = sumf;
@@ -4003,13 +4004,13 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);

        sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+            GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
    }

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4071,7 +4072,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

        }

-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
    }

    *s = sumf;
@@ -4079,7 +4080,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
--- a/ggml/src/ggml-cpu/arch/arm/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp
--- a/ggml/src/ggml-cpu/arch/loongarch/quants.c
+++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -474,7 +475,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        // Quantize these floats
        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );

@@ -548,7 +549,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        // Quantize these floats
        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
        const __m256 mul = __lasx_xvreplfr2vr_s( id );

@@ -576,7 +577,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        // Compute the sum of the quants and set y[i].s
        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));

        // Convert int32 to int16
        ni0 = lsx_packs_w( ni0, ni1 );
@@ -667,7 +668,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);

@@ -699,7 +700,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    for (; ib + 1 < nb; ib += 2) {

        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );

        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);

@@ -717,7 +718,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );

        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);

@@ -766,7 +767,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -797,10 +798,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

    // Main loop
    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
@@ -834,7 +835,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -865,7 +866,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -902,7 +903,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -934,16 +935,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

    // Main loop
    for (; ib < nb; ++ib) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d));

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
        qx = __lasx_xvor_v(qx, bxhi);

-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d));
        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);

        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -973,7 +974,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -1003,7 +1004,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);

@@ -1023,7 +1024,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -1047,8 +1048,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1116,8 +1117,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -1170,7 +1171,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
        // Set up scales
@@ -1294,7 +1295,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1330,8 +1331,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

   for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1438,9 +1439,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1593,9 +1594,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1624,7 +1625,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1713,7 +1714,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1780,7 +1781,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __m256 accumf = (__m256)__lasx_xvldi(0);
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        __m256i sumi1 = __lasx_xvldi(0);
@@ -1820,7 +1821,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        int32_t bsum = 0;
@@ -1895,7 +1896,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    __m256 accumf = (__m256)__lasx_xvldi(0);
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;

@@ -1980,7 +1981,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,7 +2050,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __m256 accumf = (__m256)__lasx_xvldi(0);
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -2108,7 +2109,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    float sumf = 0;
    for (int i = 0; i < nb; i++) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const int8_t  * q8 = y[i].qs;
        const uint8_t * qs = x[i].qs;
        const uint8_t * qh = x[i].qh;
@@ -2168,7 +2169,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __m256 accumf = (__m256)__lasx_xvldi(0);
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2213,7 +2214,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2279,7 +2280,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __m256 accumf = (__m256)__lasx_xvldi(0);
    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -2340,7 +2341,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2451,7 +2452,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
        accum1 += d * sumi1;
    }
@@ -2484,7 +2485,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
@@ -2530,9 +2531,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
        const __m256i p_1 = lasx_madd_h(p16_1, mone);
        const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                __lasx_xvffint_s_w(p_2), accum2);
    }

@@ -2540,7 +2541,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2595,7 +2596,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
            sumi1 = __lasx_xvadd_w(p_1, sumi1);
            sumi2 = __lasx_xvadd_w(p_2, sumi2);
        }
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
    }

@@ -2604,7 +2605,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
--- a/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
+++ b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
@@ -0,0 +1,82 @@
+# include "ggml-backend-impl.h"
+
+#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#endif
+
+#include <string>
+
+struct powerpc_features {
+    std::string platform = "";
+    int power_version    = -1;
+
+    bool has_vsx         = false;
+
+    powerpc_features() {
+#if defined(__linux__)
+        unsigned long auxval = getauxval(AT_PLATFORM);
+        if (auxval) {
+            platform = std::string(reinterpret_cast<const char*>(auxval));
+            // TBD: Do systems exist that return this in uppercase?
+            if (platform.substr(0, 5) == "power") {
+                // Extractt a numeric suffix, if one exists
+                int vpos = -1;
+                for (int i = platform.length() - 1; i >= 0; i--) {
+                    if (std::isdigit(platform[i])) {
+                        vpos = i;
+                    } else {
+                        break;
+                    }
+                }
+                if (vpos > -1) {
+                    power_version = std::stoi(platform.substr(vpos));
+                }
+            }
+        }
+#endif
+        if (power_version >= 9) {
+            has_vsx = true;
+        }
+    }
+};
+
+static int ggml_backend_cpu_powerpc_score() {
+    int score = 1;
+    powerpc_features pf;
+
+// Platform scores
+#if defined(GGML_USE_POWER7)
+    if (pf.power_version < 7) { return 0; }
+    score += 1<<1;
+#endif
+#if defined(GGML_USE_POWER8)
+    if (pf.power_version < 8) { return 0; }
+    score += 1<<2;
+#endif
+#if defined(GGML_USE_POWER9)
+    if (pf.power_version < 9) { return 0; }
+    score += 1<<3;
+#endif
+#if defined(GGML_USE_POWER10)
+    if (pf.power_version < 10) { return 0; }
+    score += 1<<4;
+#endif
+#if defined(GGML_USE_POWER11)
+    if (pf.power_version < 11) { return 0; }
+    score += 1<<5;
+#endif
+
+// Feature scores
+#if defined(GGML_USE_VSX)
+    if (!pf.has_vsx) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
+
+#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
--- a/ggml/src/ggml-cpu/arch/powerpc/quants.c
+++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -67,7 +68,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float id = d ? 1.0f/d : 0.0f;
        const vector float vid = vec_splats(id);

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        for (int j = 0; j < 8; j++) {
            const vector float v  = vec_round(vec_mul(srcv[j], vid));
@@ -112,7 +113,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float id = d ? 1.0f/d : 0.0f;
        const vector float vid = vec_splats(id);

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        vector int accv = vec_splats(0);

@@ -127,7 +128,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        accv = vec_add(accv, vec_sld(accv, accv, 4));
        accv = vec_add(accv, vec_sld(accv, accv, 8));
-        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0));
    }

 #else
@@ -170,8 +171,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -214,7 +215,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -249,12 +250,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
        vsumf0 = vec_madd(vxmin, vys, vsumf0);

        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -291,7 +292,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -326,8 +327,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
@@ -379,7 +380,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -415,12 +416,12 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
        vsumf0 = vec_madd(vxmin, vys, vsumf0);

        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
@@ -470,7 +471,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -502,8 +503,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
@@ -542,7 +543,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -574,11 +575,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
        vector float vdmin = vec_mul(vxmin, vyd);

        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -708,8 +709,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -770,7 +771,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -962,7 +963,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1005,11 +1006,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
        vector float vdmin = vec_mul(vxmin, vyd);

        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -1177,9 +1178,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1222,11 +1223,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
        vector float vdmin = vec_mul(vxmin, vyd);

        UNUSED(kmask1);
@@ -1394,9 +1395,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1432,7 +1433,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -1591,7 +1592,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1659,7 +1660,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -1742,7 +1743,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        int32_t bsum = 0;
@@ -1790,7 +1791,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -1871,7 +1872,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -1939,7 +1940,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -2033,7 +2034,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    float sumf = 0;
    for (int i = 0; i < nb; i++) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const int8_t  * q8 = y[i].qs;
        const uint8_t * qs = x[i].qs;
        const uint8_t * qh = x[i].qh;
@@ -2096,7 +2097,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -2176,7 +2177,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2236,7 +2237,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -2329,7 +2330,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2394,7 +2395,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    vector float vsumf3 = vec_splats(0.0f);

    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
        vector float vyd = vec_splats(y[i].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -2505,7 +2506,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
@@ -2546,8 +2547,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
        __builtin_prefetch(y[ib].qs, 0, 1);


-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
        vector float vd = vec_mul(vxd, vyd);

        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -2582,7 +2583,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2620,7 +2621,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    for (int ibl = 0; ibl < nb; ++ibl) {

-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d));
        vector float vyd = vec_splats(y[ibl].d);
        vector float vd = vec_mul(vxd, vyd);

@@ -2697,7 +2698,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
--- a/ggml/src/ggml-cpu/arch/riscv/quants.c
+++ b/ggml/src/ggml-cpu/arch/riscv/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -45,7 +46,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);

@@ -85,7 +86,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d  = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);

@@ -102,7 +103,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        // set y[i].s
        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = GGML_FP32_TO_FP16(sum*d);
+        y[i].s = GGML_CPU_FP32_TO_FP16(sum*d);
    }

 #else
@@ -160,7 +161,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);

-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

 #endif
@@ -177,7 +178,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -225,7 +226,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);

-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

 #endif
@@ -242,7 +243,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -293,7 +294,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
        int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);

-        sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

 #endif
@@ -316,7 +317,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -366,7 +367,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
        int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);

-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

 #endif
@@ -389,7 +390,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -427,7 +428,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

 #endif
@@ -438,7 +439,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -465,8 +466,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const uint8_t * q2 = x[i].qs;
        const  int8_t * q8 = y[i].qs;
        const uint8_t * sc = x[i].scales;
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
        uint8_t *patmp = atmp;
        int vsums;
        int tmp;
@@ -569,8 +570,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            const int8_t *  q8 = y[i].qs;
            const uint8_t * sc = x[i].scales;

-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

            size_t vl = 16;

@@ -644,8 +645,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            const uint8_t * q2 = x[i].qs;
            const  int8_t * q8 = y[i].qs;
            const uint8_t * sc = x[i].scales;
-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
            uint8_t *patmp = atmp;
            int vsums;
            int tmp;
@@ -750,8 +751,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -916,7 +917,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            q3 += 32;    q8 += 128;   scale += 8;
        }

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        sumf += d * isum;
    }

@@ -1017,7 +1018,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

            }

-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

            sumf += d*sum_t;

@@ -1134,7 +1135,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                q3 += 32;    q8 += 128;   scale += 8;
            }

-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
            sumf += d * isum;
        }
        break;
@@ -1202,7 +1203,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1239,8 +1240,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int tmp, tmp2, sumi;
        __asm__ __volatile__(
@@ -1361,8 +1362,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

            size_t vl = 8;

-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

            vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
            vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
@@ -1422,8 +1423,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        break;
    case 128:
        for (int i = 0; i < nb; ++i) {
-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

            int tmp, tmp2, sumi;
            __asm__ __volatile__(
@@ -1580,9 +1581,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1627,8 +1628,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const uint8_t * GGML_RESTRICT hm = x[i].qh;
        const  int8_t * GGML_RESTRICT q8 = y[i].qs;

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;

        vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl);
        vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl);
@@ -1749,9 +1750,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1778,7 +1779,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

        const uint8_t * restrict q6 = x[i].ql;
        const uint8_t * restrict qh = x[i].qh;
@@ -1862,7 +1863,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    case 256:
        for (int i = 0; i < nb; ++i) {

-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

            const uint8_t * GGML_RESTRICT q6 = x[i].ql;
            const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1943,7 +1944,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    case 128:
        for (int i = 0; i < nb; ++i) {

-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;

            const uint8_t * restrict q6 = x[i].ql;
            const uint8_t * restrict qh = x[i].qh;
@@ -2058,7 +2059,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
--- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp
@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"

 #include <cmath>
@@ -90,16 +91,16 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);

                // vector version needs Zvfhmin extension
-                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                const float b_scales[8] = {
-                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                };
                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
@@ -129,7 +130,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                        }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                    }
                }
            }
@@ -181,20 +182,20 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo

                    // vector version needs Zvfhmin extension
                    const float a_scales[4] = {
-                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
                    };
                    const float b_scales[8] = {
-                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                    };
                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);

@@ -382,7 +383,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                        }
                    }
                }
--- a/ggml/src/ggml-cpu/arch/s390/quants.c
+++ b/ggml/src/ggml-cpu/arch/s390/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -49,7 +50,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f / d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        for (int j = 0; j < 8; j++) {
            const __vector float v = vec_mul(srcv[j], vec_splats(id));
@@ -94,7 +95,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f / d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        __vector int32_t acc = vec_splats(0);

@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
            acc = vec_add(acc, vi);
        }

-        y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
    }
 #else
    GGML_UNUSED(nb);
@@ -164,7 +165,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);

        const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
-        const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));

        acc = vec_madd(v_xy, v_d, acc);
    }
@@ -185,7 +186,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -219,7 +220,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        __builtin_prefetch(x[ib].qs, 0, 1);
        __builtin_prefetch(y[ib].qs, 0, 1);

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        const uint8x16_t v_x = vec_xl(0, x[ib].qs);
        const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
@@ -231,7 +232,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
        const float32x4_t v_xy = vec_float(v_xy_);

-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));

        acc = vec_madd(v_xy, v_d, acc);
    }
@@ -252,7 +253,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
        const float32x4_t v_xy = vec_float(v_xy_);
-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));

        acc = vec_madd(v_xy, v_d, acc);
    }
@@ -305,7 +306,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -348,7 +349,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sum = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * restrict x0l = x[i].qs;
        const uint8_t * restrict x0h = x[i].hmask;
@@ -497,7 +498,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -537,8 +538,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -647,9 +648,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -698,8 +699,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -819,9 +820,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -859,7 +860,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    int8x16_t  v_y[4];

    for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT x0l = x[i].ql;
        const uint8_t * GGML_RESTRICT x0h = x[i].qh;
@@ -1004,7 +1005,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1071,7 +1072,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 //    float sumf = 0;

 //    for (int i = 0; i < nb; ++i) {
-//        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //        const int8_t   * GGML_RESTRICT q8 = y[i].qs;

@@ -1121,7 +1122,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

 //     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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 //         int32_t bsum = 0;
@@ -1182,12 +1183,12 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
        const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);

-        sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
    }

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1257,7 +1258,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
            sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
        }

-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
    }

    *s = sumf;
@@ -1265,7 +1266,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
--- a/ggml/src/ggml-cpu/arch/wasm/quants.c
+++ b/ggml/src/ggml-cpu/arch/wasm/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -65,7 +66,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        for (int j = 0; j < 8; j++) {
            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);

        v128_t accv = wasm_i32x4_splat(0);

@@ -126,7 +127,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
            accv = wasm_i32x4_add(accv, vi);
        }

-        y[i].s = GGML_FP32_TO_FP16(
+        y[i].s = GGML_CPU_FP32_TO_FP16(
                d * (wasm_i32x4_extract_lane(accv, 0) +
                     wasm_i32x4_extract_lane(accv, 1) +
                     wasm_i32x4_extract_lane(accv, 2) +
@@ -324,8 +325,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        );

        // Accumulate results with scaling
-        float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
-        float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d);
+        float scale0 = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
+        float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d);

        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
@@ -348,7 +349,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -428,7 +429,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
    }

    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -454,7 +455,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -491,7 +492,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];

-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);

        const v128_t m4b = wasm_i8x16_splat(0x0F);

@@ -538,7 +539,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
    }

    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -564,7 +565,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -620,7 +621,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));

        // Convert to float and accumulate
-        const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
+        const float scale = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
    }

@@ -635,7 +636,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -746,8 +747,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            isum += wasm_i32x4_extract_lane(isum_vec, 0);
        }

-        const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf += dall * isum - dmin * summs;
    }

@@ -768,8 +769,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -880,7 +881,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        // Accumulate results
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const v128_t v_d = wasm_f32x4_splat(d);
        v128_t v_sum = wasm_f32x4_add(
            wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
@@ -957,7 +958,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -991,8 +992,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Corrected sign

        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1136,9 +1137,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1170,8 +1171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    float sumf = 0;

    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Fixed sign

        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1331,9 +1332,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1420,7 +1421,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        wasm_v128_store(&aux32[0], acc0);
        wasm_v128_store(&aux32[4], acc1);

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) {
            sums[l] += d * aux32[l];
        }
@@ -1470,7 +1471,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
--- a/ggml/src/ggml-cpu/arch/x86/quants.c
+++ b/ggml/src/ggml-cpu/arch/x86/quants.c
@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"

 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -256,9 +257,9 @@ static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_

 // quad fp16 delta calculation
 static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
-    // GGML_FP16_TO_FP32 is faster than Intel F16C
-    return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0)));
+    // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -305,7 +306,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        // Quantize these floats
        const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
        const __m256 mul = _mm256_set1_ps( id );

@@ -401,7 +402,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        // Quantize these floats
        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
        const __m256 mul = _mm256_set1_ps( id );

@@ -425,7 +426,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

 #if defined(__AVX2__)
        // Compute the sum of the quants and set y[i].s
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));

        // Convert int32 to int16
        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -455,7 +456,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
        // Compute the sum of the quants and set y[i].s
        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));

        // Convert int32 to int16
        ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -552,7 +553,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);

@@ -613,7 +614,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );

        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);

@@ -631,7 +632,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);

        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );

        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);

@@ -680,7 +681,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
@@ -711,10 +712,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

    // Main loop
    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        const __m256 d0v = _mm256_set1_ps( d0 );
        const __m256 d1v = _mm256_set1_ps( d1 );
@@ -752,7 +753,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -783,7 +784,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -807,7 +808,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));

        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -851,7 +852,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
@@ -883,16 +884,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

    // Main loop
    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
        qx = _mm256_or_si256(qx, bxhi);

-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);

        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -910,9 +911,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi

    // Main loop
    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));

-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);

        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -926,7 +927,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        bxh = _mm_or_si128(bxh, bxhih);
        bx_0 = MM256_SET_M128I(bxh, bxl);

-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);

        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@@ -956,7 +957,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
@@ -986,7 +987,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
    // Main loop
    for (; ib < nb; ++ib) {
        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);

@@ -1025,7 +1026,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
@@ -1144,7 +1145,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
        }

        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));

        sumi0 = _mm256_sub_epi16(sumi0, ysum);
        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
@@ -1190,7 +1191,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            }
        }

-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
    }

    *s = sumf;
@@ -1244,7 +1245,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
        }

        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));

        sumi0 = _mm256_add_epi16(sumi0, sumi1);
        sumi0 = _mm256_sub_epi16(sumi0, ysum);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            }
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        sumf += (float) sumi * d;
    }
@@ -1299,8 +1300,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1366,8 +1367,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -1533,7 +1534,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1638,7 +1639,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1824,7 +1825,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1862,8 +1863,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

   for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1928,8 +1929,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

   for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,9 +2050,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2092,8 +2093,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        memcpy(utmp, x[i].scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -2170,8 +2171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2311,9 +2312,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2344,7 +2345,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2422,7 +2423,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    for (int i = 0; i < nb; ++i) {

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        const uint8_t * GGML_RESTRICT q4 = x[i].ql;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2555,7 +2556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2622,7 +2623,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        __m256i sumi1 = _mm256_setzero_si256();
@@ -2663,7 +2664,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        __m128i sumi1_0 = _mm_setzero_si128();
@@ -2717,7 +2718,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        int32_t bsum = 0;
@@ -2792,7 +2793,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;

@@ -2913,7 +2914,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;

@@ -3035,7 +3036,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3104,7 +3105,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3177,7 +3178,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3253,7 +3254,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    float sumf = 0;
    for (int i = 0; i < nb; i++) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const int8_t  * q8 = y[i].qs;
        const uint8_t * qs = x[i].qs;
        const uint8_t * qh = x[i].qh;
@@ -3313,7 +3314,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3358,7 +3359,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3414,7 +3415,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3480,7 +3481,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3565,7 +3566,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3648,7 +3649,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3753,7 +3754,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
        accum1 += d * sumi1;

@@ -3801,7 +3802,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
        accum1 += d * sumi1;

@@ -3835,7 +3836,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
@@ -3947,7 +3948,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 8; qh += 4;
        }

-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));

        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
@@ -4033,7 +4034,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qs += 8; qh += 4;
        }

-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));

        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
@@ -4083,7 +4084,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
            qh += 2;
        }

-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
    }

    *s = sumf;
@@ -4129,9 +4130,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                _mm256_cvtepi32_ps(p_2), accum2);
    }

@@ -4164,7 +4165,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4219,7 +4220,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
            sumi1 = _mm256_add_epi32(p_1, sumi1);
            sumi2 = _mm256_add_epi32(p_2, sumi2);
        }
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
    }

@@ -4267,7 +4268,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
        }
        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
    }

@@ -4276,7 +4277,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"

 #include <cmath>
@@ -39,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
    float tmp[16];

    for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
    }

    for (int i = 0; i < 8; i++) {
-        tmp[i + 8] = GGML_FP16_TO_FP32(y[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
    }

    return _mm512_loadu_ps(tmp);
@@ -54,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
    _mm_storeu_si128((__m128i*)tmphalf, x);

    for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
    }

    return _mm512_loadu_ps(tmp);
@@ -67,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
    float tmp[8];

    for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
    }

    return _mm256_loadu_ps(tmp);
@@ -76,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
    float tmp[8];

    for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
    }

    return _mm256_loadu_ps(tmp);
@@ -88,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang

    _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
    for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
    }

    return _mm256_loadu_ps(tmp);
@@ -211,7 +212,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
            id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;

            // Store the scale for the individual block
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);

            // Store the values in blocks of eight values - Aim is to use these later for block interleaving
            srcv[row_iter][0] = v0;
@@ -297,7 +298,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
            const float d = amax / ((1 << 7) - 1);
            id[row_iter] = d ? 1.0f / d : 0.0f;

-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
        }

        for (int j = 0; j < QK8_0 * 4; j++) {
@@ -647,7 +648,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);

                // Load and convert to FP32 scale from block_q8_0
-                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));

                // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
                __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
@@ -706,7 +707,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                        }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                    }
                }
            }
@@ -972,13 +973,13 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                        sumi2 = sumi2 * scales_1[j];
                        sumi += sumi1 + sumi2;
                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                }
            }
            for (int sb = 0; sb < 8; sb++) {
                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                }
            }
        }
@@ -1755,7 +1756,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                        }
                    }
                }
@@ -3259,7 +3260,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                sumi2 = sumi2 * scales_1[j];
                                sumi += sumi1 + sumi2;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                        }
                    }
                }
@@ -3268,7 +3269,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    for(int m = 0; m < 4; m++) {
                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                        }
                    }
                }
--- a/ggml/src/ggml-cpu/common.h
+++ b/ggml/src/ggml-cpu/common.h
@@ -4,6 +4,7 @@
 #include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
+#include "simd-mappings.h"

 #ifdef __cplusplus

@@ -12,11 +13,11 @@
 // convenience functions/macros for use in template calls
 // note: these won't be required after the 'traits' lookup table is used.
 static inline ggml_fp16_t f32_to_f16(float x) {
-    return GGML_FP32_TO_FP16(x);
+    return GGML_CPU_FP32_TO_FP16(x);
 }

 static inline float f16_to_f32(ggml_fp16_t x) {
-    return GGML_FP16_TO_FP32(x);
+    return GGML_CPU_FP16_TO_FP32(x);
 }

 static inline ggml_bf16_t f32_to_bf16(float x) {
--- a/ggml/src/ggml-cpu/ggml-cpu-impl.h
+++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h
@@ -62,11 +62,17 @@ struct ggml_compute_params {
 #if defined(__s390x__) && defined(__VEC__)
 #ifndef __VXE__
 #define __VXE__
-#endif
+#endif  // __VXE__
 #ifndef __VXE2__
 #define __VXE2__
-#endif
-#endif
+#endif  // __VXE2__
+#endif  // __s390x__ && __VEC__
+
+#if defined(__s390x__) && defined(GGML_NNPA)
+#ifndef __NNPA__
+#define __NNPA__
+#endif  // __NNPA__
+#endif  // __s390x__ && GGML_NNPA

 #if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
@@ -371,7 +377,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
 #endif

-typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef signed   char char8x16_t  __attribute__((vector_size(16)));
 typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));

 typedef int8_t  int8x16_t __attribute__((vector_size(16)));
@@ -382,10 +388,10 @@ typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
 typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
 typedef uint32_t uint32x4_t __attribute__((vector_size(16)));

-typedef float float32x4_t __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute((vector_size(16)));
+typedef float  float32x4_t  __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));

-typedef signed long long long64x2_t __attribute((vector_size(16)));
+typedef signed   long long long64x2_t  __attribute__((vector_size(16)));
 typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));

 typedef struct ggml_uint8x16x2_t {
@@ -503,31 +509,9 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);

+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
+int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
+
 #ifdef __cplusplus
 }
 #endif
-
-#define GGML_DO_PRAGMA_(x) _Pragma (#x)
-#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
-#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
-// Note for Apple targets:
-// - clang: aliases are not supported on darwin
-// - all native kernels need to be implemented in both x86 and arm files
-// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
-# define GGML_WEAK_ALIAS(name, alias)
-#elif defined(__GNUC__)
-// GCC/Clang on *nix
-# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
-#elif defined(_MSC_VER) && defined(_WIN64)
-// MSVC
-// Note: C name mangling varies across different calling conventions
-// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
-# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
-#elif defined(_MSC_VER) && defined(WIN32)
-// ref: https://github.com/ggml-org/whisper.cpp/pull/3239#issuecomment-2958224591
-# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:_" #name "=_" #alias))
-#else
-# error "Unsupported compiler for GGML_WEAK_ALIAS"
-#endif
-
-#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -72,15 +72,13 @@
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)

+// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
+float ggml_table_f32_f16[1 << 16];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
-    int has_neon;
-    int has_dotprod;
-    int has_i8mm;
-    int has_sve;
    int sve_cnt;
-    int has_sme;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
+} ggml_arm_arch_features = { 0 };
 #endif


@@ -197,6 +195,7 @@ typedef pthread_t ggml_thread_t;

 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
    [GGML_TYPE_F32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
        .vec_dot_type             = GGML_TYPE_F32,
        .nrows                    = 1,
@@ -559,6 +558,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }

+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
+    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
+    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
    cpu_set_t cpuset;
@@ -670,87 +677,15 @@ bool ggml_is_numa(void) {

 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
 #endif

 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__)
-    uint32_t hwcap = getauxval(AT_HWCAP);
-    uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
-    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
-    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
-
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-#elif defined(__APPLE__)
-    int oldp = 0;
-    size_t size = sizeof(oldp);
-    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_neon = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_dotprod = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_i8mm = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_sme = oldp;
-
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#else
-// Run-time CPU feature detection not implemented for this platform, fallback to compile time
-#if defined(__ARM_NEON)
-    ggml_arm_arch_features.has_neon = 1;
-#else
-    ggml_arm_arch_features.has_neon = 0;
-#endif
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    ggml_arm_arch_features.has_i8mm = 1;
-#else
-    ggml_arm_arch_features.has_i8mm = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.has_sve = 1;
-    ggml_arm_arch_features.sve_cnt = 16;
-#else
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
-    ggml_arm_arch_features.has_sme = 1;
-#else
-    ggml_arm_arch_features.has_sme = 0;
-#endif
-#endif
 }
-#endif
+
+#endif // __ARM_ARCH

 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
    GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -805,7 +740,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
            {
                assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                }
            } break;
        case GGML_TYPE_BF16:
@@ -864,7 +799,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
            {
                assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                }
            } break;
        case GGML_TYPE_BF16:
@@ -915,7 +850,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
        case GGML_TYPE_F16:
            {
                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
            }
        case GGML_TYPE_BF16:
            {
@@ -960,7 +895,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
        case GGML_TYPE_F16:
            {
                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
            } break;
        case GGML_TYPE_BF16:
            {
@@ -989,7 +924,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
        case GGML_TYPE_I32:
            return ((int32_t *) data)[0];
        case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
        case GGML_TYPE_BF16:
            return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
        case GGML_TYPE_F32:
@@ -1016,7 +951,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
            } break;
        case GGML_TYPE_F16:
            {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
            } break;
        case GGML_TYPE_BF16:
            {
@@ -1054,7 +989,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
            }
        case GGML_TYPE_F16:
            {
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
            }
        case GGML_TYPE_BF16:
            {
@@ -1093,7 +1028,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
            } break;
        case GGML_TYPE_F16:
            {
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
            } break;
        case GGML_TYPE_BF16:
            {
@@ -1120,7 +1055,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
        case GGML_TYPE_I32:
            return ((int32_t *) data)[0];
        case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
        case GGML_TYPE_BF16:
            return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
        case GGML_TYPE_F32:
@@ -1147,7 +1082,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
            } break;
        case GGML_TYPE_F16:
            {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
            } break;
        case GGML_TYPE_BF16:
            {
@@ -1258,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
    }
 }

-static void ggml_compute_forward_mul_mat(
+void ggml_compute_forward_mul_mat(
        const struct ggml_compute_params * params,
              struct ggml_tensor * dst) {

@@ -1883,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_get_rows_back(params, tensor);
            } break;
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_compute_forward_set_rows(params, tensor);
+            } break;
        case GGML_OP_DIAG:
            {
                ggml_compute_forward_diag(params, tensor);
@@ -1927,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_im2col_back_f32(params, tensor);
            } break;
+        case GGML_OP_CONV_2D:
+            {
+                ggml_compute_forward_conv_2d(params, tensor);
+            } break;
        case GGML_OP_CONV_2D_DW:
            {
                ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -1959,6 +1902,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_pad_reflect_1d(params, tensor);
            } break;
+        case GGML_OP_ROLL:
+            {
+                ggml_compute_forward_roll(params, tensor);
+            } break;
        case GGML_OP_ARANGE:
            {
                ggml_compute_forward_arange(params, tensor);
@@ -2006,6 +1953,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_unary(params, tensor);
            } break;
+        case GGML_OP_GLU:
+            {
+                ggml_compute_forward_glu(params, tensor);
+            } break;
        case GGML_OP_GET_REL_POS:
            {
                ggml_compute_forward_get_rel_pos(params, tensor);
@@ -2216,6 +2167,20 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                    GGML_ABORT("fatal error");
            }
            break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(node)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    {
+                        n_tasks = n_threads;
+                    } break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            break;
        case GGML_OP_SILU_BACK:
        case GGML_OP_MUL:
        case GGML_OP_DIV:
@@ -2232,6 +2197,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                n_tasks = n_threads;
            } break;
        case GGML_OP_GET_ROWS:
+        case GGML_OP_SET_ROWS:
            {
                // FIXME: get_rows can use additional threads, but the cost of launching additional threads
                // decreases performance with GPU offloading
@@ -2268,6 +2234,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
            } break;
        case GGML_OP_IM2COL:
        case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D:
        case GGML_OP_CONV_2D_DW:
        case GGML_OP_CONV_TRANSPOSE_1D:
        case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2283,6 +2250,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD:
        case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_ROLL:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_ARGSORT:
@@ -2785,6 +2753,10 @@ struct ggml_cplan ggml_graph_plan(
                            GGML_ABORT("fatal error");
                        }
                    } break;
+                case GGML_OP_CONV_2D:
+                    {
+                        cur = GGML_IM2COL_WORK_SIZE;
+                    } break;
                case GGML_OP_CONV_TRANSPOSE_2D:
                    {
                        const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -3185,6 +3157,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
    return ggml_graph_compute(cgraph, &cplan);
 }

+void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
+    memcpy(y, x, n * sizeof(float));
+}
+
 void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
    int64_t i = 0;
 #if defined(__F16C__)
@@ -3205,9 +3181,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
        __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
        _mm_storel_epi64((__m128i *)(y + i), y_vec);
    }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
+        float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
+        uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
+    for (; i + 3 < n; i += 4) {
+        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
 #endif
    for (; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
    }
 }

@@ -3231,9 +3222,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
        __m128 y_vec = _mm_cvtph_ps(x_vec);
        _mm_storeu_ps(y + i, y_vec);
    }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i + 0));
+        vec_xst(v_yl, 0, (float *)(y + i + 4));
+    }
+    for (; i + 3 < n; i += 4) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i));
+    }
 #endif
+
    for (; i < n; ++i) {
-        y[i] = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP16_TO_FP32(x[i]);
    }
 }

@@ -3433,9 +3440,17 @@ int ggml_cpu_has_vxe(void) {
 #endif
 }

+int ggml_cpu_has_nnpa(void) {
+#if defined(GGML_NNPA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return ggml_arm_arch_features.has_neon;
+    return 1;
 #else
    return 0;
 #endif
@@ -3443,7 +3458,7 @@ int ggml_cpu_has_neon(void) {

 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
    return 0;
 #endif
@@ -3451,7 +3466,7 @@ int ggml_cpu_has_dotprod(void) {

 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
    return 0;
 #endif
@@ -3459,7 +3474,7 @@ int ggml_cpu_has_sve(void) {

 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
    return 0;
 #endif
@@ -3475,14 +3490,14 @@ int ggml_cpu_get_sve_cnt(void) {

 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
    return 0;
 #endif
 }

 void ggml_cpu_init(void) {
-    // needed to initialize f16 tables
+    // needed to initialize ggml_time
    {
        struct ggml_init_params params = { 0, NULL, false };
        struct ggml_context * ctx = ggml_init(params);
@@ -3503,9 +3518,10 @@ void ggml_cpu_init(void) {
                    uint16_t u16;
                    ggml_fp16_t fp16;
                } u = {i};
-                float f = GGML_FP16_TO_FP32(u.fp16);
-                ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
-                ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
+                float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
+                ggml_table_f32_f16[i] = f;
+                ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f));
+                ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
            }

            const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st

    switch (op->op) {
        case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
            return
                op->type != GGML_TYPE_IQ3_XXS &&
                op->type != GGML_TYPE_IQ3_S   &&
@@ -578,6 +579,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
        if (ggml_cpu_has_vxe()) {
            features.push_back({ "VXE", "1" });
        }
+        if (ggml_cpu_has_nnpa()) {
+            features.push_back({ "NNPA", "1" });
+        }
        if (ggml_cpu_has_wasm_simd()) {
            features.push_back({ "WASM_SIMD", "1" });
        }
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -52,8 +52,8 @@
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
+#include "simd-mappings.h"

-#include <atomic>
 #include <array>
 #include <type_traits>

@@ -63,7 +63,7 @@
 #define NOINLINE __attribute__((__noinline__))
 #endif

-#if defined(__ARM_NEON) || defined(__AVX512F__)
+#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
 #define VECTOR_REGISTERS 32
 #else
 #define VECTOR_REGISTERS 16
@@ -74,7 +74,7 @@
 namespace {

 inline float unhalf(ggml_fp16_t d) {
-    return GGML_FP16_TO_FP32(d);
+    return GGML_CPU_FP16_TO_FP32(d);
 }

 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -110,6 +110,12 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
 inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC

+#if defined(__VXE__) || defined(__VXE2__)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
+#endif
+
 #if defined(__MMA__)
 typedef vector unsigned char vec_t;
 typedef __vector_quad acc_t;
@@ -163,6 +169,13 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
 #endif
 #endif

+#if defined(__VXE__) || defined(__VXE2__)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vec_madd(a, b, c);
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM

@@ -179,6 +192,13 @@ inline float hsum(float16x8_t x) {
 }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC

+#if defined(__VXE__) || defined(__VXE2__)
+inline float hsum(float32x4_t x) {
+    float32x4_t tmp = x + vec_reve(x);
+    return tmp[0] + tmp[1];
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 inline float hsum(__m128 x) {
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
@@ -228,6 +248,21 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) {
 #endif // _MSC_VER
 #endif // __ARM_NEON

+#if defined(__VXE__) || defined(__VXE2__)
+template <> inline float32x4_t load(const ggml_fp16_t * p) {
+    float tmp[4];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_CPU_FP16_TO_FP32(p[i]);
+    }
+
+    return vec_xl(0, (const float *)(tmp));
+}
+template <> inline float32x4_t load(const float * p) {
+    return vec_xl(0, p);
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 template <> inline __m128 load(const float *p) {
    return _mm_loadu_ps(p);
@@ -394,8 +429,6 @@ class tinyBLAS {

    template <int RM, int RN, int BM>
    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
-        static std::atomic<int64_t> current_chunk;
-
        GGML_ASSERT(m % (RM * BM) == 0);
        const int64_t ytiles = m / (RM * BM);
        const int64_t xtiles = (n + RN -1) / RN;
@@ -410,7 +443,7 @@ class tinyBLAS {
        if (params->ith == 0) {
            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
        }

        ggml_barrier(params->threadpool);
@@ -439,8 +472,7 @@ class tinyBLAS {
                GGML_ASSERT(jj == jj2);
            }

-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
        }

        ggml_barrier(params->threadpool);
@@ -3323,6 +3355,14 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
            (const float *)B, ldb,
            (float *)C, ldc};
        return tb.matmul(m, n);
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__MMA__)
        if (k % 8)
            return false;
@@ -3414,6 +3454,16 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                (float *)C, ldc};
            return tb.matmul(m, n);
        }
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #endif
        return false;
    }
--- a/ggml/src/ggml-cpu/llamafile/sgemm.h
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.h
@@ -1,6 +1,11 @@
 #pragma once
 #include <stdint.h>
 #include <stdbool.h>
+
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -20,6 +20,9 @@

 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);

+// Work buffer size for im2col operations in CONV2D
+#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -53,6 +56,7 @@ void ggml_compute_forward_permute(const struct ggml_compute_params * params, str
 void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -64,6 +68,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
 void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -72,6 +77,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
 void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -92,6 +98,7 @@ void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, st
 void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -104,6 +111,7 @@ void ggml_compute_forward_custom(const struct ggml_compute_params * params, stru
 void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);

 #ifdef __cplusplus
 }
--- a/ggml/src/ggml-cpu/quants.c
+++ b/ggml/src/ggml-cpu/quants.c
@@ -2,9 +2,12 @@
 #include "ggml-common.h"

 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "ggml-quants.h"
 #include "quants.h"

+#include "arch-fallback.h"
+
 #include <string.h>
 #include <assert.h>
 #include <float.h>
@@ -38,12 +41,10 @@ void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
    quantize_row_q8_0_ref(x, y, k);
 }
-GGML_CPU_NATIVE_IMPL(quantize_row_q8_0)

 void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
    quantize_row_q8_1_ref(x, y, k);
 }
-GGML_CPU_NATIVE_IMPL(quantize_row_q8_1)

 //
 // 2-6 bit quantization in super-blocks
@@ -104,7 +105,6 @@ void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy,
 void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
    quantize_row_q8_K_ref(x, y, k);
 }
-GGML_CPU_NATIVE_IMPL(quantize_row_q8_K)

 //===================================== Dot products =================================

@@ -138,12 +138,11 @@ void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
        }

        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_0_q8_0)

 // TODO: add WASM SIMD
 void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -176,12 +175,11 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_1_q8_1)

 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
@@ -220,12 +218,11 @@ void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_0_q8_0)

 void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_1;
@@ -264,12 +261,11 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
        }

        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_1_q8_1)

 void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
@@ -295,12 +291,11 @@ void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            sumi += x[ib].qs[j]*y[ib].qs[j];
        }

-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q8_0_q8_0)

 void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(nrc == 1);
@@ -348,12 +343,11 @@ void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
            }
        }

-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq1_0_q8_K)

 void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(nrc == 1);
@@ -379,14 +373,13 @@ void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
            }
        }

-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);

        sumf += (float) sumi * d;
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq2_0_q8_K)

 void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(nrc == 1);
@@ -413,8 +406,8 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            summs += y[i].bsums[j] * (sc[j] >> 4);
        }

-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

        int isum = 0;
        int is = 0;
@@ -439,7 +432,6 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
    }
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q2_K_q8_K)

 void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -513,13 +505,12 @@ void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q3_K_q8_K)

 void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -587,15 +578,14 @@ void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_K_q8_K)

 void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy,  size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -668,15 +658,14 @@ void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
        sumf -= dmin * sumi;
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_K_q8_K)

 void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -726,13 +715,12 @@ void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
    }
    for (int l = 0; l < 8; ++l) sumf += sums[l];
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q6_K_q8_K)

 void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -752,7 +740,7 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
        int32_t bsum = 0;
@@ -775,7 +763,6 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
    }
    *s = 0.125f * sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xxs_q8_K)

 void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -792,7 +779,7 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
        const uint8_t  * GGML_RESTRICT sc = x[i].scales;
        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -826,7 +813,6 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
    *s = 0.125f * sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xs_q8_K)

 void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -844,7 +830,7 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    float sumf = 0;
    for (int i = 0; i < nb; i++) {

-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const int8_t  * q8 = y[i].qs;
        const uint8_t * qs = x[i].qs;
        const uint8_t * qh = x[i].qh;
@@ -879,7 +865,6 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,

    *s = 0.125f * sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_s_q8_K)

 void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -898,7 +883,7 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
        const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -924,7 +909,6 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
    }
    *s = 0.25f * sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_xxs_q8_K)

 void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -941,7 +925,7 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,

    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 float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
        const uint8_t * GGML_RESTRICT qs = x[i].qs;
        const uint8_t * GGML_RESTRICT qh = x[i].qh;
        const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -981,7 +965,6 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_s_q8_K)

 void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -1020,12 +1003,11 @@ void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_s_q8_K)

 void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(n % QK_K == 0);
@@ -1082,12 +1064,11 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
            qh += 2;
        }

-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
    }

    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_m_q8_K)

 void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(nrc == 1);
@@ -1107,7 +1088,7 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
    float sumf = 0;

    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
        int sumi1 = 0, sumi2 = 0;
        for (int j = 0; j < QK4_NL/2; ++j) {
            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1117,7 +1098,6 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_nl_q8_0)

 void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    assert(nrc == 1);
@@ -1134,7 +1114,7 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,

    float sumf = 0;
    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
        uint16_t h = x[ibl].scales_h;
        const uint8_t * qs = x[ibl].qs;
        const int8_t  * q8 = y[ibl].qs;
@@ -1164,7 +1144,6 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
    *s = sumf;
 }
-GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_xs_q8_K)

 // ============================ 4-bit non-linear quants

--- a/ggml/src/ggml-cpu/quants.h
+++ b/ggml/src/ggml-cpu/quants.h
@@ -84,33 +84,6 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

-#if defined(GGML_CPU_GENERIC)
-#define quantize_row_q8_0_generic quantize_row_q8_0
-#define quantize_row_q8_1_generic quantize_row_q8_1
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
-#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
-#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
-#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-#endif
-
 #ifdef __cplusplus
 }
 #endif
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -6,8 +6,11 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"

+#include "arch-fallback.h"
+
 #include <cmath>
 #include <cstring>
 #include <cassert>
@@ -70,7 +73,7 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG
            const float d = amax / ((1 << 7) - 1);
            id[row_iter] = d ? 1.0f / d : 0.0f;

-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
        }

        for (int j = 0; j < QK8_0 * 4; j++) {
@@ -83,7 +86,6 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x4)

 void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
    assert(QK8_0 == 32);
@@ -109,7 +111,7 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
            const float d = amax / ((1 << 7) - 1);
            id[row_iter] = d ? 1.0f / d : 0.0f;

-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
        }

        for (int j = 0; j < QK8_0 * 4; j++) {
@@ -122,7 +124,6 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x8)

 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
    assert(QK_K == 256);
@@ -174,7 +175,6 @@ void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GG
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_K_4x8)

 } // extern "C"

@@ -237,14 +237,13 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                }
            }
        }
        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x4_q8_0)

 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -282,14 +281,13 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                }
            }
        }
        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x8_q8_0)

 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -328,7 +326,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                        }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                    }
                }
            }
@@ -336,7 +334,6 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_8x8_q8_0)

 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK_K;
@@ -400,13 +397,13 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                        sumi2 = sumi2 * scales_1[j];
                        sumi += sumi1 + sumi2;
                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                }
            }
            for (int sb = 0; sb < 8; sb++) {
                uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                }
            }
        }
@@ -415,7 +412,6 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_K_8x8_q8_K)

 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -454,7 +450,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                        }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                    }
                }
            }
@@ -462,7 +458,6 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemv_iq4_nl_4x4_q8_0)

 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -506,7 +501,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                                }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                            }
                        }
                    }
@@ -519,7 +514,6 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x4_q8_0)

 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -562,7 +556,7 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                        }
                    }
                }
@@ -574,7 +568,6 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x8_q8_0)

 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -617,7 +610,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                        }
                    }
                }
@@ -629,7 +622,6 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_8x8_q8_0)

 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK_K;
@@ -697,7 +689,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                sumi2 = sumi2 * scales_1[j];
                                sumi += sumi1 + sumi2;
                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                        }
                    }
                }
@@ -706,7 +698,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                    for(int m = 0; m < 4; m++) {
                        const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                        for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                        }
                    }
                }
@@ -719,7 +711,6 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_K_8x8_q8_K)

 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
    const int qk = QK8_0;
@@ -763,7 +754,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                                }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                            }
                        }
                    }
@@ -776,7 +767,6 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
        }
    }
 }
-GGML_CPU_NATIVE_IMPL(ggml_gemm_iq4_nl_4x4_q8_0)

 } // extern "C"

@@ -1174,13 +1164,24 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
        // not realy a GGML_TYPE_Q8_0 but same size.
        switch (op->op) {
            case GGML_OP_MUL_MAT:
-                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                return true;
+                {
+                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                    return true;
+                }
            case GGML_OP_MUL_MAT_ID:
-                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                size = GGML_PAD(size, sizeof(int64_t));  // + padding for next bloc.
-                size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
-                return true;
+                {
+                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
+
+                    const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
+                    const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
+
+                    const size_t sizeof_mmid_row_mapping = sizeof(int64_t);
+
+                    size += sizeof_mmid_row_mapping*ne02*(ne12 + 1);
+
+                    return true;
+                }
            default:
                // GGML_ABORT("fatal error");
                break;
@@ -1316,14 +1317,17 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
            int32_t i2;
        };

-        GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
-                                      n_as * ne12 * sizeof(mmid_row_mapping)));
+        GGML_ASSERT(params->wsize >=
+                (GGML_PAD(nbw3, sizeof(int64_t)) +
+                 n_as*(ne12 + 1)*sizeof(mmid_row_mapping))
+                );

-        auto * wdata             = (char *)     params->wdata;
-        auto * wdata_src1_end    = (char *)     wdata + GGML_PAD(nbw3, sizeof(int64_t));
-        auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+        auto * wdata          = (char *)params->wdata;
+        auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));

-        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as);  // [n_as][ne12]
+        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
+        auto * matrix_row_counts = (int64_t *) (wdata_src1_end);                                        // [n_as]
+        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]

        // src1: float32 => param type
        for (int64_t i12 = 0; i12 < ne12; ++i12) {
@@ -1408,44 +1412,45 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
    }
 };

-// instance for Q4
-static const tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
-static const tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
-static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
-static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
-
-// instance for IQ4
-static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
-
 }  // namespace ggml::cpu::repack

 static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
+
+    // instance for Q4
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+
+    // instance for IQ4
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+
    if (cur->type == GGML_TYPE_Q4_0) {
        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
            if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::repack::q4_0_8x8_q8_0;
+                return &q4_0_8x8_q8_0;
            }
        }
        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
            if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::q4_0_4x8_q8_0;
+                return &q4_0_4x8_q8_0;
            }
        }
        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
            if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::q4_0_4x4_q8_0;
+                return &q4_0_4x4_q8_0;
            }
        }
    } else if (cur->type == GGML_TYPE_Q4_K) {
        if (ggml_cpu_has_avx2()) {
            if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::repack::q4_K_8x8_q8_K;
+                return &q4_K_8x8_q8_K;
            }
        }
    } else if (cur->type == GGML_TYPE_IQ4_NL) {
        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
            if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::iq4_nl_4x4_q8_0;
+                return &iq4_nl_4x4_q8_0;
            }
        }
    }
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -64,10 +64,6 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro
 extern "C" {
 #endif

-// Workaround for clang:
-// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
-// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
-#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -81,7 +77,6 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-#endif // !defined(__clang__)

 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -98,22 +93,6 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);

-#if defined(GGML_CPU_GENERIC)
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#endif
-
 #if defined(__cplusplus)
 } // extern "C"
 #endif
--- a/ggml/src/ggml-cpu/simd-mappings.h
+++ b/ggml/src/ggml-cpu/simd-mappings.h
@@ -2,10 +2,167 @@

 #include "ggml-cpu-impl.h"

+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 //
 // simd mappings
 //

+// FP16 to FP32 conversion
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+//
+// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
+// for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
+//
+#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+
+    static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
+        __fp16 tmp;
+        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __fp16 tmp = f;
+        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+        return res;
+    }
+#elif defined(__F16C__)
+    #ifdef _MSC_VER
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+#elif defined(__POWER9_VECTOR__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) {
+        double d;
+        ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+#elif defined(__riscv) && defined(__riscv_zfhmin)
+    static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        __asm__(
+            "fmv.h.x %[f], %[h]\n\t"
+            "fcvt.s.h %[f], %[f]"
+            : [f] "=&f" (f)
+            : [h] "r" (h)
+        );
+        return f;
+    }
+
+    static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __asm__(
+            "fcvt.h.s %[f], %[f]\n\t"
+            "fmv.x.h %[h], %[f]"
+            : [h] "=&r" (res)
+            : [f] "f" (f)
+        );
+        return res;
+    }
+
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+#elif defined(__NNPA__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) {
+        uint16x8_t v_h = vec_splats(h);
+        uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
+        return vec_extend_to_fp32_hi(v_hd, 0)[0];
+    }
+
+    static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) {
+        float32x4_t v_f = vec_splats(f);
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
+        uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
+        return vec_extract(v_h, 0);
+    }
+#endif
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_CPU_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_CPU_FP32_TO_FP16)
+#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+
 // we define a common set of C macros which map to specific intrinsics based on the current architecture
 // we then implement the fundamental computation operations below using only these macros
 // adding support for new architectures requires to define the corresponding SIMD macros
@@ -32,7 +189,7 @@
 #define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
 #define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
 #define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
 #define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
@@ -415,7 +572,7 @@ static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
    float tmp[8];

    for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
    }

    return _mm256_loadu_ps(tmp);
@@ -426,7 +583,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
    _mm256_storeu_ps(arr, y);

    for (int i = 0; i < 8; i++)
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
 }
 #define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
 #define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
@@ -574,10 +731,10 @@ static inline unsigned char ggml_endian_byte(int i) {
 inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) {
    float tmp[4];

-    tmp[0] = GGML_FP16_TO_FP32(p[0]);
-    tmp[1] = GGML_FP16_TO_FP32(p[1]);
-    tmp[2] = GGML_FP16_TO_FP32(p[2]);
-    tmp[3] = GGML_FP16_TO_FP32(p[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]);

    return wasm_v128_load(tmp);
 }
@@ -587,10 +744,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {

    wasm_v128_store(tmp, x);

-    p[0] = GGML_FP32_TO_FP16(tmp[0]);
-    p[1] = GGML_FP32_TO_FP16(tmp[1]);
-    p[2] = GGML_FP32_TO_FP16(tmp[2]);
-    p[3] = GGML_FP32_TO_FP16(tmp[3]);
+    p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]);
+    p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]);
+    p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]);
+    p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]);
 }

 #define GGML_F16x4             v128_t
@@ -690,10 +847,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
    float tmp[4];

-    tmp[0] = GGML_FP16_TO_FP32(x[0]);
-    tmp[1] = GGML_FP16_TO_FP32(x[1]);
-    tmp[2] = GGML_FP16_TO_FP32(x[2]);
-    tmp[3] = GGML_FP16_TO_FP32(x[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);

    return _mm_loadu_ps(tmp);
 }
@@ -703,10 +860,10 @@ static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {

    _mm_storeu_ps(arr, y);

-    x[0] = GGML_FP32_TO_FP16(arr[0]);
-    x[1] = GGML_FP32_TO_FP16(arr[1]);
-    x[2] = GGML_FP32_TO_FP16(arr[2]);
-    x[3] = GGML_FP32_TO_FP16(arr[3]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }

 #define GGML_F32Cx4             __m128
@@ -828,7 +985,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F32x4_ZERO    __lsx_vldi(0)
 #define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
 #define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
-#define GGML_F32x4_STORE((x),(y))   __lsx_vst((y), (x), 0)
+#define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
@@ -874,10 +1031,10 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
    float tmp[4];

-    tmp[0] = GGML_FP16_TO_FP32(x[0]);
-    tmp[1] = GGML_FP16_TO_FP32(x[1]);
-    tmp[2] = GGML_FP16_TO_FP32(x[2]);
-    tmp[3] = GGML_FP16_TO_FP32(x[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);

    return __lsx_vld(tmp, 0);
 }
@@ -887,10 +1044,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {

    __lsx_vst(y, arr, 0);

-    x[0] = GGML_FP32_TO_FP16(arr[0]);
-    x[1] = GGML_FP32_TO_FP16(arr[1]);
-    x[2] = GGML_FP32_TO_FP16(arr[2]);
-    x[3] = GGML_FP32_TO_FP16(arr[3]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }

 #define GGML_F32Cx4             __m128
@@ -922,7 +1079,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F32_STEP 32
 #define GGML_F32_EPR  4

-#define GGML_F32x4              __vector float
+#define GGML_F32x4              float32x4_t
 #define GGML_F32x4_ZERO         vec_splats(0.0f)
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
@@ -944,10 +1101,8 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
    for (int i = 0; i < offset; ++i) {              \
        x[i] = vec_add(x[i], x[offset + i]);        \
    }                                               \
-    res = vec_extract(x[0], 0) +                    \
-          vec_extract(x[0], 1) +                    \
-          vec_extract(x[0], 2) +                    \
-          vec_extract(x[0], 3);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }

 #define GGML_F32_VEC        GGML_F32x4
@@ -964,28 +1119,45 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_STEP GGML_F32_STEP
 #define GGML_F16_EPR  GGML_F32_EPR

-static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
+static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
+#if defined(__NNPA__)
+    uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x);
+    uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
+    return vec_extend_to_fp32_hi(v_xd, 0);
+#else
    float tmp[4];

    for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
    }

    // note: keep type-cast here to prevent compiler bugs
    // see: https://github.com/ggml-org/llama.cpp/issues/12846
    return vec_xl(0, (const float *)(tmp));
+#endif
 }

-static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
+#if defined(__NNPA__)
+    float32x4_t v_zero = vec_splats(0.0f);
+    uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0);
+    uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0);
+
+    x[0] = vec_extract(v_x, 0);
+    x[1] = vec_extract(v_x, 1);
+    x[2] = vec_extract(v_x, 2);
+    x[3] = vec_extract(v_x, 3);
+#else
    float arr[4];

    // note: keep type-cast here to prevent compiler bugs
    // see: https://github.com/ggml-org/llama.cpp/issues/12846
-    vec_xst(y, 0, (float *)(arr));
+    vec_xst(v_y, 0, (float *)(arr));

    for (int i = 0; i < 4; i++) {
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
    }
+#endif
 }

 #define GGML_F16_VEC                GGML_F32x4
@@ -1006,3 +1178,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
 #define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
 #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
 #endif
+
+#ifdef __cplusplus
+}
+#endif
--- a/ggml/src/ggml-cpu/vec.cpp
+++ b/ggml/src/ggml-cpu/vec.cpp
@@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
        for (int i = 0; i < np; i += ggml_f32_step) {
            ax1 = GGML_F32_VEC_LOAD(x + i);
            ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);

            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
+            sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);

            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+            sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);

            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+            sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);

            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+            sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);

            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+            sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);

            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+            sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);

            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
+            sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
        }
        // leftovers
        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
@@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
        for (int i = np; i < np2; i += ggml_f32_epr) {
            ax1 = GGML_F32_VEC_LOAD(x + i);
            ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
        }
        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
        if (np2 < n) {
@@ -219,11 +219,11 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G

    // leftovers
    for (int i = np; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
    }
 #else
    for (int i = 0; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
    }
 #endif

@@ -254,6 +254,30 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
    }
 }

+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = ggml_silu_f32(x[i]) * g[i];
+    }
+}
+
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
    int i = 0;
    ggml_float sum = 0;
--- a/ggml/src/ggml-cpu/vec.h
+++ b/ggml/src/ggml-cpu/vec.h
@@ -58,7 +58,7 @@ inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf
 inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
 inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
    for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
    }
 }
 inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
@@ -67,7 +67,7 @@ inline static void ggml_vec_acc1_f32(const int n, float * y, const float   v)
 inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] - y[i]; }
 inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
    for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i]));
    }
 }
 inline static void ggml_vec_set_f32 (const int n, float * x, const float   v)                  { for (int i = 0; i < n; ++i) x[i]  = v;           }
@@ -75,20 +75,20 @@ inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)
 inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = -x[i];       }
 inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i]));
+        y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_FP16_TO_FP32(x[i]));
    }
 }

 inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
 inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
    for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i]));
    }
 }
 inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
 inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
    for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i]));
    }
 }

@@ -131,13 +131,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
    // leftovers
    for (int i = np; i < n; ++i) {
        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
        }
    }
 #else
    for (int i = 0; i < n; ++i) {
        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
        }
    }
 #endif
@@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const

            ax1 = GGML_F32_VEC_LOAD(x + i);
            ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);

            GGML_F32_VEC_STORE(y + i, ay1);

            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+            ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);

            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);

            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+            ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);

            GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);

            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+            ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);

            GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);

            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+            ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);

            GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);

            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+            ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);

            GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);

            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+            ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);

            GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);

            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+            ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);

            GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
        }
@@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
        for (int i = np; i < np2; i += ggml_f32_epr) {
            ax1 = GGML_F32_VEC_LOAD(x + i);
            ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);

            GGML_F32_VEC_STORE(y + i, ay1);
        }
@@ -280,12 +280,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,

    // leftovers
    for (int i = np; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
    }
 #else
    // scalar
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
    }
 #endif
 }
@@ -351,6 +351,45 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
 #endif
 }

+inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, const float s, const float b) {
+#if defined(GGML_USE_ACCELERATE)
+    vDSP_vsmsa(x, 1, &s, &b, y, 1, n);
+#elif defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+        // scalar ; TODO: Write SVE code
+        for (int i = 0; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
+
+        GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);
+        GGML_F32_VEC vb = GGML_F32_VEC_SET1(b);
+
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], vs, vb);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #endif
+#else
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = x[i]*s + b;
+    }
+#endif
+}
+
 //inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) { for (int i = 0; i < n; ++i) y[i] *= v;          }
 inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
@@ -430,12 +469,12 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float

    // leftovers
    for (int i = np; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
    }
 #else
    // scalar
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
    }
 #endif
 }
@@ -444,103 +483,103 @@ inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) {
 inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
 inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v*v);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v*v);
    }
 }
 inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
 inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);  }
 inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_sin_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]);  }
 inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_cos_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]);  }
 inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
 inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
    }
 }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
+        y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
    }
 }
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
 inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : 0.f);
    }
 }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
    }
 }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
 inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i]))));
+        y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_FP16_TO_FP32(x[i]))));
    }
 }
 // TODO: optimize performance
 inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
    }
 }
 inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
+        y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
    }
 }
 inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
 inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i])));
    }
 }

@@ -562,9 +601,9 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp

 inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float xi = GGML_FP16_TO_FP32(x[i]);
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
        float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
-        y[i] = GGML_FP32_TO_FP16(res);
+        y[i] = GGML_CPU_FP32_TO_FP16(res);
    }
 }

@@ -577,9 +616,9 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
        } else if (x[i] >= 10.0f) {
            y[i] = x[i];
        } else {
-            ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+            ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
            memcpy(&t, &fp16, sizeof(uint16_t));
-            y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+            y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]);
        }
    }
 }
@@ -613,9 +652,9 @@ inline static float ggml_gelu_quick_f32(float x) {
 inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
    uint16_t t;
    for (int i = 0; i < n; ++i) {
-        ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
        memcpy(&t, &fp16, sizeof(uint16_t));
-        y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
    }
 }
 #else
@@ -628,8 +667,8 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *

 inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
    }
 }

@@ -638,8 +677,8 @@ inline static float ggml_silu_f32(float x) {
    return x/(1.0f + expf(-x));
 }
 inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
-    float v = GGML_FP16_TO_FP32(x);
-    return GGML_FP32_TO_FP16(v/(1.0f + expf(-v)));
+    float v = GGML_CPU_FP16_TO_FP32(x);
+    return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v)));
 }

 #if __FINITE_MATH_ONLY__
@@ -888,9 +927,9 @@ inline static float ggml_silu_backward_f32(float x, float dy) {
 }

 inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
-    const float v = GGML_FP16_TO_FP32(x);
+    const float v = GGML_CPU_FP16_TO_FP32(x);
    const float s = 1.0f/(1.0f + expf(-v));
-    return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
+    return GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
 }

 inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
@@ -905,6 +944,100 @@ inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, con
    }
 }

+inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = (x[i] > 0.f) ? x[i] * g[i] : 0.f;
+    }
+}
+
+inline static void ggml_vec_reglu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v * GGML_CPU_FP16_TO_FP32(g[i]) : 0.f);
+    }
+}
+
+#ifdef GGML_GELU_FP16
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        if (x[i] <= -10.0f) {
+            y[i] = 0.0f;
+        } else if (x[i] >= 10.0f) {
+            y[i] = x[i] * g[i];
+        } else {
+            ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+            memcpy(&t, &fp16, sizeof(uint16_t));
+            y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]) * g[i];
+        }
+    }
+}
+#else
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[i16[i]]) * v);
+    }
+}
+
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g);
+
+inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        float w = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+        float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+    }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+        memcpy(&t, &fp16, sizeof(uint16_t));
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
+    }
+}
+#else
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
+    }
+}
+
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
    ggml_float sum = 0.0;
@@ -928,7 +1061,7 @@ inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float
 inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) {
    float sum = 0.0f;
    for (int i = 0; i < n; ++i) {
-        sum += GGML_FP16_TO_FP32(x[i]);
+        sum += GGML_CPU_FP16_TO_FP32(x[i]);
    }
    *s = sum;
 }
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -19,10 +19,10 @@
 #endif
 #include "ggml-common.h"

-#include <cstdio>
 #include <array>
 #include <cassert>
 #include <cfloat>
+#include <cstdio>
 #include <string>
 #include <vector>

@@ -76,11 +76,9 @@
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)

 // Moore Threads
-#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
-
-#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD

 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
@@ -177,6 +175,23 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif

+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes)                                                       \
+        do {                                                                                                   \
+            static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = { false };                         \
+            const int   id                                                = ggml_cuda_get_device();            \
+            if (!shared_memory_limit_raised[id]) {                                                             \
+                CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+                shared_memory_limit_raised[id] = true;                                                         \
+            }                                                                                                  \
+        } while (0)
+#else
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+        do {                                             \
+            GGML_UNUSED(nbytes);                         \
+        } while (0)
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
@@ -203,13 +218,13 @@ typedef float2 dfloat2;
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610

-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)

-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))

 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -219,9 +234,9 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE

-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)

 static bool fp16_available(const int cc) {
    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
@@ -233,7 +248,8 @@ static bool fast_fp16_available(const int cc) {

 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fast_fp16_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }

 // Any FP16 tensor core instructions are available for ggml code.
@@ -241,15 +257,35 @@ static bool fp16_mma_available(const int cc) {
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
    return false;
 #else
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
+    if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) ||
+        GGML_CUDA_CC_IS_MTHREADS(cc)) {
+        return true;
+    } else if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+    } else {
+        return false;
+    }
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }

 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+}
+
+static bool bf16_mma_hardware_available(const int cc) {
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+}
+
+static bool fp32_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_CDNA(cc);
 }

 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
@@ -262,11 +298,11 @@ static bool cp_async_available(const int cc) {
 }

 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
    return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }

 [[noreturn]]
@@ -362,6 +398,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }

+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -767,21 +823,7 @@ struct ggml_backend_cuda_context {
        name(GGML_CUDA_NAME + std::to_string(device)) {
    }

-    ~ggml_backend_cuda_context() {
-        if (copy_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(copy_event));
-        }
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-                if (streams[i][j] != nullptr) {
-                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-                }
-            }
-            if (cublas_handles[i] != nullptr) {
-                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-            }
-        }
-    }
+    ~ggml_backend_cuda_context();

    cudaStream_t stream(int device, int stream) {
        if (streams[device][stream] == nullptr) {
--- a/ggml/src/ggml-cuda/conv2d-dw.cu
+++ b/ggml/src/ggml-cuda/conv2d-dw.cu
@@ -0,0 +1,161 @@
+#include "conv2d-dw.cuh"
+
+struct conv_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) {
+    kernel_bounds bounds;
+    bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.y_max =
+        min(params.kernel_h,
+            (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    bounds.x_max =
+        min(params.kernel_w,
+            (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    return bounds;
+}
+
+__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+struct whcn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h +
+               y * params.out_w + x;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        out_x = global_idx % params.out_w;
+        out_y = (global_idx / params.out_w) % params.out_h;
+        c     = (global_idx / (params.out_w * params.out_h)) % params.channels;
+        n     = global_idx / (params.out_w * params.out_h * params.channels);
+    }
+};
+
+struct cwhn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return (ky * params.kernel_w + kx) * params.channels + c;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) +
+               x * params.channels + c;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        c     = global_idx % params.channels;
+        out_x = (global_idx / params.channels) % params.out_w;
+        out_y = (global_idx / (params.channels * params.out_w)) % params.out_h;
+        n     = global_idx / (params.channels * params.out_w * params.out_h);
+    }
+};
+
+template <typename T, typename Layout>
+__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output,
+                                 const int in_w, const int in_h, const int out_w, const int out_h,
+                                 const int kernel_w, const int kernel_h, const int stride_x, const int stride_y,
+                                 const int padding_x, const int padding_y, const int dilation_x, const int dilation_y,
+                                 const int channels, const int batches) {
+    const int global_idx     = blockIdx.x * blockDim.x + threadIdx.x;
+    const int total_elements = batches * channels * out_h * out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    conv_params params = { in_w,     in_h,      out_w,     out_h,      kernel_w,   kernel_h, stride_x,
+                           stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches };
+
+    int batch_idx, channel_idx, out_y_idx, out_x_idx;
+    Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx);
+
+    T accumulator = 0;
+    kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params);
+
+    for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) {
+        int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y);
+
+        for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) {
+            int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x);
+
+            const T input_val  = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)];
+            const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)];
+
+            accumulator += input_val * kernel_val;
+        }
+    }
+
+    output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator;
+}
+
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    cudaStream_t st = ctx.stream();
+
+    const int total  = batches * channels * out_h * out_w;
+    const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE;
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_kernel<float, whcn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_kernel<float, cwhn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv_2d_dw");
+    }
+}
--- a/ggml/src/ggml-cuda/conv2d-dw.cuh
+++ b/ggml/src/ggml-cuda/conv2d-dw.cuh
@@ -0,0 +1,5 @@
+#pragma once
+#include "common.cuh"
+
+#define CUDA_CONV2D_DW_BLOCK_SIZE 256
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml/src/ggml-cuda/conv2d-transpose.cu
+++ b/ggml/src/ggml-cuda/conv2d-transpose.cu
@@ -0,0 +1,91 @@
+#include <algorithm>
+
+#include "conv2d-transpose.cuh"
+#include "ggml.h"
+
+__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
+                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
+                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
+                                        const int c_in, const int c_out, const int batches) {
+    const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    const int total_elements = out_w * out_h * c_out * batches;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int out_x_idx = global_idx % out_w;
+    const int out_y_idx = (global_idx / out_w) % out_h;
+    const int c_idx     = (global_idx / (out_w * out_h)) % c_out;
+    const int n_idx     = global_idx / (out_w * out_h * c_out);
+
+    float accumulator = 0;
+    // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds
+
+    for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
+        for (int kh = 0; kh < kernel_h; ++kh) {
+            int in_y = out_y_idx - kh;
+            if (in_y < 0 || in_y % stride) continue;
+            in_y /= stride;
+            if (in_y >= in_h) continue;
+
+            for (int kw = 0; kw < kernel_w; ++kw) {
+                int in_x = out_x_idx - kw;
+                if (in_x < 0 || in_x % stride) continue;
+                in_x /= stride;
+                if (in_x >= in_w) continue;
+
+                const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
+                const int kernel_idx =
+                    (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
+
+                float input_val = input[input_idx];
+                half  kern_val  = kernel[kernel_idx];
+
+                accumulator += input_val * (float) kern_val;
+            }
+        }
+    }
+
+    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator;
+}
+
+//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in)
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * input_data  = (const float *) input->data;
+    float *       output_data = (float *) dst->data;
+    const half * kernel_data = (const half *) kernel->data;
+
+    const int input_w      = input->ne[0];
+    const int input_h      = input->ne[1];
+    const int output_w     = dst->ne[0];
+    const int output_h     = dst->ne[1];
+    const int channels_in  = input->ne[2];
+    const int channels_out = kernel->ne[2];
+    const int kernel_w     = kernel->ne[0];
+    const int kernel_h     = kernel->ne[1];
+    const int stride       = dst->op_params[0];
+    const int batches      = input->ne[3];
+
+    GGML_ASSERT(channels_in == kernel->ne[3]);
+    GGML_ASSERT(stride > 0);
+
+    cudaStream_t st = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(input));
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int total  = (output_w * output_h * channels_out * batches);
+    const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
+
+    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
+        channels_in, channels_out, batches);
+}
--- a/ggml/src/ggml-cuda/conv2d-transpose.cuh
+++ b/ggml/src/ggml-cuda/conv2d-transpose.cuh
@@ -0,0 +1,4 @@
+#include "common.cuh"
+
+#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -728,3 +728,25 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
            return nullptr;
    }
 }
+
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, float>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cuda<nv_bfloat16, float>;
+        default:
+            return nullptr;
+    }
+}
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -22,5 +22,10 @@ using to_t_nc_cuda_t = void (*)(const void * x, T * y,
    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
    int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);

+typedef to_t_nc_cuda_t<float> to_fp32_nc_cuda_t;
 typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
--- a/ggml/src/ggml-cuda/cross-entropy-loss.cu
+++ b/ggml/src/ggml-cuda/cross-entropy-loss.cu
@@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
    ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);

    if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
        cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
    } else {
        cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@@ -175,13 +169,7 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten
    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;

    if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
        cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
    } else {
        cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -32,7 +32,9 @@ typedef void (* fattn_kernel_t)(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -851,7 +853,8 @@ void launch_fattn(
        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
+        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
        Q->nb[1], Q->nb[2], Q->nb[3],
        nb11, nb12, nb13,
        nb21, nb22, nb23,
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -1223,7 +1223,9 @@ static __global__ void flash_attn_ext_f16(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -1288,7 +1290,8 @@ static __global__ void flash_attn_ext_f16(

        const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
        const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-        const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+        const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+            (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
        float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);

        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1327,7 +1330,8 @@ static __global__ void flash_attn_ext_f16(

    const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
    const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-    const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+    const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+        (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
    float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);

    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1348,8 +1352,8 @@ static __global__ void flash_attn_ext_f16(
    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -6,7 +6,7 @@

 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
        const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f16(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -64,7 +66,7 @@ static __global__ void flash_attn_tile_ext_f16(
    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
+    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);

    const int stride_KV2 = nb11 / sizeof(half2);

@@ -288,8 +290,8 @@ static __global__ void flash_attn_tile_ext_f16(
    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -6,7 +6,7 @@

 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
        const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f32(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -58,8 +60,8 @@ static __global__ void flash_attn_tile_ext_f32(
        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
        GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -76,7 +78,7 @@ static __global__ void flash_attn_tile_ext_f32(
    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
+    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);

    const int stride_KV2 = nb11 / sizeof(half2);

@@ -297,14 +299,14 @@ static __global__ void flash_attn_tile_ext_f32(
    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
    NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f16(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -68,7 +70,7 @@ static __global__ void flash_attn_vec_ext_f16(
    K += nb12*(blockIdx.z / gqa_ratio);
    V += nb22*(blockIdx.z / gqa_ratio);

-    const half * maskh = (const half   *)  mask + ne11*ic0;
+    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);

    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
    const half  slopeh = __float2half(slopef);
@@ -342,8 +344,8 @@ static __global__ void flash_attn_vec_ext_f16(
    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f32(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -51,8 +53,8 @@ static __global__ void flash_attn_vec_ext_f32(
        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
        GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -79,7 +81,8 @@ static __global__ void flash_attn_vec_ext_f32(
    Q += nb02* blockIdx.z              + nb01*ic0;
    K += nb12*(blockIdx.z / gqa_ratio);
    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
-    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);

    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);

@@ -334,13 +337,15 @@ static __global__ void flash_attn_vec_ext_f32(
    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
-    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
    NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -9,7 +9,11 @@
 #ifdef FP16_MMA_AVAILABLE
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 #include <mma.h>
+#ifdef GGML_USE_MUSA
+namespace wmma = mtmusa::wmma;
+#else // GGML_USE_MUSA
 namespace wmma = nvcuda::wmma;
+#endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
 #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
@@ -42,7 +46,9 @@ static __global__ void flash_attn_ext_f16(
        const int ne12,
        const int ne13,
        const int ne31,
+        const int ne32,
        const int nb31,
+        const int nb32,
        const int nb01,
        const int nb02,
        const int nb03,
@@ -90,11 +96,11 @@ static __global__ void flash_attn_ext_f16(
    constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);

    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q + nb02* blockIdx.z              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.z / gqa_ratio));
-    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
-    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+    const float * Q_f   = (const float *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half  * V_h   = (const half  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const half2 * mask2 = (const half2 *)  maskh;

    const int stride_Q  = nb01 / sizeof(float);
    const int stride_KV = nb11 / sizeof(half);
@@ -436,7 +442,7 @@ static __global__ void flash_attn_ext_f16(
    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -168,6 +168,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
+        case GGML_TYPE_I32:
+            get_rows_cuda_float((const int32_t *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
        case GGML_TYPE_BF16:
            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
@@ -210,6 +214,10 @@ void get_rows_cuda(
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
+        case GGML_TYPE_I32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (int32_t *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
        case GGML_TYPE_F16:
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -11,6 +11,8 @@
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
+#include "ggml-cuda/conv2d-dw.cuh"
+#include "ggml-cuda/conv2d-transpose.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
@@ -35,6 +37,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
@@ -47,6 +50,7 @@
 #include <atomic>
 #include <charconv>
 #include <cinttypes>
+#include <condition_variable>
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
@@ -54,9 +58,8 @@
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdint.h>
-#include <stdio.h>
 #include <stdarg.h>
+#include <stdio.h>
 #include <stdlib.h>
 #include <string>
 #include <vector>
@@ -97,8 +100,7 @@ int ggml_cuda_get_device() {
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
    ggml_cuda_set_device(device);
    cudaError_t err;
-    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
-    {
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr) {
        err = cudaMallocManaged(ptr, size);
 #if defined(GGML_USE_HIP)
        if (err == hipSuccess) {
@@ -116,9 +118,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
            err = cudaMalloc(ptr, size);
        }
 #endif // defined(GGML_USE_HIP)
-    }
-    else
-    {
+    } else {
        err = cudaMalloc(ptr, size);
    }
    return err;
@@ -514,6 +514,33 @@ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(i
    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }

+// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
+// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured
+
+static std::mutex ggml_cuda_lock;
+static std::condition_variable ggml_cuda_lock_cv;
+static std::atomic<int> ggml_cuda_lock_counter;
+
+ggml_backend_cuda_context::~ggml_backend_cuda_context() {
+    std::unique_lock<std::mutex> lock(ggml_cuda_lock);
+    ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; });
+
+    if (copy_event != nullptr) {
+        CUDA_CHECK(cudaEventDestroy(copy_event));
+    }
+    for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+        for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+            if (streams[i][j] != nullptr) {
+                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+            }
+        }
+        if (cublas_handles[i] != nullptr) {
+            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+        }
+    }
+}
+
+
 // cuda buffer

 struct ggml_backend_cuda_buffer_context {
@@ -1200,9 +1227,12 @@ static void ggml_cuda_op_mul_mat_cublas(

    const int cc = ggml_cuda_info().devices[id].cc;

+    const bool supports_bf16 = GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+
    const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;

-    if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
+    if (supports_bf16 && src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
        ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
        if (src1->type != GGML_TYPE_BF16) {
            const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type);
@@ -1230,7 +1260,7 @@ static void ggml_cuda_op_mul_mat_cublas(

        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
        to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-    } else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
+    } else if (fast_fp16_hardware_available(cc) && use_fp16) {
        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
        if (src0->type != GGML_TYPE_F16) {
@@ -1719,7 +1749,7 @@ static void ggml_cuda_op_mul_mat(
 }

 static __global__ void k_compute_batched_ptrs(
-        const half * src0_as_f16, const half * src1_as_f16, char * dst,
+        const void * src0_as_f16, const void * src1_as_f16, char * dst,
        const void ** ptrs_src, void ** ptrs_dst,
        int64_t ne12, int64_t ne13,
        int64_t ne23,
@@ -1742,83 +1772,131 @@ static __global__ void k_compute_batched_ptrs(
    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }

-static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+// Type traits for mapping ggml types to CUDA/cuBLAS types
+template<ggml_type T>
+struct batched_mul_mat_traits;
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F32> {
+    using cuda_type = float;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_32F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F32;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp32_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_BF16> {
+    using cuda_type = nv_bfloat16;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_16BF;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_BF16;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_bf16_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F16> {
+    using cuda_type = half;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_16F;
+    static inline const cudaDataType_t data_type = CUDA_R_16F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F16;
+    static inline const half alpha = 1.0;
+    static inline const half beta = 0.0;
+    static inline const void* get_alpha() { static const half val = alpha; return &val; }
+    static inline const void* get_beta() { static const half val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp16_nc_cuda(src_type); }
+};
+
+template<ggml_type src0_type>
+static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    using traits = batched_mul_mat_traits<src0_type>;
+    using cuda_t = typename traits::cuda_type;
+
    GGML_ASSERT(!ggml_is_transposed(src0));
    GGML_ASSERT(!ggml_is_transposed(src1));
-
    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == src0_type);
+    GGML_ASSERT(ggml_is_contiguous(dst));

    // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
    // As long as dst is contiguous this does not matter though.
-    GGML_ASSERT(ggml_is_contiguous(dst));

    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t ne_dst = ggml_nelements(dst);
-
    cudaStream_t main_stream = ctx.stream();
-
    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));

-    const half * src0_f16 = (const half *) src0->data;
    float * dst_ddf = (float *) dst->data;
-
-    const half * src1_f16 = (const half *) src1->data;
    const size_t ts_src1 = ggml_type_size(src1->type);
    GGML_ASSERT(nb10 == ts_src1);
    int64_t s11 = nb11 / ts_src1;
    int64_t s12 = nb12 / ts_src1;
    int64_t s13 = nb13 / ts_src1;
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());

-    // convert src1 to fp16
-    if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
+    const cuda_t * src0_ptr = nullptr;
+    const cuda_t * src1_ptr = nullptr;
+
+    ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
+    ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
+
+    // Handle src0
+    src0_ptr = (const cuda_t *) src0->data;
+
+    // Handle src1 - convert if necessary
+    if (src1->type == src0_type) {
+        src1_ptr = (const cuda_t *) src1->data;
+    } else {
+        // Convert src1 to target type using traits conversion functions
        const int64_t ne_src1 = ggml_nelements(src1);
-        src1_f16_alloc.alloc(ne_src1);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
+        src1_alloc.alloc(ne_src1);

-        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
-
-        src1_f16 = src1_f16_alloc.get();
+        const auto convert_func = traits::get_nc_converter(src1->type);
+        GGML_ASSERT(convert_func != nullptr);
+        convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+        src1_ptr = src1_alloc.get();
        s11 = ne10;
        s12 = ne11*s11;
        s13 = ne12*s12;
    }

-    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
+    // Setup destination buffer
+    ggml_cuda_pool_alloc<cuda_t> dst_temp(ctx.pool());
    char * dst_t;
-
-    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
-    cudaDataType_t      cu_data_type    = CUDA_R_16F;
-
-    // dst strides
    size_t nbd2 = dst->nb[2];
    size_t nbd3 = dst->nb[3];

-    const half  alpha_f16 = 1.0f;
-    const half  beta_f16  = 0.0f;
-
+    cublasComputeType_t cu_compute_type = traits::compute_type;
+    cudaDataType_t cu_data_type = traits::data_type;
+    cudaDataType_t cu_data_type_a = traits::data_type;
+    cudaDataType_t cu_data_type_b = traits::data_type;
+    const void * alpha = traits::get_alpha();
+    const void * beta = traits::get_beta();
    const float alpha_f32 = 1.0f;
-    const float beta_f32  = 0.0f;
-
-    const void * alpha = &alpha_f16;
-    const void * beta  = &beta_f16;
+    const float beta_f32 = 0.0f;

    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        dst_t = (char *) dst_f16.alloc(ne_dst);
-
-        nbd2 /= sizeof(float) / sizeof(half);
-        nbd3 /= sizeof(float) / sizeof(half);
+        if constexpr (src0_type == GGML_TYPE_F32) {
+            dst_t = (char *) dst_ddf;  // Direct F32 output
+        } else {
+            dst_t = (char *) dst_temp.alloc(ne_dst);
+            nbd2 /= sizeof(float) / sizeof(cuda_t);
+            nbd3 /= sizeof(float) / sizeof(cuda_t);
+        }
    } else {
        dst_t = (char *) dst_ddf;
-
        cu_compute_type = CUBLAS_COMPUTE_32F;
-        cu_data_type    = CUDA_R_32F;
-
+        cu_data_type = CUDA_R_32F;
        alpha = &alpha_f32;
-        beta  = &beta_f32;
+        beta = &beta_f32;
    }

    int id = ggml_cuda_get_device();
@@ -1826,7 +1904,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
    if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
        cu_compute_type = CUBLAS_COMPUTE_32F;
        alpha = &alpha_f32;
-        beta  = &beta_f32;
+        beta = &beta_f32;
    }

    GGML_ASSERT(ne12 % ne02 == 0);
@@ -1836,35 +1914,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
    const int64_t r2 = ne12/ne02;
    const int64_t r3 = ne13/ne03;

-#if 0
-    // use cublasGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                CUBLAS_CHECK(
-                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
-                    ne01, ne11, ne10,
-                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
-                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
-                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
-                    cu_compute_type,
-                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
        // use cublasGemmStridedBatchedEx
        CUBLAS_CHECK(
        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
-                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
-                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
+                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
                ne12*ne13,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1875,34 +1933,55 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
        ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
        ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);

+        size_t src1_stride_size = sizeof(cuda_t);
+
        dim3 block_dims(ne13, ne12);
        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_f16, src1_f16, dst_t,
+                src0_ptr, src1_ptr, dst_t,
                ptrs_src.get(), ptrs_dst.get(),
                ne12, ne13,
                ne23,
                nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
-                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
+                (src1->type == src0_type) ? nb12 : s12*src1_stride_size,
+                (src1->type == src0_type) ? nb13 : s13*src1_stride_size,
                nbd2, nbd3,
                r2, r3);
+
        CUDA_CHECK(cudaGetLastError());

        CUBLAS_CHECK(
        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
-                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
+                alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
+                       (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type,   ne0,
                ne23,
                cu_compute_type,
                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
    }
-#endif

-    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
-        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
+    // Convert output back to F32 if needed
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type != CUDA_R_32F) {
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(traits::ggml_type_val);
+        to_fp32_cuda(dst_temp.get(), dst_ddf, ne_dst, main_stream);
+    }
+}
+
+static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F32);
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F32>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_BF16>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F16>(ctx, src0, src1, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupported type");
    }
 }

@@ -1916,16 +1995,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
        && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;

    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
    bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;

-    bool any_gpus_with_slow_fp16   = false;
-    bool any_gpus_without_fp16_mma = false;
+    bool any_gpus_with_slow_fp16 = false;

    if (split) {
        ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1936,16 +2013,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                continue;
            }

-            const int cc              = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+            const int cc            = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
        }
    } else {
-        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
    }

    // debug helpers
@@ -1956,7 +2033,13 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);

-    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    //TODO update for generic tensor parallelism
+    const int cc                     = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    bool use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
+    bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
+    bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
+
+    if (!split && use_mul_mat_vec) {
        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
        ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
@@ -1964,8 +2047,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
        ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
    } else if (!split && use_mul_mat_q) {
        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
-            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
        // general KQ + KQV multi-batch without FlashAttention
        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
    } else if (use_mul_mat_vec) {
@@ -2220,6 +2303,27 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                    return false;
            }
            break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    ggml_cuda_op_reglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    ggml_cuda_op_geglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_SWIGLU:
+                    ggml_cuda_op_swiglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_ERF:
+                    ggml_cuda_op_geglu_erf(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_cuda_op_geglu_quick(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
        case GGML_OP_NORM:
            ggml_cuda_op_norm(ctx, dst);
            break;
@@ -2310,6 +2414,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
        case GGML_OP_IM2COL:
            ggml_cuda_op_im2col(ctx, dst);
            break;
+        case GGML_OP_CONV_2D_DW:
+            ggml_cuda_op_conv2d_dw(ctx, dst);
+            break;
+        case GGML_OP_CONV_TRANSPOSE_2D:
+            ggml_cuda_conv_2d_transpose_p0(ctx, dst);
+            break;
        case GGML_OP_CONV_TRANSPOSE_1D:
            ggml_cuda_op_conv_transpose_1d(ctx,dst);
            break;
@@ -2322,6 +2432,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
        case GGML_OP_SUM_ROWS:
            ggml_cuda_op_sum_rows(ctx, dst);
            break;
+        case GGML_OP_MEAN:
+            ggml_cuda_op_mean(ctx, dst);
+            break;
        case GGML_OP_SSM_CONV:
            ggml_cuda_op_ssm_conv(ctx, dst);
            break;
@@ -2664,7 +2777,9 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                    }
                }
-#endif
+#else
+                GGML_UNUSED(integrated);
+#endif // NDEBUG

                bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                if (!ok) {
@@ -2683,6 +2798,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx

            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
            graph_evaluated_or_captured = true; // CUDA graph has been captured
+
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
+                ggml_cuda_lock_cv.notify_all();
+            }
        } else {
            graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
        }
@@ -2758,7 +2878,13 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
        }
    }

-    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
+    if (use_cuda_graph && cuda_graph_update_required) {
+        // Start CUDA graph capture
+        {
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
+        }
+
        CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
    }

@@ -2991,6 +3117,18 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                    return false;
            }
            break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
        case GGML_OP_MUL_MAT:
        case GGML_OP_MUL_MAT_ID:
            {
@@ -3014,9 +3152,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                    return false;
                }
 #ifdef GGML_USE_MUSA
-                if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
-                    !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
-                    return false;
+                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+                if (b->ne[2]*b->ne[3] > 1 && !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
+                    if (GGML_CUDA_CC_IS_QY1(cc) && op->op == GGML_OP_MUL_MAT &&
+                            a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) {
+                        return false;
+                    }
+                    if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID &&
+                            a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) {
+                        return false;
+                    }
                }
 #endif // GGML_USE_MUSA
                switch (a->type) {
@@ -3043,11 +3188,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                    case GGML_TYPE_IQ4_NL:
                    case GGML_TYPE_IQ4_XS:
                    case GGML_TYPE_BF16:
-#ifdef GGML_USE_MUSA
-                        if (a->type == GGML_TYPE_Q3_K) {
-                            return false;
-                        }
-#endif // GGML_USE_MUSA
                        return true;
                    default:
                        return false;
@@ -3060,6 +3200,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                switch (op->src[0]->type) {
                    case GGML_TYPE_F16:
                    case GGML_TYPE_F32:
+                    case GGML_TYPE_BF16:
+                    case GGML_TYPE_I32:
                    case GGML_TYPE_Q4_0:
                    case GGML_TYPE_Q4_1:
                    case GGML_TYPE_Q5_0:
@@ -3189,12 +3331,26 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_COS:
        case GGML_OP_CLAMP:
        case GGML_OP_LOG:
-        case GGML_OP_SSM_SCAN:
-        case GGML_OP_SSM_CONV:
            return true;
+        case GGML_OP_SSM_SCAN: {
+            if (op->src[3]->ne[0] == 1) {
+                // Mamba2
+                // (kernel only supports (d_state == 128 || d_state == 256) && d_head % 16 == 0)
+                return (op->src[0]->ne[0] == 128 || op->src[0]->ne[0] == 256) && op->src[0]->ne[1] % 16 == 0;
+            } else {
+                // Mamba
+                // (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
+                return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
+            }
+        }
+        case GGML_OP_SSM_CONV: {
+            // assumes d_inner % threads == 0
+            return op->src[0]->ne[1] % 128 == 0;
+        }
        case GGML_OP_CONT:
            return op->src[0]->type != GGML_TYPE_BF16;
        case GGML_OP_DIAG_MASK_INF:
+            return true;
        case GGML_OP_SOFT_MAX:
            return true;
        case GGML_OP_SOFT_MAX_BACK: {
@@ -3207,16 +3363,18 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
        }
        case GGML_OP_IM2COL:
+        case GGML_OP_CONV_2D_DW:
+        case GGML_OP_CONV_TRANSPOSE_2D:
        case GGML_OP_POOL_2D:
        case GGML_OP_SUM:
        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
        case GGML_OP_ARGSORT:
        case GGML_OP_ACC:
            return true;
        case GGML_OP_GROUP_NORM:
            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
        case GGML_OP_PAD:
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3240,6 +3398,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            if (op->src[0]->ne[0] == 192) {
                return false;
            }
+            // TODO: support broadcast
+            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
+            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
            if (op->src[0]->ne[3] != 1) {
                return false;
            }
--- a/ggml/src/ggml-cuda/mean.cu
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -0,0 +1,19 @@
+#include "mean.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) src0->data;
+    float *             dst_d  = (float *) dst->data;
+    cudaStream_t        stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+}
--- a/Show More
+++ b/Show More